diff --git a/ldm/__pycache__/util.cpython-38.pyc b/ldm/__pycache__/util.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..647db21764a46cfe0dec5066d6cacfa5333f1f0f Binary files /dev/null and b/ldm/__pycache__/util.cpython-38.pyc differ diff --git a/ldm/data/.DS_Store b/ldm/data/.DS_Store new file mode 100644 index 0000000000000000000000000000000000000000..5008ddfcf53c02e82d7eee2e57c38e5672ef89f6 Binary files /dev/null and b/ldm/data/.DS_Store differ diff --git a/ldm/data/__init__.py b/ldm/data/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/ldm/data/__pycache__/__init__.cpython-38.pyc b/ldm/data/__pycache__/__init__.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..bac105feea1c8d9dcd97fdd803ee43026a664611 Binary files /dev/null and b/ldm/data/__pycache__/__init__.cpython-38.pyc differ diff --git a/ldm/data/__pycache__/image_dresscode.cpython-38.pyc b/ldm/data/__pycache__/image_dresscode.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..e372c88646394ef6080dfef627ed1ac03e7268dd Binary files /dev/null and b/ldm/data/__pycache__/image_dresscode.cpython-38.pyc differ diff --git a/ldm/data/__pycache__/image_vitonhd.cpython-38.pyc b/ldm/data/__pycache__/image_vitonhd.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..b65917b59481ceb00b3dd322e6a8cc53526a813a Binary files /dev/null and b/ldm/data/__pycache__/image_vitonhd.cpython-38.pyc differ diff --git a/ldm/data/__pycache__/viton-images.cpython-38.pyc b/ldm/data/__pycache__/viton-images.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..385a3d287fc220ee957bb05fb971d7dd26e3a875 Binary files /dev/null and b/ldm/data/__pycache__/viton-images.cpython-38.pyc differ diff --git a/ldm/data/base.py b/ldm/data/base.py new file mode 100644 index 0000000000000000000000000000000000000000..95f4c8b445022cf53f3a2506fbfe531d96bae02e --- /dev/null +++ b/ldm/data/base.py @@ -0,0 +1,23 @@ +from abc import abstractmethod +from torch.utils.data import Dataset, ConcatDataset, ChainDataset, IterableDataset + + +class Txt2ImgIterableBaseDataset(IterableDataset): + ''' + Define an interface to make the IterableDatasets for text2img data chainable + ''' + def __init__(self, num_records=0, valid_ids=None, size=256): + super().__init__() + self.num_records = num_records + self.valid_ids = valid_ids + self.sample_ids = valid_ids + self.size = size + + print(f'{self.__class__.__name__} dataset contains {self.__len__()} examples.') + + def __len__(self): + return self.num_records + + @abstractmethod + def __iter__(self): + pass \ No newline at end of file diff --git a/ldm/data/image_dresscode.py b/ldm/data/image_dresscode.py new file mode 100644 index 0000000000000000000000000000000000000000..ce468455979b70c421f47d6ad950bd32ec855fc1 --- /dev/null +++ b/ldm/data/image_dresscode.py @@ -0,0 +1,86 @@ +import os +import torch +import torchvision +import torch.utils.data as data +import torchvision.transforms.functional as F +import numpy as np +from PIL import Image + +class OpenImageDataset(data.Dataset): + def __init__(self, state, dataset_dir, type="paired"): + self.state = state # train or test + self.dataset_dir = dataset_dir # /home/sd/zjh/Dataset/DressCode + + # 确定状态 + if state == "train": + self.dataset_file = os.path.join(dataset_dir, "train_pairs.txt") + if state == "test": + if type == "unpaired": + self.dataset_file = os.path.join(dataset_dir, "test_pairs_unpaired.txt") + if type == "paired": + self.dataset_file = os.path.join(dataset_dir, "test_pairs_paired.txt") + + # 加载数据集 + self.people_list = [] + self.clothes_list = [] + with open(self.dataset_file, 'r') as f: + for line in f.readlines(): + people, clothes, category = line.strip().split() + if category == "0": + category = "upper_body" + elif category == "1": + category = "lower_body" + elif category == "2": + category = "dresses" + people_path = os.path.join(self.dataset_dir, category, "images", people) + clothes_path = os.path.join(self.dataset_dir, category, "images", clothes) + self.people_list.append(people_path) + self.clothes_list.append(clothes_path) + + + def __len__(self): + return len(self.people_list) + + def __getitem__(self, index): + people_path = self.people_list[index] + # /home/sd/zjh/Dataset/DressCode/upper_body/images/000000_0.jpg + clothes_path = self.clothes_list[index] + # /home/sd/zjh/Dataset/DressCode/upper_body/images/000000_1.jpg + dense_path = people_path.replace("images", "dense")[:-5] + "5_uv.npz" + # /home/sd/zjh/Dataset/DressCode/upper_body/dense/000000_5_uv.npz + mask_path = people_path.replace("images", "mask")[:-3] + "png" + # /home/sd/Harddisk/zjh/DressCode/upper_body/mask/000000_0.png + + # 加载图像 + img = Image.open(people_path).convert("RGB").resize((512, 512)) + img = torchvision.transforms.ToTensor()(img) + refernce = Image.open(clothes_path).convert("RGB").resize((224, 224)) + refernce = torchvision.transforms.ToTensor()(refernce) + mask = Image.open(mask_path).convert("L").resize((512, 512)) + mask = torchvision.transforms.ToTensor()(mask) + mask = 1-mask + densepose = np.load(dense_path) + densepose = torch.from_numpy(densepose['uv']) + densepose = torch.nn.functional.interpolate(densepose.unsqueeze(0), size=(512, 512), mode='bilinear', align_corners=True).squeeze(0) + + # 正则化 + img = torchvision.transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))(img) + refernce = torchvision.transforms.Normalize((0.48145466, 0.4578275, 0.40821073), + (0.26862954, 0.26130258, 0.27577711))(refernce) + # densepose = torchvision.transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))(densepose) + + # 生成 inpaint 和 hint + inpaint = img * mask + hint = torchvision.transforms.Resize((512, 512))(refernce) + hint = torch.cat((hint,densepose),dim = 0) + + + return {"GT": img, # [3, 512, 512] + "inpaint_image": inpaint, # [3, 512, 512] + "inpaint_mask": mask, # [1, 512, 512] + "ref_imgs": refernce, # [3, 224, 224] + "hint": hint # [5, 512, 512] + } + + + \ No newline at end of file diff --git a/ldm/data/image_vitonhd.py b/ldm/data/image_vitonhd.py new file mode 100644 index 0000000000000000000000000000000000000000..69f710b87a9a52d26da2cdbc5e89b92887300eb4 --- /dev/null +++ b/ldm/data/image_vitonhd.py @@ -0,0 +1,198 @@ +import os +import json +import random +import torch +import torchvision +from torchvision import transforms +import torch.utils.data as data +import torchvision.transforms.functional as TF +from PIL import Image +from typing import Literal, Tuple,List + + +class OpenImageDataset(data.Dataset): + def __init__( + self, + state: Literal["train", "test"], + dataset_dir: str, + type: Literal["paired", "unpaired"] = "paired", + ): + self.state=state + self.dataset_dir = dataset_dir + self.flip_transform = transforms.RandomHorizontalFlip(p=1) + + with open( + os.path.join(dataset_dir, state, "vitonhd_" + state + "_tagged.json"), "r" + ) as file1: + data1 = json.load(file1) + + annotation_list = [ + # "colors", + # "textures", + "sleeveLength", + "neckLine", + "item", + ] + + self.annotations_pair = {} + for k, v in data1.items(): + for elem in v: + annotation_str = "" + for template in annotation_list: + for tag in elem["tag_info"]: + if ( + tag["tag_name"] == template + and tag["tag_category"] is not None + ): + annotation_str += tag["tag_category"] + annotation_str += " " + self.annotations_pair[elem["file_name"]] = annotation_str + + + im_names = [] + c_names = [] + + if state == "train": + filename = os.path.join(dataset_dir, f"{state}_pairs.txt") + else: + filename = os.path.join(dataset_dir, f"{state}_pairs.txt") + + with open(filename, "r") as f: + for line in f.readlines(): + if state == "train": + im_name, _ = line.strip().split() + c_name = im_name + else: + if type == "paired": + im_name, _ = line.strip().split() + c_name = im_name + else: + im_name, c_name = line.strip().split() + + im_names.append(im_name) + c_names.append(c_name) + + self.im_names = im_names + self.c_names = c_names + + def __len__(self): + return len(self.im_names) + + def __getitem__(self, index): + c_name = self.c_names[index] + im_name = self.im_names[index] + + if c_name in self.annotations_pair: + cloth_annotation = self.annotations_pair[c_name] + else: + cloth_annotation = "shirts" + + # 确定路径 + img_path = os.path.join(self.dataset_dir, self.state, "image", im_name) + reference_path = os.path.join(self.dataset_dir, self.state, "cloth", c_name) + mask_path = os.path.join(self.dataset_dir, self.state, "agnostic-mask", im_name[:-4]+"_mask.png") + densepose_path = os.path.join(self.dataset_dir, self.state, "image-densepose", im_name) + + # 加载图像 + img = Image.open(img_path).convert("RGB").resize((512, 512)) + img = torchvision.transforms.ToTensor()(img) + reference = Image.open(reference_path).convert("RGB").resize((224, 224)) + reference = torchvision.transforms.ToTensor()(reference) + mask = Image.open(mask_path).convert("L").resize((512, 512)) + mask = torchvision.transforms.ToTensor()(mask) + mask = 1-mask + densepose = Image.open(densepose_path).convert("RGB").resize((512, 512)) + densepose = torchvision.transforms.ToTensor()(densepose) + + + #Data augmentation for training phase + if self.state == "train": + #Random horizontal flip + if random.random() > 0.5: + img = self.flip_transform(img) + mask = self.flip_transform(mask) + densepose = self.flip_transform(densepose) + reference = self.flip_transform(reference) + + #Color jittering + if random.random() > 0.5: + color_jitter = transforms.ColorJitter(brightness=0.5, contrast=0.3, saturation=0.5, hue=0.5) + fn_idx, b, c, s, h = transforms.ColorJitter.get_params(color_jitter.brightness, color_jitter.contrast, color_jitter.saturation, color_jitter.hue) + + img = TF.adjust_contrast(img, c) + img = TF.adjust_brightness(img, b) + img = TF.adjust_hue(img, h) + img = TF.adjust_saturation(img, s) + + reference = TF.adjust_contrast(reference, c) + reference = TF.adjust_brightness(reference, b) + reference = TF.adjust_hue(reference, h) + reference = TF.adjust_saturation(reference, s) + + #Scaling and shifting + if random.random() > 0.5: + scale_val = random.uniform(0.8, 1.2) + img = transforms.functional.affine( + img, angle=0, translate=[0, 0], scale=scale_val, shear=0 + ) + mask = transforms.functional.affine( + mask, angle=0, translate=[0, 0], scale=scale_val, shear=0 + ) + densepose = transforms.functional.affine( + densepose, angle=0, translate=[0, 0], scale=scale_val, shear=0 + ) + + if random.random() > 0.5: + shift_valx = random.uniform(-0.2, 0.2) + shift_valy = random.uniform(-0.2, 0.2) + img = transforms.functional.affine( + img, + angle=0, + translate=[shift_valx * img.shape[-1], shift_valy * img.shape[-2]], + scale=1, + shear=0 + ) + mask = transforms.functional.affine( + mask, + angle=0, + translate=[shift_valx * mask.shape[-1], shift_valy * mask.shape[-2]], + scale=1, + shear=0 + ) + densepose = transforms.functional.affine( + densepose, + angle=0, + translate=[ + shift_valx * densepose.shape[-1], + shift_valy * densepose.shape[-2] + ], + scale=1, + shear=0 + ) + + + + # 正则化 + img = torchvision.transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))(img) + reference = torchvision.transforms.Normalize((0.48145466, 0.4578275, 0.40821073), + (0.26862954, 0.26130258, 0.27577711))(reference) + densepose = torchvision.transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))(densepose) + + # 生成 inpaint 和 hint + inpaint = img * mask + hint = torchvision.transforms.Resize((512, 512))(reference) + hint = torch.cat((hint,densepose),dim = 0) + + cloth_annotation = "a photo of " + cloth_annotation + + return {"GT": img, # [3, 512, 512] + "inpaint_image": inpaint, # [3, 512, 512] + "inpaint_mask": mask, # [1, 512, 512] + "ref_imgs": reference, # [3, 224, 224] + "hint": hint, # [6, 512, 512] + "caption_cloth": cloth_annotation, + # "caption": "model is wearing " + cloth_annotation, + } + + + diff --git a/ldm/data/imagenet.py b/ldm/data/imagenet.py new file mode 100644 index 0000000000000000000000000000000000000000..56d9fe26245117c52a603bf30e30f4d2d6687f60 --- /dev/null +++ b/ldm/data/imagenet.py @@ -0,0 +1,394 @@ +import os, yaml, pickle, shutil, tarfile, glob +import cv2 +import albumentations +import PIL +import numpy as np +import torchvision.transforms.functional as TF +from omegaconf import OmegaConf +from functools import partial +from PIL import Image +from tqdm import tqdm +from torch.utils.data import Dataset, Subset + +import taming.data.utils as tdu +from taming.data.imagenet import str_to_indices, give_synsets_from_indices, download, retrieve +from taming.data.imagenet import ImagePaths + +from ldm.modules.image_degradation import degradation_fn_bsr, degradation_fn_bsr_light + + +def synset2idx(path_to_yaml="data/index_synset.yaml"): + with open(path_to_yaml) as f: + di2s = yaml.load(f) + return dict((v,k) for k,v in di2s.items()) + + +class ImageNetBase(Dataset): + def __init__(self, config=None): + self.config = config or OmegaConf.create() + if not type(self.config)==dict: + self.config = OmegaConf.to_container(self.config) + self.keep_orig_class_label = self.config.get("keep_orig_class_label", False) + self.process_images = True # if False we skip loading & processing images and self.data contains filepaths + self._prepare() + self._prepare_synset_to_human() + self._prepare_idx_to_synset() + self._prepare_human_to_integer_label() + self._load() + + def __len__(self): + return len(self.data) + + def __getitem__(self, i): + return self.data[i] + + def _prepare(self): + raise NotImplementedError() + + def _filter_relpaths(self, relpaths): + ignore = set([ + "n06596364_9591.JPEG", + ]) + relpaths = [rpath for rpath in relpaths if not rpath.split("/")[-1] in ignore] + if "sub_indices" in self.config: + indices = str_to_indices(self.config["sub_indices"]) + synsets = give_synsets_from_indices(indices, path_to_yaml=self.idx2syn) # returns a list of strings + self.synset2idx = synset2idx(path_to_yaml=self.idx2syn) + files = [] + for rpath in relpaths: + syn = rpath.split("/")[0] + if syn in synsets: + files.append(rpath) + return files + else: + return relpaths + + def _prepare_synset_to_human(self): + SIZE = 2655750 + URL = "https://heibox.uni-heidelberg.de/f/9f28e956cd304264bb82/?dl=1" + self.human_dict = os.path.join(self.root, "synset_human.txt") + if (not os.path.exists(self.human_dict) or + not os.path.getsize(self.human_dict)==SIZE): + download(URL, self.human_dict) + + def _prepare_idx_to_synset(self): + URL = "https://heibox.uni-heidelberg.de/f/d835d5b6ceda4d3aa910/?dl=1" + self.idx2syn = os.path.join(self.root, "index_synset.yaml") + if (not os.path.exists(self.idx2syn)): + download(URL, self.idx2syn) + + def _prepare_human_to_integer_label(self): + URL = "https://heibox.uni-heidelberg.de/f/2362b797d5be43b883f6/?dl=1" + self.human2integer = os.path.join(self.root, "imagenet1000_clsidx_to_labels.txt") + if (not os.path.exists(self.human2integer)): + download(URL, self.human2integer) + with open(self.human2integer, "r") as f: + lines = f.read().splitlines() + assert len(lines) == 1000 + self.human2integer_dict = dict() + for line in lines: + value, key = line.split(":") + self.human2integer_dict[key] = int(value) + + def _load(self): + with open(self.txt_filelist, "r") as f: + self.relpaths = f.read().splitlines() + l1 = len(self.relpaths) + self.relpaths = self._filter_relpaths(self.relpaths) + print("Removed {} files from filelist during filtering.".format(l1 - len(self.relpaths))) + + self.synsets = [p.split("/")[0] for p in self.relpaths] + self.abspaths = [os.path.join(self.datadir, p) for p in self.relpaths] + + unique_synsets = np.unique(self.synsets) + class_dict = dict((synset, i) for i, synset in enumerate(unique_synsets)) + if not self.keep_orig_class_label: + self.class_labels = [class_dict[s] for s in self.synsets] + else: + self.class_labels = [self.synset2idx[s] for s in self.synsets] + + with open(self.human_dict, "r") as f: + human_dict = f.read().splitlines() + human_dict = dict(line.split(maxsplit=1) for line in human_dict) + + self.human_labels = [human_dict[s] for s in self.synsets] + + labels = { + "relpath": np.array(self.relpaths), + "synsets": np.array(self.synsets), + "class_label": np.array(self.class_labels), + "human_label": np.array(self.human_labels), + } + + if self.process_images: + self.size = retrieve(self.config, "size", default=256) + self.data = ImagePaths(self.abspaths, + labels=labels, + size=self.size, + random_crop=self.random_crop, + ) + else: + self.data = self.abspaths + + +class ImageNetTrain(ImageNetBase): + NAME = "ILSVRC2012_train" + URL = "http://www.image-net.org/challenges/LSVRC/2012/" + AT_HASH = "a306397ccf9c2ead27155983c254227c0fd938e2" + FILES = [ + "ILSVRC2012_img_train.tar", + ] + SIZES = [ + 147897477120, + ] + + def __init__(self, process_images=True, data_root=None, **kwargs): + self.process_images = process_images + self.data_root = data_root + super().__init__(**kwargs) + + def _prepare(self): + if self.data_root: + self.root = os.path.join(self.data_root, self.NAME) + else: + cachedir = os.environ.get("XDG_CACHE_HOME", os.path.expanduser("~/.cache")) + self.root = os.path.join(cachedir, "autoencoders/data", self.NAME) + + self.datadir = os.path.join(self.root, "data") + self.txt_filelist = os.path.join(self.root, "filelist.txt") + self.expected_length = 1281167 + self.random_crop = retrieve(self.config, "ImageNetTrain/random_crop", + default=True) + if not tdu.is_prepared(self.root): + # prep + print("Preparing dataset {} in {}".format(self.NAME, self.root)) + + datadir = self.datadir + if not os.path.exists(datadir): + path = os.path.join(self.root, self.FILES[0]) + if not os.path.exists(path) or not os.path.getsize(path)==self.SIZES[0]: + import academictorrents as at + atpath = at.get(self.AT_HASH, datastore=self.root) + assert atpath == path + + print("Extracting {} to {}".format(path, datadir)) + os.makedirs(datadir, exist_ok=True) + with tarfile.open(path, "r:") as tar: + tar.extractall(path=datadir) + + print("Extracting sub-tars.") + subpaths = sorted(glob.glob(os.path.join(datadir, "*.tar"))) + for subpath in tqdm(subpaths): + subdir = subpath[:-len(".tar")] + os.makedirs(subdir, exist_ok=True) + with tarfile.open(subpath, "r:") as tar: + tar.extractall(path=subdir) + + filelist = glob.glob(os.path.join(datadir, "**", "*.JPEG")) + filelist = [os.path.relpath(p, start=datadir) for p in filelist] + filelist = sorted(filelist) + filelist = "\n".join(filelist)+"\n" + with open(self.txt_filelist, "w") as f: + f.write(filelist) + + tdu.mark_prepared(self.root) + + +class ImageNetValidation(ImageNetBase): + NAME = "ILSVRC2012_validation" + URL = "http://www.image-net.org/challenges/LSVRC/2012/" + AT_HASH = "5d6d0df7ed81efd49ca99ea4737e0ae5e3a5f2e5" + VS_URL = "https://heibox.uni-heidelberg.de/f/3e0f6e9c624e45f2bd73/?dl=1" + FILES = [ + "ILSVRC2012_img_val.tar", + "validation_synset.txt", + ] + SIZES = [ + 6744924160, + 1950000, + ] + + def __init__(self, process_images=True, data_root=None, **kwargs): + self.data_root = data_root + self.process_images = process_images + super().__init__(**kwargs) + + def _prepare(self): + if self.data_root: + self.root = os.path.join(self.data_root, self.NAME) + else: + cachedir = os.environ.get("XDG_CACHE_HOME", os.path.expanduser("~/.cache")) + self.root = os.path.join(cachedir, "autoencoders/data", self.NAME) + self.datadir = os.path.join(self.root, "data") + self.txt_filelist = os.path.join(self.root, "filelist.txt") + self.expected_length = 50000 + self.random_crop = retrieve(self.config, "ImageNetValidation/random_crop", + default=False) + if not tdu.is_prepared(self.root): + # prep + print("Preparing dataset {} in {}".format(self.NAME, self.root)) + + datadir = self.datadir + if not os.path.exists(datadir): + path = os.path.join(self.root, self.FILES[0]) + if not os.path.exists(path) or not os.path.getsize(path)==self.SIZES[0]: + import academictorrents as at + atpath = at.get(self.AT_HASH, datastore=self.root) + assert atpath == path + + print("Extracting {} to {}".format(path, datadir)) + os.makedirs(datadir, exist_ok=True) + with tarfile.open(path, "r:") as tar: + tar.extractall(path=datadir) + + vspath = os.path.join(self.root, self.FILES[1]) + if not os.path.exists(vspath) or not os.path.getsize(vspath)==self.SIZES[1]: + download(self.VS_URL, vspath) + + with open(vspath, "r") as f: + synset_dict = f.read().splitlines() + synset_dict = dict(line.split() for line in synset_dict) + + print("Reorganizing into synset folders") + synsets = np.unique(list(synset_dict.values())) + for s in synsets: + os.makedirs(os.path.join(datadir, s), exist_ok=True) + for k, v in synset_dict.items(): + src = os.path.join(datadir, k) + dst = os.path.join(datadir, v) + shutil.move(src, dst) + + filelist = glob.glob(os.path.join(datadir, "**", "*.JPEG")) + filelist = [os.path.relpath(p, start=datadir) for p in filelist] + filelist = sorted(filelist) + filelist = "\n".join(filelist)+"\n" + with open(self.txt_filelist, "w") as f: + f.write(filelist) + + tdu.mark_prepared(self.root) + + + +class ImageNetSR(Dataset): + def __init__(self, size=None, + degradation=None, downscale_f=4, min_crop_f=0.5, max_crop_f=1., + random_crop=True): + """ + Imagenet Superresolution Dataloader + Performs following ops in order: + 1. crops a crop of size s from image either as random or center crop + 2. resizes crop to size with cv2.area_interpolation + 3. degrades resized crop with degradation_fn + + :param size: resizing to size after cropping + :param degradation: degradation_fn, e.g. cv_bicubic or bsrgan_light + :param downscale_f: Low Resolution Downsample factor + :param min_crop_f: determines crop size s, + where s = c * min_img_side_len with c sampled from interval (min_crop_f, max_crop_f) + :param max_crop_f: "" + :param data_root: + :param random_crop: + """ + self.base = self.get_base() + assert size + assert (size / downscale_f).is_integer() + self.size = size + self.LR_size = int(size / downscale_f) + self.min_crop_f = min_crop_f + self.max_crop_f = max_crop_f + assert(max_crop_f <= 1.) + self.center_crop = not random_crop + + self.image_rescaler = albumentations.SmallestMaxSize(max_size=size, interpolation=cv2.INTER_AREA) + + self.pil_interpolation = False # gets reset later if incase interp_op is from pillow + + if degradation == "bsrgan": + self.degradation_process = partial(degradation_fn_bsr, sf=downscale_f) + + elif degradation == "bsrgan_light": + self.degradation_process = partial(degradation_fn_bsr_light, sf=downscale_f) + + else: + interpolation_fn = { + "cv_nearest": cv2.INTER_NEAREST, + "cv_bilinear": cv2.INTER_LINEAR, + "cv_bicubic": cv2.INTER_CUBIC, + "cv_area": cv2.INTER_AREA, + "cv_lanczos": cv2.INTER_LANCZOS4, + "pil_nearest": PIL.Image.NEAREST, + "pil_bilinear": PIL.Image.BILINEAR, + "pil_bicubic": PIL.Image.BICUBIC, + "pil_box": PIL.Image.BOX, + "pil_hamming": PIL.Image.HAMMING, + "pil_lanczos": PIL.Image.LANCZOS, + }[degradation] + + self.pil_interpolation = degradation.startswith("pil_") + + if self.pil_interpolation: + self.degradation_process = partial(TF.resize, size=self.LR_size, interpolation=interpolation_fn) + + else: + self.degradation_process = albumentations.SmallestMaxSize(max_size=self.LR_size, + interpolation=interpolation_fn) + + def __len__(self): + return len(self.base) + + def __getitem__(self, i): + example = self.base[i] + image = Image.open(example["file_path_"]) + + if not image.mode == "RGB": + image = image.convert("RGB") + + image = np.array(image).astype(np.uint8) + + min_side_len = min(image.shape[:2]) + crop_side_len = min_side_len * np.random.uniform(self.min_crop_f, self.max_crop_f, size=None) + crop_side_len = int(crop_side_len) + + if self.center_crop: + self.cropper = albumentations.CenterCrop(height=crop_side_len, width=crop_side_len) + + else: + self.cropper = albumentations.RandomCrop(height=crop_side_len, width=crop_side_len) + + image = self.cropper(image=image)["image"] + image = self.image_rescaler(image=image)["image"] + + if self.pil_interpolation: + image_pil = PIL.Image.fromarray(image) + LR_image = self.degradation_process(image_pil) + LR_image = np.array(LR_image).astype(np.uint8) + + else: + LR_image = self.degradation_process(image=image)["image"] + + example["image"] = (image/127.5 - 1.0).astype(np.float32) + example["LR_image"] = (LR_image/127.5 - 1.0).astype(np.float32) + + return example + + +class ImageNetSRTrain(ImageNetSR): + def __init__(self, **kwargs): + super().__init__(**kwargs) + + def get_base(self): + with open("data/imagenet_train_hr_indices.p", "rb") as f: + indices = pickle.load(f) + dset = ImageNetTrain(process_images=False,) + return Subset(dset, indices) + + +class ImageNetSRValidation(ImageNetSR): + def __init__(self, **kwargs): + super().__init__(**kwargs) + + def get_base(self): + with open("data/imagenet_val_hr_indices.p", "rb") as f: + indices = pickle.load(f) + dset = ImageNetValidation(process_images=False,) + return Subset(dset, indices) diff --git a/ldm/data/lsun.py b/ldm/data/lsun.py new file mode 100644 index 0000000000000000000000000000000000000000..593bedc6e6fccf1d096f589315c23cacd6b97e62 --- /dev/null +++ b/ldm/data/lsun.py @@ -0,0 +1,92 @@ +import os +import numpy as np +import PIL +from PIL import Image +from torch.utils.data import Dataset +from torchvision import transforms + + +class LSUNBase(Dataset): + def __init__(self, + txt_file, + data_root, + size=None, + interpolation="bicubic", + flip_p=0.5 + ): + self.data_paths = txt_file + self.data_root = data_root + with open(self.data_paths, "r") as f: + self.image_paths = f.read().splitlines() + self._length = len(self.image_paths) + self.labels = { + "relative_file_path_": [l for l in self.image_paths], + "file_path_": [os.path.join(self.data_root, l) + for l in self.image_paths], + } + + self.size = size + self.interpolation = {"linear": PIL.Image.LINEAR, + "bilinear": PIL.Image.BILINEAR, + "bicubic": PIL.Image.BICUBIC, + "lanczos": PIL.Image.LANCZOS, + }[interpolation] + self.flip = transforms.RandomHorizontalFlip(p=flip_p) + + def __len__(self): + return self._length + + def __getitem__(self, i): + example = dict((k, self.labels[k][i]) for k in self.labels) + image = Image.open(example["file_path_"]) + if not image.mode == "RGB": + image = image.convert("RGB") + + # default to score-sde preprocessing + img = np.array(image).astype(np.uint8) + crop = min(img.shape[0], img.shape[1]) + h, w, = img.shape[0], img.shape[1] + img = img[(h - crop) // 2:(h + crop) // 2, + (w - crop) // 2:(w + crop) // 2] + + image = Image.fromarray(img) + if self.size is not None: + image = image.resize((self.size, self.size), resample=self.interpolation) + + image = self.flip(image) + image = np.array(image).astype(np.uint8) + example["image"] = (image / 127.5 - 1.0).astype(np.float32) + return example + + +class LSUNChurchesTrain(LSUNBase): + def __init__(self, **kwargs): + super().__init__(txt_file="data/lsun/church_outdoor_train.txt", data_root="data/lsun/churches", **kwargs) + + +class LSUNChurchesValidation(LSUNBase): + def __init__(self, flip_p=0., **kwargs): + super().__init__(txt_file="data/lsun/church_outdoor_val.txt", data_root="data/lsun/churches", + flip_p=flip_p, **kwargs) + + +class LSUNBedroomsTrain(LSUNBase): + def __init__(self, **kwargs): + super().__init__(txt_file="data/lsun/bedrooms_train.txt", data_root="data/lsun/bedrooms", **kwargs) + + +class LSUNBedroomsValidation(LSUNBase): + def __init__(self, flip_p=0.0, **kwargs): + super().__init__(txt_file="data/lsun/bedrooms_val.txt", data_root="data/lsun/bedrooms", + flip_p=flip_p, **kwargs) + + +class LSUNCatsTrain(LSUNBase): + def __init__(self, **kwargs): + super().__init__(txt_file="data/lsun/cat_train.txt", data_root="data/lsun/cats", **kwargs) + + +class LSUNCatsValidation(LSUNBase): + def __init__(self, flip_p=0., **kwargs): + super().__init__(txt_file="data/lsun/cat_val.txt", data_root="data/lsun/cats", + flip_p=flip_p, **kwargs) diff --git a/ldm/lr_scheduler.py b/ldm/lr_scheduler.py new file mode 100644 index 0000000000000000000000000000000000000000..9c832cdaa38f2bb63d414ac7ebf3979cd700246d --- /dev/null +++ b/ldm/lr_scheduler.py @@ -0,0 +1,81 @@ +import numpy as np + + +class LambdaWarmUpCosineScheduler: + """ + note: use with a base_lr of 1.0 + """ + def __init__(self, warm_up_steps, lr_min, lr_max, lr_start, max_decay_steps, verbosity_interval=0): + self.lr_warm_up_steps = warm_up_steps + self.lr_start = lr_start + self.lr_min = lr_min + self.lr_max = lr_max + self.lr_max_decay_steps = max_decay_steps + self.last_lr = 0. + self.verbosity_interval = verbosity_interval + + def schedule(self, n, **kwargs): + if self.verbosity_interval > 0: + if n % self.verbosity_interval == 0: print(f"current step: {n}, recent lr-multiplier: {self.last_lr}") + if n < self.lr_warm_up_steps: + lr = (self.lr_max - self.lr_start) / self.lr_warm_up_steps * n + self.lr_start + self.last_lr = lr + return lr + else: + t = (n - self.lr_warm_up_steps) / (self.lr_max_decay_steps - self.lr_warm_up_steps) + t = min(t, 1.0) + lr = self.lr_min + 0.5 * (self.lr_max - self.lr_min) * ( + 1 + np.cos(t * np.pi)) + self.last_lr = lr + return lr + + def __call__(self, n, **kwargs): + return self.schedule(n,**kwargs) + + +class LambdaWarmUpCosineScheduler2: + """ + supports repeated iterations, configurable via lists + note: use with a base_lr of 1.0. + """ + def __init__(self, warm_up_steps, f_min, f_max, f_start, cycle_lengths, verbosity_interval=0): + assert len(warm_up_steps) == len(f_min) == len(f_max) == len(f_start) == len(cycle_lengths) + self.lr_warm_up_steps = warm_up_steps # [10000] + self.f_start = f_start # 1.e-6 + self.f_min = f_min # 1. + self.f_max = f_max # 1. + self.cycle_lengths = cycle_lengths # [10000000000000] + self.cum_cycles = np.cumsum([0] + list(self.cycle_lengths)) + self.last_f = 0. + self.verbosity_interval = verbosity_interval # 0 + + def find_in_interval(self, n): + interval = 0 + for cl in self.cum_cycles[1:]: + if n <= cl: + return interval + interval += 1 + + + def __call__(self, n, **kwargs): + return self.schedule(n, **kwargs) + + +class LambdaLinearScheduler(LambdaWarmUpCosineScheduler2): + + def schedule(self, n, **kwargs): + cycle = self.find_in_interval(n) + n = n - self.cum_cycles[cycle] + if self.verbosity_interval > 0: + if n % self.verbosity_interval == 0: print(f"current step: {n}, recent lr-multiplier: {self.last_f}, " + f"current cycle {cycle}") + + if n < self.lr_warm_up_steps[cycle]: + f = (self.f_max[cycle] - self.f_start[cycle]) / self.lr_warm_up_steps[cycle] * n + self.f_start[cycle] + self.last_f = f + return f + else: + f = self.f_min[cycle] + (self.f_max[cycle] - self.f_min[cycle]) * (self.cycle_lengths[cycle] - n) / (self.cycle_lengths[cycle]) + self.last_f = f + return f + diff --git a/ldm/models/.DS_Store b/ldm/models/.DS_Store new file mode 100644 index 0000000000000000000000000000000000000000..5008ddfcf53c02e82d7eee2e57c38e5672ef89f6 Binary files /dev/null and b/ldm/models/.DS_Store differ diff --git a/ldm/models/__pycache__/autoencoder.cpython-38.pyc b/ldm/models/__pycache__/autoencoder.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..fe6c7fe280bf5e6130eb199cb851b4da2122cf28 Binary files /dev/null and b/ldm/models/__pycache__/autoencoder.cpython-38.pyc differ diff --git a/ldm/models/autoencoder.py b/ldm/models/autoencoder.py new file mode 100644 index 0000000000000000000000000000000000000000..7eab8f00588c5d822040dcc5c26b3dd923ae66c1 --- /dev/null +++ b/ldm/models/autoencoder.py @@ -0,0 +1,408 @@ +import torch +import pytorch_lightning as pl +import torch.nn.functional as F +from contextlib import contextmanager + +from taming.modules.vqvae.quantize import VectorQuantizer2 as VectorQuantizer + +from ldm.modules.diffusionmodules.model import Encoder, Decoder +from ldm.modules.distributions.distributions import DiagonalGaussianDistribution + +from ldm.util import instantiate_from_config + + +class VQModel(pl.LightningModule): + def __init__(self, + ddconfig, + lossconfig, + n_embed, + embed_dim, + ckpt_path=None, + ignore_keys=[], + image_key="image", + colorize_nlabels=None, + monitor=None, + batch_resize_range=None, + scheduler_config=None, + lr_g_factor=1.0, + remap=None, + sane_index_shape=False, # tell vector quantizer to return indices as bhw + use_ema=False + ): + super().__init__() + self.embed_dim = embed_dim + self.n_embed = n_embed + self.image_key = image_key + self.encoder = Encoder(**ddconfig) + self.decoder = Decoder(**ddconfig) + self.loss = instantiate_from_config(lossconfig) + self.quantize = VectorQuantizer(n_embed, embed_dim, beta=0.25, + remap=remap, + sane_index_shape=sane_index_shape) + self.quant_conv = torch.nn.Conv2d(ddconfig["z_channels"], embed_dim, 1) + self.post_quant_conv = torch.nn.Conv2d(embed_dim, ddconfig["z_channels"], 1) + if colorize_nlabels is not None: + assert type(colorize_nlabels)==int + self.register_buffer("colorize", torch.randn(3, colorize_nlabels, 1, 1)) + if monitor is not None: + self.monitor = monitor + self.batch_resize_range = batch_resize_range + if self.batch_resize_range is not None: + print(f"{self.__class__.__name__}: Using per-batch resizing in range {batch_resize_range}.") + + self.use_ema = use_ema + if self.use_ema: + self.model_ema = LitEma(self) + print(f"Keeping EMAs of {len(list(self.model_ema.buffers()))}.") + + if ckpt_path is not None: + self.init_from_ckpt(ckpt_path, ignore_keys=ignore_keys) + self.scheduler_config = scheduler_config + self.lr_g_factor = lr_g_factor + + @contextmanager + def ema_scope(self, context=None): + if self.use_ema: + self.model_ema.store(self.parameters()) + self.model_ema.copy_to(self) + if context is not None: + print(f"{context}: Switched to EMA weights") + try: + yield None + finally: + if self.use_ema: + self.model_ema.restore(self.parameters()) + if context is not None: + print(f"{context}: Restored training weights") + + def init_from_ckpt(self, path, ignore_keys=list()): + sd = torch.load(path, map_location="cpu")["state_dict"] + keys = list(sd.keys()) + for k in keys: + for ik in ignore_keys: + if k.startswith(ik): + print("Deleting key {} from state_dict.".format(k)) + del sd[k] + missing, unexpected = self.load_state_dict(sd, strict=False) + print(f"Restored from {path} with {len(missing)} missing and {len(unexpected)} unexpected keys") + if len(missing) > 0: + print(f"Missing Keys: {missing}") + print(f"Unexpected Keys: {unexpected}") + + def on_train_batch_end(self, *args, **kwargs): + if self.use_ema: + self.model_ema(self) + + def encode(self, x): + h = self.encoder(x) + h = self.quant_conv(h) + quant, emb_loss, info = self.quantize(h) + return quant, emb_loss, info + + def encode_to_prequant(self, x): + h = self.encoder(x) + h = self.quant_conv(h) + return h + + def decode(self, quant): + quant = self.post_quant_conv(quant) + dec = self.decoder(quant) + return dec + + def decode_code(self, code_b): + quant_b = self.quantize.embed_code(code_b) + dec = self.decode(quant_b) + return dec + + def forward(self, input, return_pred_indices=False): + quant, diff, (_,_,ind) = self.encode(input) + dec = self.decode(quant) + if return_pred_indices: + return dec, diff, ind + return dec, diff + + def get_input(self, batch, k): + x = batch[k] + if len(x.shape) == 3: + x = x[..., None] + x = x.permute(0, 3, 1, 2).to(memory_format=torch.contiguous_format).float() + if self.batch_resize_range is not None: + lower_size = self.batch_resize_range[0] + upper_size = self.batch_resize_range[1] + if self.global_step <= 4: + # do the first few batches with max size to avoid later oom + new_resize = upper_size + else: + new_resize = np.random.choice(np.arange(lower_size, upper_size+16, 16)) + if new_resize != x.shape[2]: + x = F.interpolate(x, size=new_resize, mode="bicubic") + x = x.detach() + return x + + def training_step(self, batch, batch_idx, optimizer_idx): + # https://github.com/pytorch/pytorch/issues/37142 + # try not to fool the heuristics + x = self.get_input(batch, self.image_key) + xrec, qloss, ind = self(x, return_pred_indices=True) + + if optimizer_idx == 0: + # autoencode + aeloss, log_dict_ae = self.loss(qloss, x, xrec, optimizer_idx, self.global_step, + last_layer=self.get_last_layer(), split="train", + predicted_indices=ind) + + self.log_dict(log_dict_ae, prog_bar=False, logger=True, on_step=True, on_epoch=True) + return aeloss + + if optimizer_idx == 1: + # discriminator + discloss, log_dict_disc = self.loss(qloss, x, xrec, optimizer_idx, self.global_step, + last_layer=self.get_last_layer(), split="train") + self.log_dict(log_dict_disc, prog_bar=False, logger=True, on_step=True, on_epoch=True) + return discloss + + def validation_step(self, batch, batch_idx): + log_dict = self._validation_step(batch, batch_idx) + with self.ema_scope(): + log_dict_ema = self._validation_step(batch, batch_idx, suffix="_ema") + return log_dict + + def _validation_step(self, batch, batch_idx, suffix=""): + x = self.get_input(batch, self.image_key) + xrec, qloss, ind = self(x, return_pred_indices=True) + aeloss, log_dict_ae = self.loss(qloss, x, xrec, 0, + self.global_step, + last_layer=self.get_last_layer(), + split="val"+suffix, + predicted_indices=ind + ) + + discloss, log_dict_disc = self.loss(qloss, x, xrec, 1, + self.global_step, + last_layer=self.get_last_layer(), + split="val"+suffix, + predicted_indices=ind + ) + rec_loss = log_dict_ae[f"val{suffix}/rec_loss"] + self.log(f"val{suffix}/rec_loss", rec_loss, + prog_bar=True, logger=True, on_step=False, on_epoch=True, sync_dist=True) + self.log(f"val{suffix}/aeloss", aeloss, + prog_bar=True, logger=True, on_step=False, on_epoch=True, sync_dist=True) + if version.parse(pl.__version__) >= version.parse('1.4.0'): + del log_dict_ae[f"val{suffix}/rec_loss"] + self.log_dict(log_dict_ae) + self.log_dict(log_dict_disc) + return self.log_dict + + def configure_optimizers(self): + lr_d = self.learning_rate + lr_g = self.lr_g_factor*self.learning_rate + print("lr_d", lr_d) + print("lr_g", lr_g) + opt_ae = torch.optim.Adam(list(self.encoder.parameters())+ + list(self.decoder.parameters())+ + list(self.quantize.parameters())+ + list(self.quant_conv.parameters())+ + list(self.post_quant_conv.parameters()), + lr=lr_g, betas=(0.5, 0.9)) + opt_disc = torch.optim.Adam(self.loss.discriminator.parameters(), + lr=lr_d, betas=(0.5, 0.9)) + + if self.scheduler_config is not None: + scheduler = instantiate_from_config(self.scheduler_config) + + print("Setting up LambdaLR scheduler...") + scheduler = [ + { + 'scheduler': LambdaLR(opt_ae, lr_lambda=scheduler.schedule), + 'interval': 'step', + 'frequency': 1 + }, + { + 'scheduler': LambdaLR(opt_disc, lr_lambda=scheduler.schedule), + 'interval': 'step', + 'frequency': 1 + }, + ] + return [opt_ae, opt_disc], scheduler + return [opt_ae, opt_disc], [] + + def get_last_layer(self): + return self.decoder.conv_out.weight + + def log_images(self, batch, only_inputs=False, plot_ema=False, **kwargs): + log = dict() + x = self.get_input(batch, self.image_key) + x = x.to(self.device) + if only_inputs: + log["inputs"] = x + return log + xrec, _ = self(x) + if x.shape[1] > 3: + # colorize with random projection + assert xrec.shape[1] > 3 + x = self.to_rgb(x) + xrec = self.to_rgb(xrec) + log["inputs"] = x + log["reconstructions"] = xrec + if plot_ema: + with self.ema_scope(): + xrec_ema, _ = self(x) + if x.shape[1] > 3: xrec_ema = self.to_rgb(xrec_ema) + log["reconstructions_ema"] = xrec_ema + return log + + def to_rgb(self, x): + assert self.image_key == "segmentation" + if not hasattr(self, "colorize"): + self.register_buffer("colorize", torch.randn(3, x.shape[1], 1, 1).to(x)) + x = F.conv2d(x, weight=self.colorize) + x = 2.*(x-x.min())/(x.max()-x.min()) - 1. + return x + + +class VQModelInterface(VQModel): + def __init__(self, embed_dim, *args, **kwargs): + super().__init__(embed_dim=embed_dim, *args, **kwargs) + self.embed_dim = embed_dim + + def encode(self, x): + h = self.encoder(x) + h = self.quant_conv(h) + return h + + def decode(self, h, force_not_quantize=False): + # also go through quantization layer + if not force_not_quantize: + quant, emb_loss, info = self.quantize(h) + else: + quant = h + quant = self.post_quant_conv(quant) + dec = self.decoder(quant) + return dec + + +class AutoencoderKL(pl.LightningModule): + def __init__(self, + embed_dim, # 4 + monitor, # "val/rec_loss" + ddconfig, # {...} + lossconfig, # {target: torch.nn.Identity} + ckpt_path=None, + ignore_keys=[], + image_key="image", + colorize_nlabels=None): + super().__init__() + self.image_key = image_key # "image" + self.encoder = Encoder(**ddconfig) + self.decoder = Decoder(**ddconfig) + self.loss = instantiate_from_config(lossconfig) + self.quant_conv = torch.nn.Conv2d(2*ddconfig["z_channels"], 2*embed_dim, 1) # 8 -> 8 + self.post_quant_conv = torch.nn.Conv2d(embed_dim, ddconfig["z_channels"], 1) # 4 -> 4 + self.embed_dim = embed_dim + self.monitor = monitor + + def encode(self, x): + h = self.encoder(x) + moments = self.quant_conv(h) + posterior = DiagonalGaussianDistribution(moments) + return posterior + + def decode(self, z): + z = self.post_quant_conv(z) + dec = self.decoder(z) + return dec + + def forward(self, input, sample_posterior=True): + posterior = self.encode(input) + if sample_posterior: + z = posterior.sample() + else: + z = posterior.mode() + dec = self.decode(z) + return dec, posterior + + def get_input(self, batch, k): + x = batch[k] + if len(x.shape) == 3: + x = x[..., None] + x = x.permute(0, 3, 1, 2).to(memory_format=torch.contiguous_format).float() + return x + + def training_step(self, batch, batch_idx, optimizer_idx): + inputs = self.get_input(batch, self.image_key) + reconstructions, posterior = self(inputs) + + if optimizer_idx == 0: + # train encoder+decoder+logvar + aeloss, log_dict_ae = self.loss(inputs, reconstructions, posterior, optimizer_idx, self.global_step, + last_layer=self.get_last_layer(), split="train") + self.log("aeloss", aeloss, prog_bar=True, logger=True, on_step=True, on_epoch=True) + self.log_dict(log_dict_ae, prog_bar=False, logger=True, on_step=True, on_epoch=False) + return aeloss + + if optimizer_idx == 1: + # train the discriminator + discloss, log_dict_disc = self.loss(inputs, reconstructions, posterior, optimizer_idx, self.global_step, + last_layer=self.get_last_layer(), split="train") + + self.log("discloss", discloss, prog_bar=True, logger=True, on_step=True, on_epoch=True) + self.log_dict(log_dict_disc, prog_bar=False, logger=True, on_step=True, on_epoch=False) + return discloss + + def validation_step(self, batch, batch_idx): + inputs = self.get_input(batch, self.image_key) + reconstructions, posterior = self(inputs) + aeloss, log_dict_ae = self.loss(inputs, reconstructions, posterior, 0, self.global_step, + last_layer=self.get_last_layer(), split="val") + + discloss, log_dict_disc = self.loss(inputs, reconstructions, posterior, 1, self.global_step, + last_layer=self.get_last_layer(), split="val") + + self.log("val/rec_loss", log_dict_ae["val/rec_loss"]) + self.log_dict(log_dict_ae) + self.log_dict(log_dict_disc) + return self.log_dict + + def configure_optimizers(self): + lr = self.learning_rate + opt_ae = torch.optim.Adam(list(self.encoder.parameters())+ + list(self.decoder.parameters())+ + list(self.quant_conv.parameters())+ + list(self.post_quant_conv.parameters()), + lr=lr, betas=(0.5, 0.9)) + opt_disc = torch.optim.Adam(self.loss.discriminator.parameters(), + lr=lr, betas=(0.5, 0.9)) + return [opt_ae, opt_disc], [] + + def get_last_layer(self): + return self.decoder.conv_out.weight + + @torch.no_grad() + def log_images(self, batch, only_inputs=False, **kwargs): + log = dict() + x = self.get_input(batch, self.image_key) + x = x.to(self.device) + if not only_inputs: + xrec, posterior = self(x) + if x.shape[1] > 3: + # colorize with random projection + assert xrec.shape[1] > 3 + x = self.to_rgb(x) + xrec = self.to_rgb(xrec) + log["samples"] = self.decode(torch.randn_like(posterior.sample())) + log["reconstructions"] = xrec + log["inputs"] = x + return log + + def to_rgb(self, x): + assert self.image_key == "segmentation" + if not hasattr(self, "colorize"): + self.register_buffer("colorize", torch.randn(3, x.shape[1], 1, 1).to(x)) + x = F.conv2d(x, weight=self.colorize) + x = 2.*(x-x.min())/(x.max()-x.min()) - 1. + return x + + + + diff --git a/ldm/models/diffusion/.DS_Store b/ldm/models/diffusion/.DS_Store new file mode 100644 index 0000000000000000000000000000000000000000..5008ddfcf53c02e82d7eee2e57c38e5672ef89f6 Binary files /dev/null and b/ldm/models/diffusion/.DS_Store differ diff --git a/ldm/models/diffusion/__init__.py b/ldm/models/diffusion/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/ldm/models/diffusion/__pycache__/__init__.cpython-38.pyc b/ldm/models/diffusion/__pycache__/__init__.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..2a1c508cc7c7f9531fbc8eaef6cb032aece89f9f Binary files /dev/null and b/ldm/models/diffusion/__pycache__/__init__.cpython-38.pyc differ diff --git a/ldm/models/diffusion/__pycache__/control.cpython-38.pyc b/ldm/models/diffusion/__pycache__/control.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..0411760c407c3123378c8e4c1f573cc244ab105d Binary files /dev/null and b/ldm/models/diffusion/__pycache__/control.cpython-38.pyc differ diff --git a/ldm/models/diffusion/__pycache__/ddim.cpython-38.pyc b/ldm/models/diffusion/__pycache__/ddim.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..a2fa44b5c25a324125857cc3412f0acc701f7331 Binary files /dev/null and b/ldm/models/diffusion/__pycache__/ddim.cpython-38.pyc differ diff --git a/ldm/models/diffusion/__pycache__/ddpm.cpython-38.pyc b/ldm/models/diffusion/__pycache__/ddpm.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..effef2d0a27a421cd40011d312e5ba438d392f3a Binary files /dev/null and b/ldm/models/diffusion/__pycache__/ddpm.cpython-38.pyc differ diff --git a/ldm/models/diffusion/classifier.py b/ldm/models/diffusion/classifier.py new file mode 100644 index 0000000000000000000000000000000000000000..8194f15e70546a38c85b2152c761e21706f3327e --- /dev/null +++ b/ldm/models/diffusion/classifier.py @@ -0,0 +1,267 @@ +import os +import torch +import pytorch_lightning as pl +from omegaconf import OmegaConf +from torch.nn import functional as F +from torch.optim import AdamW +from torch.optim.lr_scheduler import LambdaLR +from copy import deepcopy +from einops import rearrange +from glob import glob +from natsort import natsorted + +from ldm.modules.diffusionmodules.openaimodel import EncoderUNetModel, UNetModel +from ldm.util import log_txt_as_img, default, ismap, instantiate_from_config + +__models__ = { + 'class_label': EncoderUNetModel, + 'segmentation': UNetModel +} + + +def disabled_train(self, mode=True): + """Overwrite model.train with this function to make sure train/eval mode + does not change anymore.""" + return self + + +class NoisyLatentImageClassifier(pl.LightningModule): + + def __init__(self, + diffusion_path, + num_classes, + ckpt_path=None, + pool='attention', + label_key=None, + diffusion_ckpt_path=None, + scheduler_config=None, + weight_decay=1.e-2, + log_steps=10, + monitor='val/loss', + *args, + **kwargs): + super().__init__(*args, **kwargs) + self.num_classes = num_classes + # get latest config of diffusion model + diffusion_config = natsorted(glob(os.path.join(diffusion_path, 'configs', '*-project.yaml')))[-1] + self.diffusion_config = OmegaConf.load(diffusion_config).model + self.diffusion_config.params.ckpt_path = diffusion_ckpt_path + self.load_diffusion() + + self.monitor = monitor + self.numd = self.diffusion_model.first_stage_model.encoder.num_resolutions - 1 + self.log_time_interval = self.diffusion_model.num_timesteps // log_steps + self.log_steps = log_steps + + self.label_key = label_key if not hasattr(self.diffusion_model, 'cond_stage_key') \ + else self.diffusion_model.cond_stage_key + + assert self.label_key is not None, 'label_key neither in diffusion model nor in model.params' + + if self.label_key not in __models__: + raise NotImplementedError() + + self.load_classifier(ckpt_path, pool) + + self.scheduler_config = scheduler_config + self.use_scheduler = self.scheduler_config is not None + self.weight_decay = weight_decay + + def init_from_ckpt(self, path, ignore_keys=list(), only_model=False): + sd = torch.load(path, map_location="cpu") + if "state_dict" in list(sd.keys()): + sd = sd["state_dict"] + keys = list(sd.keys()) + for k in keys: + for ik in ignore_keys: + if k.startswith(ik): + print("Deleting key {} from state_dict.".format(k)) + del sd[k] + missing, unexpected = self.load_state_dict(sd, strict=False) if not only_model else self.model.load_state_dict( + sd, strict=False) + print(f"Restored from {path} with {len(missing)} missing and {len(unexpected)} unexpected keys") + if len(missing) > 0: + print(f"Missing Keys: {missing}") + if len(unexpected) > 0: + print(f"Unexpected Keys: {unexpected}") + + def load_diffusion(self): + model = instantiate_from_config(self.diffusion_config) + self.diffusion_model = model.eval() + self.diffusion_model.train = disabled_train + for param in self.diffusion_model.parameters(): + param.requires_grad = False + + def load_classifier(self, ckpt_path, pool): + model_config = deepcopy(self.diffusion_config.params.unet_config.params) + model_config.in_channels = self.diffusion_config.params.unet_config.params.out_channels + model_config.out_channels = self.num_classes + if self.label_key == 'class_label': + model_config.pool = pool + + self.model = __models__[self.label_key](**model_config) + if ckpt_path is not None: + print('#####################################################################') + print(f'load from ckpt "{ckpt_path}"') + print('#####################################################################') + self.init_from_ckpt(ckpt_path) + + @torch.no_grad() + def get_x_noisy(self, x, t, noise=None): + noise = default(noise, lambda: torch.randn_like(x)) + continuous_sqrt_alpha_cumprod = None + if self.diffusion_model.use_continuous_noise: + continuous_sqrt_alpha_cumprod = self.diffusion_model.sample_continuous_noise_level(x.shape[0], t + 1) + # todo: make sure t+1 is correct here + + return self.diffusion_model.q_sample(x_start=x, t=t, noise=noise, + continuous_sqrt_alpha_cumprod=continuous_sqrt_alpha_cumprod) + + def forward(self, x_noisy, t, *args, **kwargs): + return self.model(x_noisy, t) + + @torch.no_grad() + def get_input(self, batch, k): + x = batch[k] + if len(x.shape) == 3: + x = x[..., None] + x = rearrange(x, 'b h w c -> b c h w') + x = x.to(memory_format=torch.contiguous_format).float() + return x + + @torch.no_grad() + def get_conditioning(self, batch, k=None): + if k is None: + k = self.label_key + assert k is not None, 'Needs to provide label key' + + targets = batch[k].to(self.device) + + if self.label_key == 'segmentation': + targets = rearrange(targets, 'b h w c -> b c h w') + for down in range(self.numd): + h, w = targets.shape[-2:] + targets = F.interpolate(targets, size=(h // 2, w // 2), mode='nearest') + + # targets = rearrange(targets,'b c h w -> b h w c') + + return targets + + def compute_top_k(self, logits, labels, k, reduction="mean"): + _, top_ks = torch.topk(logits, k, dim=1) + if reduction == "mean": + return (top_ks == labels[:, None]).float().sum(dim=-1).mean().item() + elif reduction == "none": + return (top_ks == labels[:, None]).float().sum(dim=-1) + + def on_train_epoch_start(self): + # save some memory + self.diffusion_model.model.to('cpu') + + @torch.no_grad() + def write_logs(self, loss, logits, targets): + log_prefix = 'train' if self.training else 'val' + log = {} + log[f"{log_prefix}/loss"] = loss.mean() + log[f"{log_prefix}/acc@1"] = self.compute_top_k( + logits, targets, k=1, reduction="mean" + ) + log[f"{log_prefix}/acc@5"] = self.compute_top_k( + logits, targets, k=5, reduction="mean" + ) + + self.log_dict(log, prog_bar=False, logger=True, on_step=self.training, on_epoch=True) + self.log('loss', log[f"{log_prefix}/loss"], prog_bar=True, logger=False) + self.log('global_step', self.global_step, logger=False, on_epoch=False, prog_bar=True) + lr = self.optimizers().param_groups[0]['lr'] + self.log('lr_abs', lr, on_step=True, logger=True, on_epoch=False, prog_bar=True) + + def shared_step(self, batch, t=None): + x, *_ = self.diffusion_model.get_input(batch, k=self.diffusion_model.first_stage_key) + targets = self.get_conditioning(batch) + if targets.dim() == 4: + targets = targets.argmax(dim=1) + if t is None: + t = torch.randint(0, self.diffusion_model.num_timesteps, (x.shape[0],), device=self.device).long() + else: + t = torch.full(size=(x.shape[0],), fill_value=t, device=self.device).long() + x_noisy = self.get_x_noisy(x, t) + logits = self(x_noisy, t) + + loss = F.cross_entropy(logits, targets, reduction='none') + + self.write_logs(loss.detach(), logits.detach(), targets.detach()) + + loss = loss.mean() + return loss, logits, x_noisy, targets + + def training_step(self, batch, batch_idx): + loss, *_ = self.shared_step(batch) + return loss + + def reset_noise_accs(self): + self.noisy_acc = {t: {'acc@1': [], 'acc@5': []} for t in + range(0, self.diffusion_model.num_timesteps, self.diffusion_model.log_every_t)} + + def on_validation_start(self): + self.reset_noise_accs() + + @torch.no_grad() + def validation_step(self, batch, batch_idx): + loss, *_ = self.shared_step(batch) + + for t in self.noisy_acc: + _, logits, _, targets = self.shared_step(batch, t) + self.noisy_acc[t]['acc@1'].append(self.compute_top_k(logits, targets, k=1, reduction='mean')) + self.noisy_acc[t]['acc@5'].append(self.compute_top_k(logits, targets, k=5, reduction='mean')) + + return loss + + def configure_optimizers(self): + optimizer = AdamW(self.model.parameters(), lr=self.learning_rate, weight_decay=self.weight_decay) + + if self.use_scheduler: + scheduler = instantiate_from_config(self.scheduler_config) + + print("Setting up LambdaLR scheduler...") + scheduler = [ + { + 'scheduler': LambdaLR(optimizer, lr_lambda=scheduler.schedule), + 'interval': 'step', + 'frequency': 1 + }] + return [optimizer], scheduler + + return optimizer + + @torch.no_grad() + def log_images(self, batch, N=8, *args, **kwargs): + log = dict() + x = self.get_input(batch, self.diffusion_model.first_stage_key) + log['inputs'] = x + + y = self.get_conditioning(batch) + + if self.label_key == 'class_label': + y = log_txt_as_img((x.shape[2], x.shape[3]), batch["human_label"]) + log['labels'] = y + + if ismap(y): + log['labels'] = self.diffusion_model.to_rgb(y) + + for step in range(self.log_steps): + current_time = step * self.log_time_interval + + _, logits, x_noisy, _ = self.shared_step(batch, t=current_time) + + log[f'inputs@t{current_time}'] = x_noisy + + pred = F.one_hot(logits.argmax(dim=1), num_classes=self.num_classes) + pred = rearrange(pred, 'b h w c -> b c h w') + + log[f'pred@t{current_time}'] = self.diffusion_model.to_rgb(pred) + + for key in log: + log[key] = log[key][:N] + + return log diff --git a/ldm/models/diffusion/control.py b/ldm/models/diffusion/control.py new file mode 100644 index 0000000000000000000000000000000000000000..df3dc85b6f0f7dbf3ec94debcf5597c945b917db --- /dev/null +++ b/ldm/models/diffusion/control.py @@ -0,0 +1,460 @@ +import random +import torch +import torchvision +import torch.nn as nn +import torch.nn.functional as F +from torch.optim.lr_scheduler import CosineAnnealingLR + +from einops import rearrange +from ldm.modules.diffusionmodules.util import ( + conv_nd, + linear, + zero_module, + timestep_embedding, +) +from ldm.models.diffusion.ddpm import DDPM +from ldm.modules.attention import SpatialTransformer +from ldm.modules.diffusionmodules.openaimodel import UNetModel, TimestepEmbedSequential, ResBlock, Downsample, AttentionBlock +from ldm.util import instantiate_from_config, default +from ldm.models.diffusion.ddim import DDIMSampler + +import torch +from torch.optim.optimizer import Optimizer +from torch.optim.lr_scheduler import LambdaLR + + +def disabled_train(self, mode=True): + return self + + +# ============================================================= +# 可训练部分 ControlNet +# ============================================================= +class ControlNet(nn.Module): + def __init__( + self, + in_channels, # 9 + model_channels, # 320 + hint_channels, # 20 + attention_resolutions, # [4,2,1] + num_res_blocks, # 2 + channel_mult=(1, 2, 4, 8), # [1,2,4,4] + num_head_channels=-1, # 64 + transformer_depth=1, # 1 + context_dim=None, # 768 + use_checkpoint=False, # True + dropout=0, + conv_resample=True, + dims=2, + num_heads=-1, + use_scale_shift_norm=False): + super(ControlNet, self).__init__() + self.dims = dims + self.in_channels = in_channels + self.model_channels = model_channels + self.num_res_blocks = len(channel_mult) * [num_res_blocks] + self.attention_resolutions = attention_resolutions + self.dropout = dropout + self.channel_mult = channel_mult + self.use_checkpoint = use_checkpoint + self.dtype = torch.float32 + self.num_heads = num_heads + self.num_head_channels = num_head_channels + + # time 编码器 + time_embed_dim = model_channels * 4 + self.time_embed = nn.Sequential( + linear(model_channels, time_embed_dim), + nn.SiLU(), + linear(time_embed_dim, time_embed_dim), + ) + + # input 编码器 + self.input_blocks = nn.ModuleList( + [ + TimestepEmbedSequential( + conv_nd(dims, in_channels, model_channels, 3, padding=1) + ) + ] + ) + self.zero_convs = nn.ModuleList([self.make_zero_conv(model_channels)]) + + # hint 编码器 + self.input_hint_block = TimestepEmbedSequential( + conv_nd(dims, hint_channels, 16, 3, padding=1), + nn.SiLU(), + conv_nd(dims, 16, 16, 3, padding=1), + nn.SiLU(), + conv_nd(dims, 16, 32, 3, padding=1, stride=2), + nn.SiLU(), + conv_nd(dims, 32, 32, 3, padding=1), + nn.SiLU(), + conv_nd(dims, 32, 96, 3, padding=1, stride=2), + nn.SiLU(), + conv_nd(dims, 96, 96, 3, padding=1), + nn.SiLU(), + conv_nd(dims, 96, 256, 3, padding=1, stride=2), + nn.SiLU(), + zero_module(conv_nd(dims, 256, model_channels, 3, padding=1)) + ) + + # UNet + input_block_chans = [model_channels] + ch = model_channels + ds = 1 + for level, mult in enumerate(channel_mult): + for nr in range(self.num_res_blocks[level]): + layers = [ + ResBlock( + ch, + time_embed_dim, + dropout, + out_channels=mult * model_channels, + dims=dims, + use_checkpoint=use_checkpoint, + use_scale_shift_norm=use_scale_shift_norm, + ) + ] + ch = mult * model_channels + if ds in attention_resolutions: + num_heads = ch // num_head_channels + dim_head = num_head_channels + disabled_sa = False + + layers.append( + SpatialTransformer( + ch, num_heads, dim_head, depth=transformer_depth, context_dim=context_dim) + ) + self.input_blocks.append(TimestepEmbedSequential(*layers)) + self.zero_convs.append(self.make_zero_conv(ch)) + input_block_chans.append(ch) + if level != len(channel_mult) - 1: + out_ch = ch + self.input_blocks.append( + TimestepEmbedSequential( + Downsample(ch, conv_resample, dims=dims, out_channels=out_ch) + ) + ) + ch = out_ch + input_block_chans.append(ch) + self.zero_convs.append(self.make_zero_conv(ch)) + ds *= 2 + num_heads = ch // num_head_channels + dim_head = num_head_channels + self.middle_block = TimestepEmbedSequential( + ResBlock( + ch, + time_embed_dim, + dropout, + dims=dims, + use_checkpoint=use_checkpoint, + use_scale_shift_norm=use_scale_shift_norm, + ), + SpatialTransformer(ch, num_heads, dim_head, depth=transformer_depth, context_dim=context_dim), + ResBlock( + ch, + time_embed_dim, + dropout, + dims=dims, + use_checkpoint=use_checkpoint, + use_scale_shift_norm=use_scale_shift_norm, + ), + ) + self.middle_block_out = self.make_zero_conv(ch) + + def make_zero_conv(self, channels): + return TimestepEmbedSequential(zero_module(conv_nd(self.dims, channels, channels, 1, padding=0))) + + def forward(self, x, hint, timesteps, reference_dino): + # 处理输入 + context = reference_dino + t_emb = timestep_embedding(timesteps, self.model_channels) + emb = self.time_embed(t_emb) + guided_hint = self.input_hint_block(hint, emb) + + # 预测 control + outs = [] + h = x.type(self.dtype) + for module, zero_conv in zip(self.input_blocks, self.zero_convs): + if guided_hint is not None: + h = module(h, emb, context) + h += guided_hint + guided_hint = None + else: + h = module(h, emb, context) + outs.append(zero_conv(h, emb, context)) + h = self.middle_block(h, emb, context) + outs.append(self.middle_block_out(h, emb, context)) + + return outs + +# ============================================================= +# 固定参数部分 ControlledUnetModel +# ============================================================= +class ControlledUnetModel(UNetModel): + def forward(self, x, timesteps=None, context=None, control=None): + hs = [] + + # UNet 的上半部分 + with torch.no_grad(): + t_emb = timestep_embedding(timesteps, self.model_channels) + emb = self.time_embed(t_emb) + h = x.type(self.dtype) + for module in self.input_blocks: + h = module(h, emb, context) + hs.append(h) + h = self.middle_block(h, emb, context) + + # 注入 control + if control is not None: + h += control.pop() + + # UNet 的下半部分 + for i, module in enumerate(self.output_blocks): + h = torch.cat([h, hs.pop() + control.pop()], dim=1) + h = module(h, emb, context) + + # 输出 + h = h.type(x.dtype) + h = self.out(h) + + return h + +# ============================================================= +# 主干网络 ControlLDM +# ============================================================= +class ControlLDM(DDPM): + def __init__(self, + control_stage_config, # ControlNet + first_stage_config, # AutoencoderKL + cond_stage_config, # FrozenCLIPImageEmbedder + condi_stage_config, # FrozenCLIPTextEmbedder + scale_factor=1.0, # 0.18215 + *args, **kwargs): + self.num_timesteps_cond = 1 + super().__init__(*args, **kwargs) # self.model 和 self.register_buffer + self.control_model = instantiate_from_config(control_stage_config) # self.control_model + self.instantiate_first_stage(first_stage_config) # self.first_stage_model 调用 AutoencoderKL + self.instantiate_cond_stage(cond_stage_config) # self.cond_stage_model 调用 FrozenCLIPImageEmbedder + self.instantiate_condi_stage(condi_stage_config) # self.condi_stage_model FrozenCLIPTextEmbedder + self.proj_out=nn.Linear(1024, 768) # 全连接层 + self.scale_factor = scale_factor # 0.18215 + self.learnable_vector = nn.Parameter(torch.randn((1,1,768)), requires_grad=False) + self.trainable_vector = nn.Parameter(torch.randn((1,1,768)), requires_grad=True) + + self.dinov2_vitl14 = torch.hub.load('facebookresearch/dinov2', 'dinov2_vitl14') + self.dinov2_vitl14.eval() + self.dinov2_vitl14.train = disabled_train + for param in self.dinov2_vitl14.parameters(): + param.requires_grad = False + self.linear = nn.Linear(1024, 768) + + # AutoencoderKL 不训练 + def instantiate_first_stage(self, config): + model = instantiate_from_config(config) + self.first_stage_model = model.eval() + self.first_stage_model.train = disabled_train + for param in self.first_stage_model.parameters(): + param.requires_grad = False + + # FrozenCLIPImageEmbedder 不训练 + def instantiate_cond_stage(self, config): + model = instantiate_from_config(config) + self.cond_stage_model = model.eval() + self.cond_stage_model.train = disabled_train + for param in self.cond_stage_model.parameters(): + param.requires_grad = False + + def instantiate_condi_stage(self, config): + model = instantiate_from_config(config) + self.condi_stage_model = model.eval() + self.condi_stage_model.train = disabled_train + for param in self.condi_stage_model.parameters(): + param.required_grad = False + + # 训练 + def training_step(self, batch, batch_idx): + z_new, reference, hint, cloth_annotation= self.get_input(batch) # 加载数据 + loss= self(z_new, reference, hint, cloth_annotation) # 计算损失 + self.log("loss", # 记录损失 + loss, + prog_bar=True, + logger=True, + on_step=True, + on_epoch=True) + self.log('lr_abs', # 记录学习率 + self.optimizers().param_groups[0]['lr'], + prog_bar=True, + logger=True, + on_step=True, + on_epoch=False) + return loss + + # 加载数据 + @torch.no_grad() + def get_input(self, batch): + + # 加载原始数据 + x, inpaint, mask, reference, hint, cloth_annotation = super().get_input(batch) + + # AutoencoderKL 处理真值 + encoder_posterior = self.first_stage_model.encode(x) + z = self.scale_factor * (encoder_posterior.sample()).detach() + + # AutoencoderKL 处理 inpaint + encoder_posterior_inpaint = self.first_stage_model.encode(inpaint) + z_inpaint = self.scale_factor * (encoder_posterior_inpaint.sample()).detach() + + # Resize mask + mask_resize = torchvision.transforms.Resize([z.shape[-2],z.shape[-1]])(mask) + + # 整理 z_new + z_new = torch.cat((z,z_inpaint,mask_resize),dim=1) + out = [z_new, reference, hint, cloth_annotation] + + return out + + # 计算损失 + def forward(self, z_new, reference, hint, cloth_annotation): + + # 随机时间 t + t = torch.randint(0, self.num_timesteps, (z_new.shape[0],), device=self.device).long() + + # CLIP 处理 reference + reference_clip = self.cond_stage_model.encode(reference) + reference_clip = self.proj_out(reference_clip) + + # CLIP text reference + reference_clip_text = self.condi_stage_model.encode(cloth_annotation) + + #apply CrossAttention to combine features + cross_att = CrossAttention().to('cuda') + reference_clip = cross_att(reference_clip, reference_clip_text, reference_clip_text) + + # DINO 处理 reference + dino = self.dinov2_vitl14(reference,is_training=True) + dino1 = dino["x_norm_clstoken"].unsqueeze(1) + dino2 = dino["x_norm_patchtokens"] + reference_dino = torch.cat((dino1, dino2), dim=1) + reference_dino = self.linear(reference_dino) + + # 随机加噪 + noise = torch.randn_like(z_new[:,:4,:,:]) + x_noisy = self.q_sample(x_start=z_new[:,:4,:,:], t=t, noise=noise) + x_noisy = torch.cat((x_noisy, z_new[:,4:,:,:]),dim=1) + + # 预测噪声 + if random.uniform(0, 1)<0.2: + model_output = self.apply_model(x_noisy, hint, t, reference_clip, reference_dino) + else: + model_output = self.apply_model(x_noisy, hint, t, reference_clip, reference_dino) + + # 计算损失 + loss = self.get_loss(model_output, noise, mean=False).mean([1, 2, 3]) + loss = loss.mean() + + return loss + + # 预测噪声 + def apply_model(self, x_noisy, hint, t, reference_clip, reference_dino): + + # 预测 control + control = self.control_model(x_noisy, hint, t, reference_dino) + + # 调用 PBE + model_output = self.model(x_noisy, t, reference_clip, control) + + return model_output + + # 优化器 + def configure_optimizers(self): + # 学习率设置 + lr = self.learning_rate + params = list(self.control_model.parameters())+list(self.linear.parameters()) + opt = torch.optim.AdamW(params, lr=lr) + + return opt + + # 采样 + @torch.no_grad() + def sample_log(self, batch, ddim_steps=50, ddim_eta=0.): + z_new, reference, hint, cloth_annotation = self.get_input(batch) + x, _, mask, _, _, _ = super().get_input(batch) + log = dict() + + # log["reference"] = reference + # reconstruction = 1. / self.scale_factor * z_new[:,:4,:,:] + # log["reconstruction"] = self.first_stage_model.decode(reconstruction) + log["mask"] = mask + + test_model_kwargs = {} + test_model_kwargs['inpaint_image'] = z_new[:,4:8,:,:] + test_model_kwargs['inpaint_mask'] = z_new[:,8:,:,:] + ddim_sampler = DDIMSampler(self) + shape = (self.channels, self.image_size, self.image_size) + samples, _ = ddim_sampler.sample(ddim_steps, + reference.shape[0], + shape, + hint, + reference, + verbose=False, + eta=ddim_eta, + test_model_kwargs=test_model_kwargs) + samples = 1. / self.scale_factor * samples + x_samples = self.first_stage_model.decode(samples[:,:4,:,:]) + # log["samples"] = x_samples + + x = torchvision.transforms.Resize([512, 512])(x) + reference = torchvision.transforms.Resize([512, 512])(reference) + x_samples = torchvision.transforms.Resize([512, 512])(x_samples) + log["grid"] = torch.cat((x, reference, x_samples), dim=2) + + return log + + + +# CrossAttention class applies cross-attention between two embeddings: an image embedding and a text embedding. +class CrossAttention(nn.Module): + def __init__( + self, + embed_dim: int=768, + num_heads: int=8 + ): + """ + Initializes a CrossAttention layer using multi-head attention. + + Args: + embed_dim (int): Dimensionality of the embeddings, which should match + the size of both reference_clip and reference_clip_text. + num_heads (int): Number of attention heads. Using multiple heads allows + the model to focus on different parts of the input embeddings. + """ + super(CrossAttention, self).__init__() + self.cross_attn = nn.MultiheadAttention(embed_dim=embed_dim, num_heads=num_heads, batch_first=True) + + def forward(self, query, key, value): + """ + Applies cross-attention to the query, key, and value inputs. + + Args: + query (Tensor): The query tensor (in this case, reference_clip). + Shape should be [batch_size, seq_length, embed_dim]. + key (Tensor): The key tensor (in this case, reference_clip_text). + Shape should be [batch_size, seq_length, embed_dim]. + value (Tensor): The value tensor (in this case, reference_clip_text). + Shape should be [batch_size, seq_length, embed_dim]. + + Returns: + Tensor: The attention output after combining reference_clip and + reference_clip_text through cross-attention. + Shape is [batch_size, seq_length, embed_dim]. + """ + + query = query.to('cuda') + key = key.to('cuda') + value = value.to('cuda') + + # Apply cross-attention, where `query` attends to `key` and `value`. + # `attn_output` contains the resulting embeddings, and `attn_weights` contains the attention weights. + attn_output, attn_weights = self.cross_attn(query, key, value) + return attn_output diff --git a/ldm/models/diffusion/ddim.py b/ldm/models/diffusion/ddim.py new file mode 100644 index 0000000000000000000000000000000000000000..9a2e6970fd9f460d47b626986249cae5d19c3fcc --- /dev/null +++ b/ldm/models/diffusion/ddim.py @@ -0,0 +1,265 @@ +"""SAMPLING ONLY.""" + +import torch +import numpy as np +from tqdm import tqdm +from functools import partial + +from ldm.modules.diffusionmodules.util import make_ddim_sampling_parameters, make_ddim_timesteps, noise_like, \ + extract_into_tensor + + +class DDIMSampler(object): + def __init__(self, model, schedule="linear", **kwargs): + super().__init__() + self.model = model + self.ddpm_num_timesteps = model.num_timesteps + self.schedule = schedule + + def register_buffer(self, name, attr): + if type(attr) == torch.Tensor: + if attr.device != torch.device("cuda"): + attr = attr.to(torch.device("cuda")) + setattr(self, name, attr) + + def make_schedule(self, ddim_num_steps, ddim_discretize="uniform", ddim_eta=0., verbose=True): + self.ddim_timesteps = make_ddim_timesteps(ddim_discr_method=ddim_discretize, num_ddim_timesteps=ddim_num_steps, + num_ddpm_timesteps=self.ddpm_num_timesteps,verbose=verbose) + alphas_cumprod = self.model.alphas_cumprod + assert alphas_cumprod.shape[0] == self.ddpm_num_timesteps, 'alphas have to be defined for each timestep' + to_torch = lambda x: x.clone().detach().to(torch.float32).to(self.model.device) + + self.register_buffer('betas', to_torch(self.model.betas)) + self.register_buffer('alphas_cumprod', to_torch(alphas_cumprod)) + self.register_buffer('alphas_cumprod_prev', to_torch(self.model.alphas_cumprod_prev)) + + # calculations for diffusion q(x_t | x_{t-1}) and others + self.register_buffer('sqrt_alphas_cumprod', to_torch(np.sqrt(alphas_cumprod.cpu()))) + self.register_buffer('sqrt_one_minus_alphas_cumprod', to_torch(np.sqrt(1. - alphas_cumprod.cpu()))) + self.register_buffer('log_one_minus_alphas_cumprod', to_torch(np.log(1. - alphas_cumprod.cpu()))) + self.register_buffer('sqrt_recip_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod.cpu()))) + self.register_buffer('sqrt_recipm1_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod.cpu() - 1))) + + # ddim sampling parameters + ddim_sigmas, ddim_alphas, ddim_alphas_prev = make_ddim_sampling_parameters(alphacums=alphas_cumprod.cpu(), + ddim_timesteps=self.ddim_timesteps, + eta=ddim_eta,verbose=verbose) + + self.register_buffer('ddim_sigmas', ddim_sigmas) + self.register_buffer('ddim_alphas', ddim_alphas) + self.register_buffer('ddim_alphas_prev', ddim_alphas_prev) + self.register_buffer('ddim_sqrt_one_minus_alphas', np.sqrt(1. - ddim_alphas)) + sigmas_for_original_sampling_steps = ddim_eta * torch.sqrt( + (1 - self.alphas_cumprod_prev) / (1 - self.alphas_cumprod) * ( + 1 - self.alphas_cumprod / self.alphas_cumprod_prev)) + self.register_buffer('ddim_sigmas_for_original_num_steps', sigmas_for_original_sampling_steps) + + @torch.no_grad() + def sample(self, + S, + batch_size, + shape, + pose, + conditioning=None, + callback=None, + normals_sequence=None, + img_callback=None, + quantize_x0=False, + eta=0., + mask=None, + x0=None, + temperature=1., + noise_dropout=0., + score_corrector=None, + corrector_kwargs=None, + verbose=True, + x_T=None, + log_every_t=100, + unconditional_guidance_scale=1., + unconditional_conditioning=None, + **kwargs + ): + if conditioning is not None: + if isinstance(conditioning, dict): + cbs = conditioning[list(conditioning.keys())[0]].shape[0] + if cbs != batch_size: + print(f"Warning: Got {cbs} conditionings but batch-size is {batch_size}") + else: + if conditioning.shape[0] != batch_size: + print(f"Warning: Got {conditioning.shape[0]} conditionings but batch-size is {batch_size}") + + self.make_schedule(ddim_num_steps=S, ddim_eta=eta, verbose=verbose) + # sampling + C, H, W = shape + size = (batch_size, C, H, W) + print(f'Data shape for DDIM sampling is {size}, eta {eta}') + + samples, intermediates = self.ddim_sampling(conditioning, size, pose, + callback=callback, + img_callback=img_callback, + quantize_denoised=quantize_x0, + mask=mask, x0=x0, + ddim_use_original_steps=False, + noise_dropout=noise_dropout, + temperature=temperature, + score_corrector=score_corrector, + corrector_kwargs=corrector_kwargs, + x_T=x_T, + log_every_t=log_every_t, + unconditional_guidance_scale=unconditional_guidance_scale, + unconditional_conditioning=unconditional_conditioning, + **kwargs + ) + return samples, intermediates + + @torch.no_grad() + def ddim_sampling(self, cond, shape, pose, + x_T=None, ddim_use_original_steps=False, + callback=None, timesteps=None, quantize_denoised=False, + mask=None, x0=None, img_callback=None, log_every_t=100, + temperature=1., noise_dropout=0., score_corrector=None, corrector_kwargs=None, + unconditional_guidance_scale=1., unconditional_conditioning=None,**kwargs): + device = self.model.betas.device + b = shape[0] + if x_T is None: + img = torch.randn(shape, device=device) + else: + img = x_T + + if timesteps is None: + timesteps = self.ddpm_num_timesteps if ddim_use_original_steps else self.ddim_timesteps + elif timesteps is not None and not ddim_use_original_steps: + subset_end = int(min(timesteps / self.ddim_timesteps.shape[0], 1) * self.ddim_timesteps.shape[0]) - 1 + timesteps = self.ddim_timesteps[:subset_end] + + intermediates = {'x_inter': [img], 'pred_x0': [img]} + time_range = reversed(range(0,timesteps)) if ddim_use_original_steps else np.flip(timesteps) + total_steps = timesteps if ddim_use_original_steps else timesteps.shape[0] + print(f"Running DDIM Sampling with {total_steps} timesteps") + + iterator = tqdm(time_range, desc='DDIM Sampler', total=total_steps) + + for i, step in enumerate(iterator): + index = total_steps - i - 1 + ts = torch.full((b,), step, device=device, dtype=torch.long) + + outs = self.p_sample_ddim(img, cond, ts, pose, index=index, use_original_steps=ddim_use_original_steps, + quantize_denoised=quantize_denoised, temperature=temperature, + noise_dropout=noise_dropout, score_corrector=score_corrector, + corrector_kwargs=corrector_kwargs, + unconditional_guidance_scale=unconditional_guidance_scale, + unconditional_conditioning=unconditional_conditioning,**kwargs) + img, pred_x0 = outs + if callback: callback(i) + if img_callback: img_callback(pred_x0, i) + + if index % log_every_t == 0 or index == total_steps - 1: + intermediates['x_inter'].append(img) + intermediates['pred_x0'].append(pred_x0) + + return img, intermediates + + @torch.no_grad() + def p_sample_ddim(self, x, c, t, pose, index, repeat_noise=False, use_original_steps=False, quantize_denoised=False, + temperature=1., noise_dropout=0., score_corrector=None, corrector_kwargs=None, + unconditional_guidance_scale=1., unconditional_conditioning=None,**kwargs): + b, *_, device = *x.shape, x.device + + if 'test_model_kwargs' in kwargs: + kwargs=kwargs['test_model_kwargs'] + x = torch.cat([x, kwargs['inpaint_image'], kwargs['inpaint_mask']],dim=1) + elif 'rest' in kwargs: + x = torch.cat((x, kwargs['rest']), dim=1) + else: + raise Exception("kwargs must contain either 'test_model_kwargs' or 'rest' key") + if unconditional_conditioning is None or unconditional_guidance_scale == 1.: + reference_clip = self.model.cond_stage_model.encode(c) + reference_clip= self.model.proj_out(reference_clip) + dino = self.model.dinov2_vitl14(c,is_training=True) + dino1 = dino["x_norm_clstoken"].unsqueeze(1) + dino2 = dino["x_norm_patchtokens"] + reference_dino = torch.cat((dino1, dino2), dim=1) + reference_dino = self.model.linear(reference_dino) + control = self.model.control_model(x, pose, t, reference_dino) + e_t = self.model.model(x, t, reference_clip, control) + else: + x_in = torch.cat([x] * 2) + t_in = torch.cat([t] * 2) + c_in = torch.cat([unconditional_conditioning, c]) + e_t_uncond, e_t = self.model.apply_model(x_in, t_in, c_in).chunk(2) + e_t = e_t_uncond + unconditional_guidance_scale * (e_t - e_t_uncond) + + if score_corrector is not None: + assert self.model.parameterization == "eps" + e_t = score_corrector.modify_score(self.model, e_t, x, t, c, **corrector_kwargs) + + alphas = self.model.alphas_cumprod if use_original_steps else self.ddim_alphas + alphas_prev = self.model.alphas_cumprod_prev if use_original_steps else self.ddim_alphas_prev + sqrt_one_minus_alphas = self.model.sqrt_one_minus_alphas_cumprod if use_original_steps else self.ddim_sqrt_one_minus_alphas + sigmas = self.model.ddim_sigmas_for_original_num_steps if use_original_steps else self.ddim_sigmas + # select parameters corresponding to the currently considered timestep + a_t = torch.full((b, 1, 1, 1), alphas[index], device=device) + a_prev = torch.full((b, 1, 1, 1), alphas_prev[index], device=device) + sigma_t = torch.full((b, 1, 1, 1), sigmas[index], device=device) + sqrt_one_minus_at = torch.full((b, 1, 1, 1), sqrt_one_minus_alphas[index],device=device) + + # current prediction for x_0 + if x.shape[1]!=4: + pred_x0 = (x[:,:4,:,:] - sqrt_one_minus_at * e_t) / a_t.sqrt() + else: + pred_x0 = (x - sqrt_one_minus_at * e_t) / a_t.sqrt() + + if quantize_denoised: + pred_x0, _, *_ = self.model.first_stage_model.quantize(pred_x0) + # direction pointing to x_t + dir_xt = (1. - a_prev - sigma_t**2).sqrt() * e_t + noise = sigma_t * noise_like(dir_xt.shape, device, repeat_noise) * temperature + if noise_dropout > 0.: + noise = torch.nn.functional.dropout(noise, p=noise_dropout) + x_prev = a_prev.sqrt() * pred_x0 + dir_xt + noise + + # noise + # noise = torch.randn_like(kwargs['inpaint_image']) + # x_noisy = self.model.q_sample(x_start=kwargs['inpaint_image'], t=t, noise=noise) + # x_prev = x_prev*(1-kwargs['inpaint_mask'])+x_noisy*(kwargs['inpaint_mask']) + + + return x_prev, pred_x0 + + @torch.no_grad() + def stochastic_encode(self, x0, t, use_original_steps=False, noise=None): + # fast, but does not allow for exact reconstruction + # t serves as an index to gather the correct alphas + if use_original_steps: + sqrt_alphas_cumprod = self.sqrt_alphas_cumprod + sqrt_one_minus_alphas_cumprod = self.sqrt_one_minus_alphas_cumprod + else: + sqrt_alphas_cumprod = torch.sqrt(self.ddim_alphas) + sqrt_one_minus_alphas_cumprod = self.ddim_sqrt_one_minus_alphas + + if noise is None: + noise = torch.randn_like(x0) + return (extract_into_tensor(sqrt_alphas_cumprod, t, x0.shape) * x0 + + extract_into_tensor(sqrt_one_minus_alphas_cumprod, t, x0.shape) * noise) + + @torch.no_grad() + def decode(self, x_latent, cond, t_start, unconditional_guidance_scale=1.0, unconditional_conditioning=None, + use_original_steps=False): + + timesteps = np.arange(self.ddpm_num_timesteps) if use_original_steps else self.ddim_timesteps + timesteps = timesteps[:t_start] + + time_range = np.flip(timesteps) + total_steps = timesteps.shape[0] + print(f"Running DDIM Sampling with {total_steps} timesteps") + + iterator = tqdm(time_range, desc='Decoding image', total=total_steps) + x_dec = x_latent + for i, step in enumerate(iterator): + index = total_steps - i - 1 + ts = torch.full((x_latent.shape[0],), step, device=x_latent.device, dtype=torch.long) + x_dec, _ = self.p_sample_ddim(x_dec, cond, ts, index=index, use_original_steps=use_original_steps, + unconditional_guidance_scale=unconditional_guidance_scale, + unconditional_conditioning=unconditional_conditioning) + return x_dec + diff --git a/ldm/models/diffusion/ddpm.py b/ldm/models/diffusion/ddpm.py new file mode 100644 index 0000000000000000000000000000000000000000..6cf2583c9f456be40b8c5d2c35639e87c8b39b63 --- /dev/null +++ b/ldm/models/diffusion/ddpm.py @@ -0,0 +1,144 @@ +import torch +import einops +import torch.nn as nn +import numpy as np +import pytorch_lightning as pl +from torch.optim.lr_scheduler import LambdaLR +from einops import rearrange, repeat +from functools import partial +from torchvision.utils import make_grid +from ldm.util import default, count_params, instantiate_from_config +from ldm.modules.distributions.distributions import DiagonalGaussianDistribution +from ldm.modules.diffusionmodules.util import make_beta_schedule, extract_into_tensor +from ldm.models.diffusion.ddim import DDIMSampler +from torchvision.transforms import Resize +import random + + + +def disabled_train(self, mode=True): + return self + + +class DiffusionWrapper(pl.LightningModule): + def __init__(self, unet_config): + super().__init__() + self.diffusion_model = instantiate_from_config(unet_config) + + def forward(self, x, timesteps=None, context=None, control=None): + out = self.diffusion_model(x, timesteps, context, control) + return out + + +class DDPM(pl.LightningModule): + def __init__(self, + unet_config, + linear_start=1e-4, # 0.00085 + linear_end=2e-2, # 0.0120 + log_every_t=100, # 200 + timesteps=1000, # 1000 + image_size=256, # 64 + channels=3, # 4 + u_cond_percent=0, # 0.2 + use_ema=True, # False + beta_schedule="linear", + loss_type="l2", + clip_denoised=True, + cosine_s=8e-3, + original_elbo_weight=0., + v_posterior=0., + l_simple_weight=1., + parameterization="eps", + use_positional_encodings=False, + learn_logvar=False, + logvar_init=0.): + super().__init__() + self.parameterization = parameterization + self.cond_stage_model = None + self.clip_denoised = clip_denoised + self.log_every_t = log_every_t + self.image_size = image_size + self.channels = channels + self.u_cond_percent=u_cond_percent + self.use_positional_encodings = use_positional_encodings + self.model = DiffusionWrapper(unet_config) # 调用 UNet 模型 + + self.use_ema = use_ema + self.use_scheduler = True + self.v_posterior = v_posterior + self.original_elbo_weight = original_elbo_weight + self.l_simple_weight = l_simple_weight + self.register_schedule(beta_schedule=beta_schedule, # "linear" + timesteps=timesteps, # 1000 + linear_start=linear_start, # 0.00085 + linear_end=linear_end, # 0.0120 + cosine_s=cosine_s) # 8e-3 + self.loss_type = loss_type + self.learn_logvar = learn_logvar + self.logvar = torch.full(fill_value=logvar_init, size=(self.num_timesteps,)) + + def register_schedule(self, + beta_schedule="linear", + timesteps=1000, + linear_start=0.00085, + linear_end=0.0120, + cosine_s=8e-3): + betas = make_beta_schedule(beta_schedule, + timesteps, + linear_start=linear_start, + linear_end=linear_end, + cosine_s=cosine_s) + alphas = 1. - betas + alphas_cumprod = np.cumprod(alphas, axis=0) + alphas_cumprod_prev = np.append(1., alphas_cumprod[:-1]) + + timesteps, = betas.shape + self.num_timesteps = int(timesteps) + self.linear_start = linear_start + self.linear_end = linear_end + to_torch = partial(torch.tensor, dtype=torch.float32) + self.register_buffer('betas', to_torch(betas)) + self.register_buffer('alphas_cumprod', to_torch(alphas_cumprod)) + self.register_buffer('alphas_cumprod_prev', to_torch(alphas_cumprod_prev)) + self.register_buffer('sqrt_alphas_cumprod', to_torch(np.sqrt(alphas_cumprod))) + self.register_buffer('sqrt_one_minus_alphas_cumprod', to_torch(np.sqrt(1. - alphas_cumprod))) + self.register_buffer('log_one_minus_alphas_cumprod', to_torch(np.log(1. - alphas_cumprod))) + self.register_buffer('sqrt_recip_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod))) + self.register_buffer('sqrt_recipm1_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod - 1))) + posterior_variance = (1 - self.v_posterior) * betas * (1. - alphas_cumprod_prev) / (1. - alphas_cumprod) + self.v_posterior * betas + self.register_buffer('posterior_variance', to_torch(posterior_variance)) + self.register_buffer('posterior_log_variance_clipped', to_torch(np.log(np.maximum(posterior_variance, 1e-20)))) + self.register_buffer('posterior_mean_coef1', to_torch(betas * np.sqrt(alphas_cumprod_prev) / (1. - alphas_cumprod))) + self.register_buffer('posterior_mean_coef2', to_torch((1. - alphas_cumprod_prev) * np.sqrt(alphas) / (1. - alphas_cumprod))) + lvlb_weights = self.betas ** 2 / (2 * self.posterior_variance * to_torch(alphas) * (1 - self.alphas_cumprod)) + lvlb_weights[0] = lvlb_weights[1] + self.register_buffer('lvlb_weights', lvlb_weights, persistent=False) + + def get_input(self, batch): + x = batch['GT'] + mask = batch['inpaint_mask'] + inpaint = batch['inpaint_image'] + reference = batch['ref_imgs'] + hint = batch['hint'] + cloth_annotation = batch['caption_cloth'] + + x = x.to(memory_format=torch.contiguous_format).float() + mask = mask.to(memory_format=torch.contiguous_format).float() + inpaint = inpaint.to(memory_format=torch.contiguous_format).float() + reference = reference.to(memory_format=torch.contiguous_format).float() + hint = hint.to(memory_format=torch.contiguous_format).float() + + return x, inpaint, mask, reference, hint, cloth_annotation + + def q_sample(self, x_start, t, noise=None): + noise = default(noise, lambda: torch.randn_like(x_start)) + return (extract_into_tensor(self.sqrt_alphas_cumprod, t, x_start.shape) * x_start + + extract_into_tensor(self.sqrt_one_minus_alphas_cumprod, t, x_start.shape) * noise) + + def get_loss(self, pred, target, mean=True): + if mean: + loss = torch.nn.functional.mse_loss(target, pred) + else: + loss = torch.nn.functional.mse_loss(target, pred, reduction='none') + return loss + diff --git a/ldm/models/diffusion/plms.py b/ldm/models/diffusion/plms.py new file mode 100644 index 0000000000000000000000000000000000000000..44c377be1474f8c3dc6195b20c462e3668ce60a4 --- /dev/null +++ b/ldm/models/diffusion/plms.py @@ -0,0 +1,239 @@ +"""SAMPLING ONLY.""" + +import torch +import numpy as np +from tqdm import tqdm +from functools import partial + +from ldm.modules.diffusionmodules.util import make_ddim_sampling_parameters, make_ddim_timesteps, noise_like + + +class PLMSSampler(object): + def __init__(self, model, schedule="linear", **kwargs): + super().__init__() + self.model = model + self.ddpm_num_timesteps = model.num_timesteps + self.schedule = schedule + + def register_buffer(self, name, attr): + if type(attr) == torch.Tensor: + if attr.device != torch.device("cuda"): + attr = attr.to(torch.device("cuda")) + setattr(self, name, attr) + + def make_schedule(self, ddim_num_steps, ddim_discretize="uniform", ddim_eta=0., verbose=True): + if ddim_eta != 0: + raise ValueError('ddim_eta must be 0 for PLMS') + self.ddim_timesteps = make_ddim_timesteps(ddim_discr_method=ddim_discretize, num_ddim_timesteps=ddim_num_steps, + num_ddpm_timesteps=self.ddpm_num_timesteps,verbose=verbose) + alphas_cumprod = self.model.alphas_cumprod + assert alphas_cumprod.shape[0] == self.ddpm_num_timesteps, 'alphas have to be defined for each timestep' + to_torch = lambda x: x.clone().detach().to(torch.float32).to(self.model.device) + + self.register_buffer('betas', to_torch(self.model.betas)) + self.register_buffer('alphas_cumprod', to_torch(alphas_cumprod)) + self.register_buffer('alphas_cumprod_prev', to_torch(self.model.alphas_cumprod_prev)) + + # calculations for diffusion q(x_t | x_{t-1}) and others + self.register_buffer('sqrt_alphas_cumprod', to_torch(np.sqrt(alphas_cumprod.cpu()))) + self.register_buffer('sqrt_one_minus_alphas_cumprod', to_torch(np.sqrt(1. - alphas_cumprod.cpu()))) + self.register_buffer('log_one_minus_alphas_cumprod', to_torch(np.log(1. - alphas_cumprod.cpu()))) + self.register_buffer('sqrt_recip_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod.cpu()))) + self.register_buffer('sqrt_recipm1_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod.cpu() - 1))) + + # ddim sampling parameters + ddim_sigmas, ddim_alphas, ddim_alphas_prev = make_ddim_sampling_parameters(alphacums=alphas_cumprod.cpu(), + ddim_timesteps=self.ddim_timesteps, + eta=ddim_eta,verbose=verbose) + self.register_buffer('ddim_sigmas', ddim_sigmas) + self.register_buffer('ddim_alphas', ddim_alphas) + self.register_buffer('ddim_alphas_prev', ddim_alphas_prev) + self.register_buffer('ddim_sqrt_one_minus_alphas', np.sqrt(1. - ddim_alphas)) + sigmas_for_original_sampling_steps = ddim_eta * torch.sqrt( + (1 - self.alphas_cumprod_prev) / (1 - self.alphas_cumprod) * ( + 1 - self.alphas_cumprod / self.alphas_cumprod_prev)) + self.register_buffer('ddim_sigmas_for_original_num_steps', sigmas_for_original_sampling_steps) + + @torch.no_grad() + def sample(self, + S, + batch_size, + shape, + conditioning=None, + callback=None, + normals_sequence=None, + img_callback=None, + quantize_x0=False, + eta=0., + mask=None, + x0=None, + temperature=1., + noise_dropout=0., + score_corrector=None, + corrector_kwargs=None, + verbose=True, + x_T=None, + log_every_t=100, + unconditional_guidance_scale=1., + unconditional_conditioning=None, + # this has to come in the same format as the conditioning, # e.g. as encoded tokens, ... + **kwargs + ): + if conditioning is not None: + if isinstance(conditioning, dict): + cbs = conditioning[list(conditioning.keys())[0]].shape[0] + if cbs != batch_size: + print(f"Warning: Got {cbs} conditionings but batch-size is {batch_size}") + else: + if conditioning.shape[0] != batch_size: + print(f"Warning: Got {conditioning.shape[0]} conditionings but batch-size is {batch_size}") + + self.make_schedule(ddim_num_steps=S, ddim_eta=eta, verbose=verbose) + # sampling + C, H, W = shape + size = (batch_size, C, H, W) + print(f'Data shape for PLMS sampling is {size}') + + samples, intermediates = self.plms_sampling(conditioning, size, + callback=callback, + img_callback=img_callback, + quantize_denoised=quantize_x0, + mask=mask, x0=x0, + ddim_use_original_steps=False, + noise_dropout=noise_dropout, + temperature=temperature, + score_corrector=score_corrector, + corrector_kwargs=corrector_kwargs, + x_T=x_T, + log_every_t=log_every_t, + unconditional_guidance_scale=unconditional_guidance_scale, + unconditional_conditioning=unconditional_conditioning, + **kwargs + ) + return samples, intermediates + + @torch.no_grad() + def plms_sampling(self, cond, shape, + x_T=None, ddim_use_original_steps=False, + callback=None, timesteps=None, quantize_denoised=False, + mask=None, x0=None, img_callback=None, log_every_t=100, + temperature=1., noise_dropout=0., score_corrector=None, corrector_kwargs=None, + unconditional_guidance_scale=1., unconditional_conditioning=None,**kwargs): + device = self.model.betas.device + b = shape[0] + if x_T is None: + img = torch.randn(shape, device=device) + else: + img = x_T + + if timesteps is None: + timesteps = self.ddpm_num_timesteps if ddim_use_original_steps else self.ddim_timesteps + elif timesteps is not None and not ddim_use_original_steps: + subset_end = int(min(timesteps / self.ddim_timesteps.shape[0], 1) * self.ddim_timesteps.shape[0]) - 1 + timesteps = self.ddim_timesteps[:subset_end] + + intermediates = {'x_inter': [img], 'pred_x0': [img]} + time_range = list(reversed(range(0,timesteps))) if ddim_use_original_steps else np.flip(timesteps) + total_steps = timesteps if ddim_use_original_steps else timesteps.shape[0] + print(f"Running PLMS Sampling with {total_steps} timesteps") + + iterator = tqdm(time_range, desc='PLMS Sampler', total=total_steps) + old_eps = [] + + for i, step in enumerate(iterator): + index = total_steps - i - 1 + ts = torch.full((b,), step, device=device, dtype=torch.long) + ts_next = torch.full((b,), time_range[min(i + 1, len(time_range) - 1)], device=device, dtype=torch.long) + + if mask is not None: + assert x0 is not None + img_orig = self.model.q_sample(x0, ts) # TODO: deterministic forward pass? + img = img_orig * mask + (1. - mask) * img + + outs = self.p_sample_plms(img, cond, ts, index=index, use_original_steps=ddim_use_original_steps, + quantize_denoised=quantize_denoised, temperature=temperature, + noise_dropout=noise_dropout, score_corrector=score_corrector, + corrector_kwargs=corrector_kwargs, + unconditional_guidance_scale=unconditional_guidance_scale, + unconditional_conditioning=unconditional_conditioning, + old_eps=old_eps, t_next=ts_next,**kwargs) + img, pred_x0, e_t = outs + old_eps.append(e_t) + if len(old_eps) >= 4: + old_eps.pop(0) + if callback: callback(i) + if img_callback: img_callback(pred_x0, i) + + if index % log_every_t == 0 or index == total_steps - 1: + intermediates['x_inter'].append(img) + intermediates['pred_x0'].append(pred_x0) + + return img, intermediates + + @torch.no_grad() + def p_sample_plms(self, x, c, t, index, repeat_noise=False, use_original_steps=False, quantize_denoised=False, + temperature=1., noise_dropout=0., score_corrector=None, corrector_kwargs=None, + unconditional_guidance_scale=1., unconditional_conditioning=None, old_eps=None, t_next=None,**kwargs): + b, *_, device = *x.shape, x.device + def get_model_output(x, t): + if unconditional_conditioning is None or unconditional_guidance_scale == 1.: + e_t = self.model.apply_model(x, t, c) + else: + x_in = torch.cat([x] * 2) + t_in = torch.cat([t] * 2) + c_in = torch.cat([unconditional_conditioning, c]) + e_t_uncond, e_t = self.model.apply_model(x_in, t_in, c_in).chunk(2) + e_t = e_t_uncond + unconditional_guidance_scale * (e_t - e_t_uncond) + + if score_corrector is not None: + assert self.model.parameterization == "eps" + e_t = score_corrector.modify_score(self.model, e_t, x, t, c, **corrector_kwargs) + + return e_t + + alphas = self.model.alphas_cumprod if use_original_steps else self.ddim_alphas + alphas_prev = self.model.alphas_cumprod_prev if use_original_steps else self.ddim_alphas_prev + sqrt_one_minus_alphas = self.model.sqrt_one_minus_alphas_cumprod if use_original_steps else self.ddim_sqrt_one_minus_alphas + sigmas = self.model.ddim_sigmas_for_original_num_steps if use_original_steps else self.ddim_sigmas + + def get_x_prev_and_pred_x0(e_t, index): + # select parameters corresponding to the currently considered timestep + a_t = torch.full((b, 1, 1, 1), alphas[index], device=device) + a_prev = torch.full((b, 1, 1, 1), alphas_prev[index], device=device) + sigma_t = torch.full((b, 1, 1, 1), sigmas[index], device=device) + sqrt_one_minus_at = torch.full((b, 1, 1, 1), sqrt_one_minus_alphas[index],device=device) + + # current prediction for x_0 + pred_x0 = (x - sqrt_one_minus_at * e_t) / a_t.sqrt() + if quantize_denoised: + pred_x0, _, *_ = self.model.first_stage_model.quantize(pred_x0) + # direction pointing to x_t + dir_xt = (1. - a_prev - sigma_t**2).sqrt() * e_t + noise = sigma_t * noise_like(x.shape, device, repeat_noise) * temperature + if noise_dropout > 0.: + noise = torch.nn.functional.dropout(noise, p=noise_dropout) + x_prev = a_prev.sqrt() * pred_x0 + dir_xt + noise + return x_prev, pred_x0 + kwargs=kwargs['test_model_kwargs'] + print(x.shape,kwargs['inpaint_image'].shape,kwargs['inpaint_mask'].shape) + x_new=torch.cat([x,kwargs['inpaint_image'],kwargs['inpaint_mask']],dim=1) + e_t = get_model_output(x_new, t) + if len(old_eps) == 0: + # Pseudo Improved Euler (2nd order) + x_prev, pred_x0 = get_x_prev_and_pred_x0(e_t, index) + x_prev_new=torch.cat([x_prev,kwargs['inpaint_image'],kwargs['inpaint_mask']],dim=1) + e_t_next = get_model_output(x_prev_new, t_next) + e_t_prime = (e_t + e_t_next) / 2 + elif len(old_eps) == 1: + # 2nd order Pseudo Linear Multistep (Adams-Bashforth) + e_t_prime = (3 * e_t - old_eps[-1]) / 2 + elif len(old_eps) == 2: + # 3nd order Pseudo Linear Multistep (Adams-Bashforth) + e_t_prime = (23 * e_t - 16 * old_eps[-1] + 5 * old_eps[-2]) / 12 + elif len(old_eps) >= 3: + # 4nd order Pseudo Linear Multistep (Adams-Bashforth) + e_t_prime = (55 * e_t - 59 * old_eps[-1] + 37 * old_eps[-2] - 9 * old_eps[-3]) / 24 + + x_prev, pred_x0 = get_x_prev_and_pred_x0(e_t_prime, index) + + return x_prev, pred_x0, e_t diff --git a/ldm/modules/.DS_Store b/ldm/modules/.DS_Store new file mode 100644 index 0000000000000000000000000000000000000000..5008ddfcf53c02e82d7eee2e57c38e5672ef89f6 Binary files /dev/null and b/ldm/modules/.DS_Store differ diff --git a/ldm/modules/__pycache__/attention.cpython-38.pyc b/ldm/modules/__pycache__/attention.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..a190315c6586b17ccd84ac15f376adadc1c70e40 Binary files /dev/null and b/ldm/modules/__pycache__/attention.cpython-38.pyc differ diff --git a/ldm/modules/__pycache__/x_transformer.cpython-38.pyc b/ldm/modules/__pycache__/x_transformer.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..be465a5c9ed2b7db3eb4c0e9fbff92e636db3a99 Binary files /dev/null and b/ldm/modules/__pycache__/x_transformer.cpython-38.pyc differ diff --git a/ldm/modules/attention.py b/ldm/modules/attention.py new file mode 100644 index 0000000000000000000000000000000000000000..83e8d56ca33ffe6b15e04ce3ef64d0e56f2c786b --- /dev/null +++ b/ldm/modules/attention.py @@ -0,0 +1,345 @@ +from inspect import isfunction +import math +import torch +import torch.nn.functional as F +from torch import nn, einsum +from einops import rearrange, repeat + +from ldm.modules.diffusionmodules.util import checkpoint + +try: + import xformers + import xformers.ops + XFORMERS_IS_AVAILBLE = True +except: + XFORMERS_IS_AVAILBLE = False + + +def exists(val): + return val is not None + + +def uniq(arr): + return{el: True for el in arr}.keys() + + +def default(val, d): + if exists(val): + return val + return d() if isfunction(d) else d + + +def max_neg_value(t): + return -torch.finfo(t.dtype).max + + +def init_(tensor): + dim = tensor.shape[-1] + std = 1 / math.sqrt(dim) + tensor.uniform_(-std, std) + return tensor + + +# feedforward +class GEGLU(nn.Module): + def __init__(self, dim_in, dim_out): + super().__init__() + self.proj = nn.Linear(dim_in, dim_out * 2) + + def forward(self, x): + x, gate = self.proj(x).chunk(2, dim=-1) + return x * F.gelu(gate) + + +class FeedForward(nn.Module): + def __init__(self, dim, dim_out=None, mult=4, glu=False, dropout=0.): + super().__init__() + inner_dim = int(dim * mult) + dim_out = default(dim_out, dim) + project_in = nn.Sequential( + nn.Linear(dim, inner_dim), + nn.GELU() + ) if not glu else GEGLU(dim, inner_dim) + + self.net = nn.Sequential( + project_in, + nn.Dropout(dropout), + nn.Linear(inner_dim, dim_out) + ) + + def forward(self, x): + return self.net(x) + + +def zero_module(module): + """ + Zero out the parameters of a module and return it. + """ + for p in module.parameters(): + p.detach().zero_() + return module + + +def Normalize(in_channels): + return torch.nn.GroupNorm(num_groups=32, num_channels=in_channels, eps=1e-6, affine=True) + + +class LinearAttention(nn.Module): + def __init__(self, dim, heads=4, dim_head=32): + super().__init__() + self.heads = heads + hidden_dim = dim_head * heads + self.to_qkv = nn.Conv2d(dim, hidden_dim * 3, 1, bias = False) + self.to_out = nn.Conv2d(hidden_dim, dim, 1) + + def forward(self, x): + b, c, h, w = x.shape + qkv = self.to_qkv(x) + q, k, v = rearrange(qkv, 'b (qkv heads c) h w -> qkv b heads c (h w)', heads = self.heads, qkv=3) + k = k.softmax(dim=-1) + context = torch.einsum('bhdn,bhen->bhde', k, v) + out = torch.einsum('bhde,bhdn->bhen', context, q) + out = rearrange(out, 'b heads c (h w) -> b (heads c) h w', heads=self.heads, h=h, w=w) + return self.to_out(out) + + +class SpatialSelfAttention(nn.Module): + def __init__(self, in_channels): + super().__init__() + self.in_channels = in_channels + + self.norm = Normalize(in_channels) + self.q = torch.nn.Conv2d(in_channels, + in_channels, + kernel_size=1, + stride=1, + padding=0) + self.k = torch.nn.Conv2d(in_channels, + in_channels, + kernel_size=1, + stride=1, + padding=0) + self.v = torch.nn.Conv2d(in_channels, + in_channels, + kernel_size=1, + stride=1, + padding=0) + self.proj_out = torch.nn.Conv2d(in_channels, + in_channels, + kernel_size=1, + stride=1, + padding=0) + + def forward(self, x): + h_ = x + h_ = self.norm(h_) + q = self.q(h_) + k = self.k(h_) + v = self.v(h_) + + # compute attention + b,c,h,w = q.shape + q = rearrange(q, 'b c h w -> b (h w) c') + k = rearrange(k, 'b c h w -> b c (h w)') + w_ = torch.einsum('bij,bjk->bik', q, k) + + w_ = w_ * (int(c)**(-0.5)) + w_ = torch.nn.functional.softmax(w_, dim=2) + + # attend to values + v = rearrange(v, 'b c h w -> b c (h w)') + w_ = rearrange(w_, 'b i j -> b j i') + h_ = torch.einsum('bij,bjk->bik', v, w_) + h_ = rearrange(h_, 'b c (h w) -> b c h w', h=h) + h_ = self.proj_out(h_) + + return x+h_ + + +class CrossAttention(nn.Module): + def __init__(self, query_dim, context_dim=None, heads=8, dim_head=64, dropout=0.): + super().__init__() + inner_dim = dim_head * heads + context_dim = default(context_dim, query_dim) + + self.scale = dim_head ** -0.5 + self.heads = heads + + self.to_q = nn.Linear(query_dim, inner_dim, bias=False) + self.to_k = nn.Linear(context_dim, inner_dim, bias=False) + self.to_v = nn.Linear(context_dim, inner_dim, bias=False) + + self.to_out = nn.Sequential( + nn.Linear(inner_dim, query_dim), + nn.Dropout(dropout) + ) + + def forward(self, x, context=None, mask=None): + h = self.heads + + q = self.to_q(x) + context = default(context, x) + k = self.to_k(context) + v = self.to_v(context) + + q, k, v = map(lambda t: rearrange(t, 'b n (h d) -> (b h) n d', h=h), (q, k, v)) + + sim = einsum('b i d, b j d -> b i j', q, k) * self.scale + + if exists(mask): + mask = rearrange(mask, 'b ... -> b (...)') + max_neg_value = -torch.finfo(sim.dtype).max + mask = repeat(mask, 'b j -> (b h) () j', h=h) + sim.masked_fill_(~mask, max_neg_value) + + # attention, what we cannot get enough of + attn = sim.softmax(dim=-1) + + out = einsum('b i j, b j d -> b i d', attn, v) + out = rearrange(out, '(b h) n d -> b n (h d)', h=h) + return self.to_out(out) + + +class MemoryEfficientCrossAttention(nn.Module): + # https://github.com/MatthieuTPHR/diffusers/blob/d80b531ff8060ec1ea982b65a1b8df70f73aa67c/src/diffusers/models/attention.py#L223 + def __init__(self, query_dim, context_dim=None, heads=8, dim_head=64, dropout=0.0): + super().__init__() + print(f"Setting up {self.__class__.__name__}. Query dim is {query_dim}, context_dim is {context_dim} and using " + f"{heads} heads.") + inner_dim = dim_head * heads + context_dim = default(context_dim, query_dim) + + self.heads = heads + self.dim_head = dim_head + + self.to_q = nn.Linear(query_dim, inner_dim, bias=False) + self.to_k = nn.Linear(context_dim, inner_dim, bias=False) + self.to_v = nn.Linear(context_dim, inner_dim, bias=False) + + self.to_out = nn.Sequential(nn.Linear(inner_dim, query_dim), nn.Dropout(dropout)) + self.attention_op: Optional[Any] = None + + def forward(self, x, context=None, mask=None): + q = self.to_q(x) + context = default(context, x) + k = self.to_k(context) + v = self.to_v(context) + + b, _, _ = q.shape + q, k, v = map( + lambda t: t.unsqueeze(3) + .reshape(b, t.shape[1], self.heads, self.dim_head) + .permute(0, 2, 1, 3) + .reshape(b * self.heads, t.shape[1], self.dim_head) + .contiguous(), + (q, k, v), + ) + + # actually compute the attention, what we cannot get enough of + out = xformers.ops.memory_efficient_attention(q, k, v, attn_bias=None, op=self.attention_op) + + if exists(mask): + raise NotImplementedError + out = ( + out.unsqueeze(0) + .reshape(b, self.heads, out.shape[1], self.dim_head) + .permute(0, 2, 1, 3) + .reshape(b, out.shape[1], self.heads * self.dim_head) + ) + return self.to_out(out) + +class BasicTransformerBlock(nn.Module): + ATTENTION_MODES = { + "softmax": CrossAttention, # vanilla attention + "softmax-xformers": MemoryEfficientCrossAttention + } + def __init__(self, dim, n_heads, d_head, dropout=0., context_dim=None, gated_ff=True, checkpoint=True, + disable_self_attn=False): + super().__init__() + attn_mode = "softmax-xformers" if XFORMERS_IS_AVAILBLE else "softmax" + assert attn_mode in self.ATTENTION_MODES + attn_cls = self.ATTENTION_MODES[attn_mode] + self.disable_self_attn = disable_self_attn + self.attn1 = attn_cls(query_dim=dim, heads=n_heads, dim_head=d_head, dropout=dropout, + context_dim=context_dim if self.disable_self_attn else None) # is a self-attention if not self.disable_self_attn + self.ff = FeedForward(dim, dropout=dropout, glu=gated_ff) + self.attn2 = attn_cls(query_dim=dim, context_dim=context_dim, + heads=n_heads, dim_head=d_head, dropout=dropout) # is self-attn if context is none + self.norm1 = nn.LayerNorm(dim) + self.norm2 = nn.LayerNorm(dim) + self.norm3 = nn.LayerNorm(dim) + self.checkpoint = checkpoint + + def forward(self, x, context=None): + return checkpoint(self._forward, (x, context), self.parameters(), self.checkpoint) + + def _forward(self, x, context=None): + x = self.attn1(self.norm1(x), context=context if self.disable_self_attn else None) + x + x = self.attn2(self.norm2(x), context=context) + x + x = self.ff(self.norm3(x)) + x + return x + + +class SpatialTransformer(nn.Module): + """ + Transformer block for image-like data. + First, project the input (aka embedding) + and reshape to b, t, d. + Then apply standard transformer action. + Finally, reshape to image + NEW: use_linear for more efficiency instead of the 1x1 convs + """ + def __init__(self, in_channels, n_heads, d_head, + depth=1, dropout=0., context_dim=None, + disable_self_attn=False, use_linear=False, + use_checkpoint=True): + super().__init__() + if exists(context_dim) and not isinstance(context_dim, list): + context_dim = [context_dim] + self.in_channels = in_channels + inner_dim = n_heads * d_head + self.norm = Normalize(in_channels) + if not use_linear: + self.proj_in = nn.Conv2d(in_channels, + inner_dim, + kernel_size=1, + stride=1, + padding=0) + else: + self.proj_in = nn.Linear(in_channels, inner_dim) + + self.transformer_blocks = nn.ModuleList( + [BasicTransformerBlock(inner_dim, n_heads, d_head, dropout=dropout, context_dim=context_dim[d], + disable_self_attn=disable_self_attn, checkpoint=use_checkpoint) + for d in range(depth)] + ) + if not use_linear: + self.proj_out = zero_module(nn.Conv2d(inner_dim, + in_channels, + kernel_size=1, + stride=1, + padding=0)) + else: + self.proj_out = zero_module(nn.Linear(in_channels, inner_dim)) + self.use_linear = use_linear + + def forward(self, x, context=None): + # note: if no context is given, cross-attention defaults to self-attention + if not isinstance(context, list): + context = [context] + b, c, h, w = x.shape + x_in = x + x = self.norm(x) + if not self.use_linear: + x = self.proj_in(x) + x = rearrange(x, 'b c h w -> b (h w) c').contiguous() + if self.use_linear: + x = self.proj_in(x) + for i, block in enumerate(self.transformer_blocks): + x = block(x, context=context[i]) + if self.use_linear: + x = self.proj_out(x) + x = rearrange(x, 'b (h w) c -> b c h w', h=h, w=w).contiguous() + if not self.use_linear: + x = self.proj_out(x) + return x + x_in \ No newline at end of file diff --git a/ldm/modules/diffusionmodules/.DS_Store b/ldm/modules/diffusionmodules/.DS_Store new file mode 100644 index 0000000000000000000000000000000000000000..5008ddfcf53c02e82d7eee2e57c38e5672ef89f6 Binary files /dev/null and b/ldm/modules/diffusionmodules/.DS_Store differ diff --git a/ldm/modules/diffusionmodules/__init__.py b/ldm/modules/diffusionmodules/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/ldm/modules/diffusionmodules/__pycache__/__init__.cpython-38.pyc b/ldm/modules/diffusionmodules/__pycache__/__init__.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..6ec06b49c745519cbc153aa945c921ef6636b502 Binary files /dev/null and b/ldm/modules/diffusionmodules/__pycache__/__init__.cpython-38.pyc differ diff --git a/ldm/modules/diffusionmodules/__pycache__/model.cpython-38.pyc b/ldm/modules/diffusionmodules/__pycache__/model.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..175906996f2806f7901cebfe5ff682af512f5356 Binary files /dev/null and b/ldm/modules/diffusionmodules/__pycache__/model.cpython-38.pyc differ diff --git a/ldm/modules/diffusionmodules/__pycache__/openaimodel.cpython-38.pyc b/ldm/modules/diffusionmodules/__pycache__/openaimodel.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..8b7b41a9270b520b2cf853748da723a36c38b599 Binary files /dev/null and b/ldm/modules/diffusionmodules/__pycache__/openaimodel.cpython-38.pyc differ diff --git a/ldm/modules/diffusionmodules/__pycache__/util.cpython-38.pyc b/ldm/modules/diffusionmodules/__pycache__/util.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..cbbe43900552e2b3a5f658abb322f53228797377 Binary files /dev/null and b/ldm/modules/diffusionmodules/__pycache__/util.cpython-38.pyc differ diff --git a/ldm/modules/diffusionmodules/model.py b/ldm/modules/diffusionmodules/model.py new file mode 100644 index 0000000000000000000000000000000000000000..4f5ad2f07702e8907a6b81454efc00e3879c8528 --- /dev/null +++ b/ldm/modules/diffusionmodules/model.py @@ -0,0 +1,835 @@ +# pytorch_diffusion + derived encoder decoder +import math +import torch +import torch.nn as nn +import numpy as np +from einops import rearrange + +from ldm.util import instantiate_from_config +from ldm.modules.attention import LinearAttention + + +def get_timestep_embedding(timesteps, embedding_dim): + """ + This matches the implementation in Denoising Diffusion Probabilistic Models: + From Fairseq. + Build sinusoidal embeddings. + This matches the implementation in tensor2tensor, but differs slightly + from the description in Section 3.5 of "Attention Is All You Need". + """ + assert len(timesteps.shape) == 1 + + half_dim = embedding_dim // 2 + emb = math.log(10000) / (half_dim - 1) + emb = torch.exp(torch.arange(half_dim, dtype=torch.float32) * -emb) + emb = emb.to(device=timesteps.device) + emb = timesteps.float()[:, None] * emb[None, :] + emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=1) + if embedding_dim % 2 == 1: # zero pad + emb = torch.nn.functional.pad(emb, (0,1,0,0)) + return emb + + +def nonlinearity(x): + # swish + return x*torch.sigmoid(x) + + +def Normalize(in_channels, num_groups=32): + return torch.nn.GroupNorm(num_groups=num_groups, num_channels=in_channels, eps=1e-6, affine=True) + + +class Upsample(nn.Module): + def __init__(self, in_channels, with_conv): + super().__init__() + self.with_conv = with_conv + if self.with_conv: + self.conv = torch.nn.Conv2d(in_channels, + in_channels, + kernel_size=3, + stride=1, + padding=1) + + def forward(self, x): + x = torch.nn.functional.interpolate(x, scale_factor=2.0, mode="nearest") + if self.with_conv: + x = self.conv(x) + return x + + +class Downsample(nn.Module): + def __init__(self, in_channels, with_conv): + super().__init__() + self.with_conv = with_conv + if self.with_conv: + # no asymmetric padding in torch conv, must do it ourselves + self.conv = torch.nn.Conv2d(in_channels, + in_channels, + kernel_size=3, + stride=2, + padding=0) + + def forward(self, x): + if self.with_conv: + pad = (0,1,0,1) + x = torch.nn.functional.pad(x, pad, mode="constant", value=0) + x = self.conv(x) + else: + x = torch.nn.functional.avg_pool2d(x, kernel_size=2, stride=2) + return x + + +class ResnetBlock(nn.Module): + def __init__(self, *, in_channels, out_channels=None, conv_shortcut=False, + dropout, temb_channels=512): + super().__init__() + self.in_channels = in_channels + out_channels = in_channels if out_channels is None else out_channels + self.out_channels = out_channels + self.use_conv_shortcut = conv_shortcut + + self.norm1 = Normalize(in_channels) + self.conv1 = torch.nn.Conv2d(in_channels, + out_channels, + kernel_size=3, + stride=1, + padding=1) + if temb_channels > 0: + self.temb_proj = torch.nn.Linear(temb_channels, + out_channels) + self.norm2 = Normalize(out_channels) + self.dropout = torch.nn.Dropout(dropout) + self.conv2 = torch.nn.Conv2d(out_channels, + out_channels, + kernel_size=3, + stride=1, + padding=1) + if self.in_channels != self.out_channels: + if self.use_conv_shortcut: + self.conv_shortcut = torch.nn.Conv2d(in_channels, + out_channels, + kernel_size=3, + stride=1, + padding=1) + else: + self.nin_shortcut = torch.nn.Conv2d(in_channels, + out_channels, + kernel_size=1, + stride=1, + padding=0) + + def forward(self, x, temb): + h = x + h = self.norm1(h) + h = nonlinearity(h) + h = self.conv1(h) + + if temb is not None: + h = h + self.temb_proj(nonlinearity(temb))[:,:,None,None] + + h = self.norm2(h) + h = nonlinearity(h) + h = self.dropout(h) + h = self.conv2(h) + + if self.in_channels != self.out_channels: + if self.use_conv_shortcut: + x = self.conv_shortcut(x) + else: + x = self.nin_shortcut(x) + + return x+h + + +class LinAttnBlock(LinearAttention): + """to match AttnBlock usage""" + def __init__(self, in_channels): + super().__init__(dim=in_channels, heads=1, dim_head=in_channels) + + +class AttnBlock(nn.Module): + def __init__(self, in_channels): + super().__init__() + self.in_channels = in_channels + + self.norm = Normalize(in_channels) + self.q = torch.nn.Conv2d(in_channels, + in_channels, + kernel_size=1, + stride=1, + padding=0) + self.k = torch.nn.Conv2d(in_channels, + in_channels, + kernel_size=1, + stride=1, + padding=0) + self.v = torch.nn.Conv2d(in_channels, + in_channels, + kernel_size=1, + stride=1, + padding=0) + self.proj_out = torch.nn.Conv2d(in_channels, + in_channels, + kernel_size=1, + stride=1, + padding=0) + + + def forward(self, x): + h_ = x + h_ = self.norm(h_) + q = self.q(h_) + k = self.k(h_) + v = self.v(h_) + + # compute attention + b,c,h,w = q.shape + q = q.reshape(b,c,h*w) + q = q.permute(0,2,1) # b,hw,c + k = k.reshape(b,c,h*w) # b,c,hw + w_ = torch.bmm(q,k) # b,hw,hw w[b,i,j]=sum_c q[b,i,c]k[b,c,j] + w_ = w_ * (int(c)**(-0.5)) + w_ = torch.nn.functional.softmax(w_, dim=2) + + # attend to values + v = v.reshape(b,c,h*w) + w_ = w_.permute(0,2,1) # b,hw,hw (first hw of k, second of q) + h_ = torch.bmm(v,w_) # b, c,hw (hw of q) h_[b,c,j] = sum_i v[b,c,i] w_[b,i,j] + h_ = h_.reshape(b,c,h,w) + + h_ = self.proj_out(h_) + + return x+h_ + + +def make_attn(in_channels, attn_type="vanilla"): + assert attn_type in ["vanilla", "linear", "none"], f'attn_type {attn_type} unknown' + print(f"making attention of type '{attn_type}' with {in_channels} in_channels") + if attn_type == "vanilla": + return AttnBlock(in_channels) + elif attn_type == "none": + return nn.Identity(in_channels) + else: + return LinAttnBlock(in_channels) + + +class Model(nn.Module): + def __init__(self, *, ch, out_ch, ch_mult=(1,2,4,8), num_res_blocks, + attn_resolutions, dropout=0.0, resamp_with_conv=True, in_channels, + resolution, use_timestep=True, use_linear_attn=False, attn_type="vanilla"): + super().__init__() + if use_linear_attn: attn_type = "linear" + self.ch = ch + self.temb_ch = self.ch*4 + self.num_resolutions = len(ch_mult) + self.num_res_blocks = num_res_blocks + self.resolution = resolution + self.in_channels = in_channels + + self.use_timestep = use_timestep + if self.use_timestep: + # timestep embedding + self.temb = nn.Module() + self.temb.dense = nn.ModuleList([ + torch.nn.Linear(self.ch, + self.temb_ch), + torch.nn.Linear(self.temb_ch, + self.temb_ch), + ]) + + # downsampling + self.conv_in = torch.nn.Conv2d(in_channels, + self.ch, + kernel_size=3, + stride=1, + padding=1) + + curr_res = resolution + in_ch_mult = (1,)+tuple(ch_mult) + self.down = nn.ModuleList() + for i_level in range(self.num_resolutions): + block = nn.ModuleList() + attn = nn.ModuleList() + block_in = ch*in_ch_mult[i_level] + block_out = ch*ch_mult[i_level] + for i_block in range(self.num_res_blocks): + block.append(ResnetBlock(in_channels=block_in, + out_channels=block_out, + temb_channels=self.temb_ch, + dropout=dropout)) + block_in = block_out + if curr_res in attn_resolutions: + attn.append(make_attn(block_in, attn_type=attn_type)) + down = nn.Module() + down.block = block + down.attn = attn + if i_level != self.num_resolutions-1: + down.downsample = Downsample(block_in, resamp_with_conv) + curr_res = curr_res // 2 + self.down.append(down) + + # middle + self.mid = nn.Module() + self.mid.block_1 = ResnetBlock(in_channels=block_in, + out_channels=block_in, + temb_channels=self.temb_ch, + dropout=dropout) + self.mid.attn_1 = make_attn(block_in, attn_type=attn_type) + self.mid.block_2 = ResnetBlock(in_channels=block_in, + out_channels=block_in, + temb_channels=self.temb_ch, + dropout=dropout) + + # upsampling + self.up = nn.ModuleList() + for i_level in reversed(range(self.num_resolutions)): + block = nn.ModuleList() + attn = nn.ModuleList() + block_out = ch*ch_mult[i_level] + skip_in = ch*ch_mult[i_level] + for i_block in range(self.num_res_blocks+1): + if i_block == self.num_res_blocks: + skip_in = ch*in_ch_mult[i_level] + block.append(ResnetBlock(in_channels=block_in+skip_in, + out_channels=block_out, + temb_channels=self.temb_ch, + dropout=dropout)) + block_in = block_out + if curr_res in attn_resolutions: + attn.append(make_attn(block_in, attn_type=attn_type)) + up = nn.Module() + up.block = block + up.attn = attn + if i_level != 0: + up.upsample = Upsample(block_in, resamp_with_conv) + curr_res = curr_res * 2 + self.up.insert(0, up) # prepend to get consistent order + + # end + self.norm_out = Normalize(block_in) + self.conv_out = torch.nn.Conv2d(block_in, + out_ch, + kernel_size=3, + stride=1, + padding=1) + + def forward(self, x, t=None, context=None): + #assert x.shape[2] == x.shape[3] == self.resolution + if context is not None: + # assume aligned context, cat along channel axis + x = torch.cat((x, context), dim=1) + if self.use_timestep: + # timestep embedding + assert t is not None + temb = get_timestep_embedding(t, self.ch) + temb = self.temb.dense[0](temb) + temb = nonlinearity(temb) + temb = self.temb.dense[1](temb) + else: + temb = None + + # downsampling + hs = [self.conv_in(x)] + for i_level in range(self.num_resolutions): + for i_block in range(self.num_res_blocks): + h = self.down[i_level].block[i_block](hs[-1], temb) + if len(self.down[i_level].attn) > 0: + h = self.down[i_level].attn[i_block](h) + hs.append(h) + if i_level != self.num_resolutions-1: + hs.append(self.down[i_level].downsample(hs[-1])) + + # middle + h = hs[-1] + h = self.mid.block_1(h, temb) + h = self.mid.attn_1(h) + h = self.mid.block_2(h, temb) + + # upsampling + for i_level in reversed(range(self.num_resolutions)): + for i_block in range(self.num_res_blocks+1): + h = self.up[i_level].block[i_block]( + torch.cat([h, hs.pop()], dim=1), temb) + if len(self.up[i_level].attn) > 0: + h = self.up[i_level].attn[i_block](h) + if i_level != 0: + h = self.up[i_level].upsample(h) + + # end + h = self.norm_out(h) + h = nonlinearity(h) + h = self.conv_out(h) + return h + + def get_last_layer(self): + return self.conv_out.weight + + +class Encoder(nn.Module): + def __init__(self, *, ch, out_ch, ch_mult=(1,2,4,8), num_res_blocks, + attn_resolutions, dropout=0.0, resamp_with_conv=True, in_channels, + resolution, z_channels, double_z=True, use_linear_attn=False, attn_type="vanilla", + **ignore_kwargs): + super().__init__() + if use_linear_attn: attn_type = "linear" + self.ch = ch + self.temb_ch = 0 + self.num_resolutions = len(ch_mult) + self.num_res_blocks = num_res_blocks + self.resolution = resolution + self.in_channels = in_channels + + # downsampling + self.conv_in = torch.nn.Conv2d(in_channels, + self.ch, + kernel_size=3, + stride=1, + padding=1) + + curr_res = resolution + in_ch_mult = (1,)+tuple(ch_mult) + self.in_ch_mult = in_ch_mult + self.down = nn.ModuleList() + for i_level in range(self.num_resolutions): + block = nn.ModuleList() + attn = nn.ModuleList() + block_in = ch*in_ch_mult[i_level] + block_out = ch*ch_mult[i_level] + for i_block in range(self.num_res_blocks): + block.append(ResnetBlock(in_channels=block_in, + out_channels=block_out, + temb_channels=self.temb_ch, + dropout=dropout)) + block_in = block_out + if curr_res in attn_resolutions: + attn.append(make_attn(block_in, attn_type=attn_type)) + down = nn.Module() + down.block = block + down.attn = attn + if i_level != self.num_resolutions-1: + down.downsample = Downsample(block_in, resamp_with_conv) + curr_res = curr_res // 2 + self.down.append(down) + + # middle + self.mid = nn.Module() + self.mid.block_1 = ResnetBlock(in_channels=block_in, + out_channels=block_in, + temb_channels=self.temb_ch, + dropout=dropout) + self.mid.attn_1 = make_attn(block_in, attn_type=attn_type) + self.mid.block_2 = ResnetBlock(in_channels=block_in, + out_channels=block_in, + temb_channels=self.temb_ch, + dropout=dropout) + + # end + self.norm_out = Normalize(block_in) + self.conv_out = torch.nn.Conv2d(block_in, + 2*z_channels if double_z else z_channels, + kernel_size=3, + stride=1, + padding=1) + + def forward(self, x): + # timestep embedding + temb = None + + # downsampling + hs = [self.conv_in(x)] + for i_level in range(self.num_resolutions): + for i_block in range(self.num_res_blocks): + h = self.down[i_level].block[i_block](hs[-1], temb) + if len(self.down[i_level].attn) > 0: + h = self.down[i_level].attn[i_block](h) + hs.append(h) + if i_level != self.num_resolutions-1: + hs.append(self.down[i_level].downsample(hs[-1])) + + # middle + h = hs[-1] + h = self.mid.block_1(h, temb) + h = self.mid.attn_1(h) + h = self.mid.block_2(h, temb) + + # end + h = self.norm_out(h) + h = nonlinearity(h) + h = self.conv_out(h) + return h + + +class Decoder(nn.Module): + def __init__(self, *, ch, out_ch, ch_mult=(1,2,4,8), num_res_blocks, + attn_resolutions, dropout=0.0, resamp_with_conv=True, in_channels, + resolution, z_channels, give_pre_end=False, tanh_out=False, use_linear_attn=False, + attn_type="vanilla", **ignorekwargs): + super().__init__() + if use_linear_attn: attn_type = "linear" + self.ch = ch + self.temb_ch = 0 + self.num_resolutions = len(ch_mult) + self.num_res_blocks = num_res_blocks + self.resolution = resolution + self.in_channels = in_channels + self.give_pre_end = give_pre_end + self.tanh_out = tanh_out + + # compute in_ch_mult, block_in and curr_res at lowest res + in_ch_mult = (1,)+tuple(ch_mult) + block_in = ch*ch_mult[self.num_resolutions-1] + curr_res = resolution // 2**(self.num_resolutions-1) + self.z_shape = (1,z_channels,curr_res,curr_res) + print("Working with z of shape {} = {} dimensions.".format( + self.z_shape, np.prod(self.z_shape))) + + # z to block_in + self.conv_in = torch.nn.Conv2d(z_channels, + block_in, + kernel_size=3, + stride=1, + padding=1) + + # middle + self.mid = nn.Module() + self.mid.block_1 = ResnetBlock(in_channels=block_in, + out_channels=block_in, + temb_channels=self.temb_ch, + dropout=dropout) + self.mid.attn_1 = make_attn(block_in, attn_type=attn_type) + self.mid.block_2 = ResnetBlock(in_channels=block_in, + out_channels=block_in, + temb_channels=self.temb_ch, + dropout=dropout) + + # upsampling + self.up = nn.ModuleList() + for i_level in reversed(range(self.num_resolutions)): + block = nn.ModuleList() + attn = nn.ModuleList() + block_out = ch*ch_mult[i_level] + for i_block in range(self.num_res_blocks+1): + block.append(ResnetBlock(in_channels=block_in, + out_channels=block_out, + temb_channels=self.temb_ch, + dropout=dropout)) + block_in = block_out + if curr_res in attn_resolutions: + attn.append(make_attn(block_in, attn_type=attn_type)) + up = nn.Module() + up.block = block + up.attn = attn + if i_level != 0: + up.upsample = Upsample(block_in, resamp_with_conv) + curr_res = curr_res * 2 + self.up.insert(0, up) # prepend to get consistent order + + # end + self.norm_out = Normalize(block_in) + self.conv_out = torch.nn.Conv2d(block_in, + out_ch, + kernel_size=3, + stride=1, + padding=1) + + def forward(self, z): + #assert z.shape[1:] == self.z_shape[1:] + self.last_z_shape = z.shape + + # timestep embedding + temb = None + + # z to block_in + h = self.conv_in(z) + + # middle + h = self.mid.block_1(h, temb) + h = self.mid.attn_1(h) + h = self.mid.block_2(h, temb) + + # upsampling + for i_level in reversed(range(self.num_resolutions)): + for i_block in range(self.num_res_blocks+1): + h = self.up[i_level].block[i_block](h, temb) + if len(self.up[i_level].attn) > 0: + h = self.up[i_level].attn[i_block](h) + if i_level != 0: + h = self.up[i_level].upsample(h) + + # end + if self.give_pre_end: + return h + + h = self.norm_out(h) + h = nonlinearity(h) + h = self.conv_out(h) + if self.tanh_out: + h = torch.tanh(h) + return h + + +class SimpleDecoder(nn.Module): + def __init__(self, in_channels, out_channels, *args, **kwargs): + super().__init__() + self.model = nn.ModuleList([nn.Conv2d(in_channels, in_channels, 1), + ResnetBlock(in_channels=in_channels, + out_channels=2 * in_channels, + temb_channels=0, dropout=0.0), + ResnetBlock(in_channels=2 * in_channels, + out_channels=4 * in_channels, + temb_channels=0, dropout=0.0), + ResnetBlock(in_channels=4 * in_channels, + out_channels=2 * in_channels, + temb_channels=0, dropout=0.0), + nn.Conv2d(2*in_channels, in_channels, 1), + Upsample(in_channels, with_conv=True)]) + # end + self.norm_out = Normalize(in_channels) + self.conv_out = torch.nn.Conv2d(in_channels, + out_channels, + kernel_size=3, + stride=1, + padding=1) + + def forward(self, x): + for i, layer in enumerate(self.model): + if i in [1,2,3]: + x = layer(x, None) + else: + x = layer(x) + + h = self.norm_out(x) + h = nonlinearity(h) + x = self.conv_out(h) + return x + + +class UpsampleDecoder(nn.Module): + def __init__(self, in_channels, out_channels, ch, num_res_blocks, resolution, + ch_mult=(2,2), dropout=0.0): + super().__init__() + # upsampling + self.temb_ch = 0 + self.num_resolutions = len(ch_mult) + self.num_res_blocks = num_res_blocks + block_in = in_channels + curr_res = resolution // 2 ** (self.num_resolutions - 1) + self.res_blocks = nn.ModuleList() + self.upsample_blocks = nn.ModuleList() + for i_level in range(self.num_resolutions): + res_block = [] + block_out = ch * ch_mult[i_level] + for i_block in range(self.num_res_blocks + 1): + res_block.append(ResnetBlock(in_channels=block_in, + out_channels=block_out, + temb_channels=self.temb_ch, + dropout=dropout)) + block_in = block_out + self.res_blocks.append(nn.ModuleList(res_block)) + if i_level != self.num_resolutions - 1: + self.upsample_blocks.append(Upsample(block_in, True)) + curr_res = curr_res * 2 + + # end + self.norm_out = Normalize(block_in) + self.conv_out = torch.nn.Conv2d(block_in, + out_channels, + kernel_size=3, + stride=1, + padding=1) + + def forward(self, x): + # upsampling + h = x + for k, i_level in enumerate(range(self.num_resolutions)): + for i_block in range(self.num_res_blocks + 1): + h = self.res_blocks[i_level][i_block](h, None) + if i_level != self.num_resolutions - 1: + h = self.upsample_blocks[k](h) + h = self.norm_out(h) + h = nonlinearity(h) + h = self.conv_out(h) + return h + + +class LatentRescaler(nn.Module): + def __init__(self, factor, in_channels, mid_channels, out_channels, depth=2): + super().__init__() + # residual block, interpolate, residual block + self.factor = factor + self.conv_in = nn.Conv2d(in_channels, + mid_channels, + kernel_size=3, + stride=1, + padding=1) + self.res_block1 = nn.ModuleList([ResnetBlock(in_channels=mid_channels, + out_channels=mid_channels, + temb_channels=0, + dropout=0.0) for _ in range(depth)]) + self.attn = AttnBlock(mid_channels) + self.res_block2 = nn.ModuleList([ResnetBlock(in_channels=mid_channels, + out_channels=mid_channels, + temb_channels=0, + dropout=0.0) for _ in range(depth)]) + + self.conv_out = nn.Conv2d(mid_channels, + out_channels, + kernel_size=1, + ) + + def forward(self, x): + x = self.conv_in(x) + for block in self.res_block1: + x = block(x, None) + x = torch.nn.functional.interpolate(x, size=(int(round(x.shape[2]*self.factor)), int(round(x.shape[3]*self.factor)))) + x = self.attn(x) + for block in self.res_block2: + x = block(x, None) + x = self.conv_out(x) + return x + + +class MergedRescaleEncoder(nn.Module): + def __init__(self, in_channels, ch, resolution, out_ch, num_res_blocks, + attn_resolutions, dropout=0.0, resamp_with_conv=True, + ch_mult=(1,2,4,8), rescale_factor=1.0, rescale_module_depth=1): + super().__init__() + intermediate_chn = ch * ch_mult[-1] + self.encoder = Encoder(in_channels=in_channels, num_res_blocks=num_res_blocks, ch=ch, ch_mult=ch_mult, + z_channels=intermediate_chn, double_z=False, resolution=resolution, + attn_resolutions=attn_resolutions, dropout=dropout, resamp_with_conv=resamp_with_conv, + out_ch=None) + self.rescaler = LatentRescaler(factor=rescale_factor, in_channels=intermediate_chn, + mid_channels=intermediate_chn, out_channels=out_ch, depth=rescale_module_depth) + + def forward(self, x): + x = self.encoder(x) + x = self.rescaler(x) + return x + + +class MergedRescaleDecoder(nn.Module): + def __init__(self, z_channels, out_ch, resolution, num_res_blocks, attn_resolutions, ch, ch_mult=(1,2,4,8), + dropout=0.0, resamp_with_conv=True, rescale_factor=1.0, rescale_module_depth=1): + super().__init__() + tmp_chn = z_channels*ch_mult[-1] + self.decoder = Decoder(out_ch=out_ch, z_channels=tmp_chn, attn_resolutions=attn_resolutions, dropout=dropout, + resamp_with_conv=resamp_with_conv, in_channels=None, num_res_blocks=num_res_blocks, + ch_mult=ch_mult, resolution=resolution, ch=ch) + self.rescaler = LatentRescaler(factor=rescale_factor, in_channels=z_channels, mid_channels=tmp_chn, + out_channels=tmp_chn, depth=rescale_module_depth) + + def forward(self, x): + x = self.rescaler(x) + x = self.decoder(x) + return x + + +class Upsampler(nn.Module): + def __init__(self, in_size, out_size, in_channels, out_channels, ch_mult=2): + super().__init__() + assert out_size >= in_size + num_blocks = int(np.log2(out_size//in_size))+1 + factor_up = 1.+ (out_size % in_size) + print(f"Building {self.__class__.__name__} with in_size: {in_size} --> out_size {out_size} and factor {factor_up}") + self.rescaler = LatentRescaler(factor=factor_up, in_channels=in_channels, mid_channels=2*in_channels, + out_channels=in_channels) + self.decoder = Decoder(out_ch=out_channels, resolution=out_size, z_channels=in_channels, num_res_blocks=2, + attn_resolutions=[], in_channels=None, ch=in_channels, + ch_mult=[ch_mult for _ in range(num_blocks)]) + + def forward(self, x): + x = self.rescaler(x) + x = self.decoder(x) + return x + + +class Resize(nn.Module): + def __init__(self, in_channels=None, learned=False, mode="bilinear"): + super().__init__() + self.with_conv = learned + self.mode = mode + if self.with_conv: + print(f"Note: {self.__class__.__name} uses learned downsampling and will ignore the fixed {mode} mode") + raise NotImplementedError() + assert in_channels is not None + # no asymmetric padding in torch conv, must do it ourselves + self.conv = torch.nn.Conv2d(in_channels, + in_channels, + kernel_size=4, + stride=2, + padding=1) + + def forward(self, x, scale_factor=1.0): + if scale_factor==1.0: + return x + else: + x = torch.nn.functional.interpolate(x, mode=self.mode, align_corners=False, scale_factor=scale_factor) + return x + +class FirstStagePostProcessor(nn.Module): + + def __init__(self, ch_mult:list, in_channels, + pretrained_model:nn.Module=None, + reshape=False, + n_channels=None, + dropout=0., + pretrained_config=None): + super().__init__() + if pretrained_config is None: + assert pretrained_model is not None, 'Either "pretrained_model" or "pretrained_config" must not be None' + self.pretrained_model = pretrained_model + else: + assert pretrained_config is not None, 'Either "pretrained_model" or "pretrained_config" must not be None' + self.instantiate_pretrained(pretrained_config) + + self.do_reshape = reshape + + if n_channels is None: + n_channels = self.pretrained_model.encoder.ch + + self.proj_norm = Normalize(in_channels,num_groups=in_channels//2) + self.proj = nn.Conv2d(in_channels,n_channels,kernel_size=3, + stride=1,padding=1) + + blocks = [] + downs = [] + ch_in = n_channels + for m in ch_mult: + blocks.append(ResnetBlock(in_channels=ch_in,out_channels=m*n_channels,dropout=dropout)) + ch_in = m * n_channels + downs.append(Downsample(ch_in, with_conv=False)) + + self.model = nn.ModuleList(blocks) + self.downsampler = nn.ModuleList(downs) + + + def instantiate_pretrained(self, config): + model = instantiate_from_config(config) + self.pretrained_model = model.eval() + # self.pretrained_model.train = False + for param in self.pretrained_model.parameters(): + param.requires_grad = False + + + @torch.no_grad() + def encode_with_pretrained(self,x): + c = self.pretrained_model.encode(x) + if isinstance(c, DiagonalGaussianDistribution): + c = c.mode() + return c + + def forward(self,x): + z_fs = self.encode_with_pretrained(x) + z = self.proj_norm(z_fs) + z = self.proj(z) + z = nonlinearity(z) + + for submodel, downmodel in zip(self.model,self.downsampler): + z = submodel(z,temb=None) + z = downmodel(z) + + if self.do_reshape: + z = rearrange(z,'b c h w -> b (h w) c') + return z + diff --git a/ldm/modules/diffusionmodules/openaimodel.py b/ldm/modules/diffusionmodules/openaimodel.py new file mode 100644 index 0000000000000000000000000000000000000000..a4fe35b942c0c35e65190aa1ddc8eab8f942b31e --- /dev/null +++ b/ldm/modules/diffusionmodules/openaimodel.py @@ -0,0 +1,707 @@ +from abc import abstractmethod +from functools import partial +import math +from typing import Iterable + +import numpy as np +import torch as th +import torch.nn as nn +import torch.nn.functional as F + +from ldm.modules.diffusionmodules.util import ( + checkpoint, + conv_nd, + linear, + avg_pool_nd, + zero_module, + normalization, + timestep_embedding, +) +from ldm.modules.attention import SpatialTransformer + + +# dummy replace +def convert_module_to_f16(x): + pass + +def convert_module_to_f32(x): + pass + + +## go +class AttentionPool2d(nn.Module): + """ + Adapted from CLIP: https://github.com/openai/CLIP/blob/main/clip/model.py + """ + + def __init__( + self, + spacial_dim: int, + embed_dim: int, + num_heads_channels: int, + output_dim: int = None, + ): + super().__init__() + self.positional_embedding = nn.Parameter(th.randn(embed_dim, spacial_dim ** 2 + 1) / embed_dim ** 0.5) + self.qkv_proj = conv_nd(1, embed_dim, 3 * embed_dim, 1) + self.c_proj = conv_nd(1, embed_dim, output_dim or embed_dim, 1) + self.num_heads = embed_dim // num_heads_channels + self.attention = QKVAttention(self.num_heads) + + def forward(self, x): + b, c, *_spatial = x.shape + x = x.reshape(b, c, -1) # NC(HW) + x = th.cat([x.mean(dim=-1, keepdim=True), x], dim=-1) # NC(HW+1) + x = x + self.positional_embedding[None, :, :].to(x.dtype) # NC(HW+1) + x = self.qkv_proj(x) + x = self.attention(x) + x = self.c_proj(x) + return x[:, :, 0] + + +class TimestepBlock(nn.Module): + @abstractmethod + def forward(self, x, emb): + """ + Apply the module to `x` given `emb` timestep embeddings. + """ + + +class TimestepEmbedSequential(nn.Sequential, TimestepBlock): + def forward(self, x, emb, context=None): + for layer in self: + if isinstance(layer, TimestepBlock): + x = layer(x, emb) + elif isinstance(layer, SpatialTransformer): + x = layer(x, context) + else: + x = layer(x) + return x + + +class Upsample(nn.Module): + """ + An upsampling layer with an optional convolution. + :param channels: channels in the inputs and outputs. + :param use_conv: a bool determining if a convolution is applied. + :param dims: determines if the signal is 1D, 2D, or 3D. If 3D, then + upsampling occurs in the inner-two dimensions. + """ + + def __init__(self, channels, use_conv, dims=2, out_channels=None, padding=1): + super().__init__() + self.channels = channels + self.out_channels = out_channels or channels + self.use_conv = use_conv + self.dims = dims + if use_conv: + self.conv = conv_nd(dims, self.channels, self.out_channels, 3, padding=padding) + + def forward(self, x): + assert x.shape[1] == self.channels + if self.dims == 3: + x = F.interpolate( + x, (x.shape[2], x.shape[3] * 2, x.shape[4] * 2), mode="nearest" + ) + else: + x = F.interpolate(x, scale_factor=2, mode="nearest") + if self.use_conv: + x = self.conv(x) + return x + +class TransposedUpsample(nn.Module): + 'Learned 2x upsampling without padding' + def __init__(self, channels, out_channels=None, ks=5): + super().__init__() + self.channels = channels + self.out_channels = out_channels or channels + + self.up = nn.ConvTranspose2d(self.channels,self.out_channels,kernel_size=ks,stride=2) + + def forward(self,x): + return self.up(x) + + +class Downsample(nn.Module): + """ + A downsampling layer with an optional convolution. + :param channels: channels in the inputs and outputs. + :param use_conv: a bool determining if a convolution is applied. + :param dims: determines if the signal is 1D, 2D, or 3D. If 3D, then + downsampling occurs in the inner-two dimensions. + """ + + def __init__(self, channels, use_conv, dims=2, out_channels=None,padding=1): + super().__init__() + self.channels = channels + self.out_channels = out_channels or channels + self.use_conv = use_conv + self.dims = dims + stride = 2 if dims != 3 else (1, 2, 2) + if use_conv: + self.op = conv_nd( + dims, self.channels, self.out_channels, 3, stride=stride, padding=padding + ) + else: + assert self.channels == self.out_channels + self.op = avg_pool_nd(dims, kernel_size=stride, stride=stride) + + def forward(self, x): + assert x.shape[1] == self.channels + return self.op(x) + + +class ResBlock(TimestepBlock): + def __init__( + self, + channels, + emb_channels, + dropout, + out_channels=None, + use_conv=False, + use_scale_shift_norm=False, + dims=2, + use_checkpoint=False, + up=False, + down=False, + ): + super().__init__() + self.channels = channels + self.emb_channels = emb_channels + self.dropout = dropout + self.out_channels = out_channels or channels + self.use_conv = use_conv + self.use_checkpoint = use_checkpoint + self.use_scale_shift_norm = use_scale_shift_norm + + self.in_layers = nn.Sequential( + normalization(channels), + nn.SiLU(), + conv_nd(dims, channels, self.out_channels, 3, padding=1), + ) + + self.updown = up or down + + if up: + self.h_upd = Upsample(channels, False, dims) + self.x_upd = Upsample(channels, False, dims) + elif down: + self.h_upd = Downsample(channels, False, dims) + self.x_upd = Downsample(channels, False, dims) + else: + self.h_upd = self.x_upd = nn.Identity() + + self.emb_layers = nn.Sequential( + nn.SiLU(), + linear( + emb_channels, + 2 * self.out_channels if use_scale_shift_norm else self.out_channels, + ), + ) + self.out_layers = nn.Sequential( + normalization(self.out_channels), + nn.SiLU(), + nn.Dropout(p=dropout), + zero_module( + conv_nd(dims, self.out_channels, self.out_channels, 3, padding=1) + ), + ) + + if self.out_channels == channels: + self.skip_connection = nn.Identity() + elif use_conv: + self.skip_connection = conv_nd( + dims, channels, self.out_channels, 3, padding=1 + ) + else: + self.skip_connection = conv_nd(dims, channels, self.out_channels, 1) + + def forward(self, x, emb): + """ + Apply the block to a Tensor, conditioned on a timestep embedding. + :param x: an [N x C x ...] Tensor of features. + :param emb: an [N x emb_channels] Tensor of timestep embeddings. + :return: an [N x C x ...] Tensor of outputs. + """ + return checkpoint( + self._forward, (x, emb), self.parameters(), self.use_checkpoint + ) + + + def _forward(self, x, emb): + if self.updown: + in_rest, in_conv = self.in_layers[:-1], self.in_layers[-1] + h = in_rest(x) + h = self.h_upd(h) + x = self.x_upd(x) + h = in_conv(h) + else: + h = self.in_layers(x) + emb_out = self.emb_layers(emb).type(h.dtype) + while len(emb_out.shape) < len(h.shape): + emb_out = emb_out[..., None] + if self.use_scale_shift_norm: + out_norm, out_rest = self.out_layers[0], self.out_layers[1:] + scale, shift = th.chunk(emb_out, 2, dim=1) + h = out_norm(h) * (1 + scale) + shift + h = out_rest(h) + else: + h = h + emb_out + h = self.out_layers(h) + return self.skip_connection(x) + h + + +class My_ResBlock(TimestepBlock): + """ + A residual block that can optionally change the number of channels. + :param channels: the number of input channels. + :param emb_channels: the number of timestep embedding channels. + :param dropout: the rate of dropout. + :param out_channels: if specified, the number of out channels. + :param use_conv: if True and out_channels is specified, use a spatial + convolution instead of a smaller 1x1 convolution to change the + channels in the skip connection. + :param dims: determines if the signal is 1D, 2D, or 3D. + :param use_checkpoint: if True, use gradient checkpointing on this module. + :param up: if True, use this block for upsampling. + :param down: if True, use this block for downsampling. + """ + + def __init__( + self, + channels, + emb_channels, + dropout, + out_channels=None, + use_conv=False, + use_scale_shift_norm=False, + dims=2, + use_checkpoint=False, + up=False, + down=False, + ): + super().__init__() + self.channels = channels + self.emb_channels = emb_channels + self.dropout = dropout + self.out_channels = out_channels or channels + self.use_conv = use_conv + self.use_checkpoint = use_checkpoint + self.use_scale_shift_norm = use_scale_shift_norm + + self.in_layers = nn.Sequential( + normalization(channels), + nn.SiLU(), + conv_nd(dims, channels, self.out_channels, 3, padding=1), + ) + + self.updown = up or down + + if up: + self.h_upd = Upsample(channels, False, dims) + self.x_upd = Upsample(channels, False, dims) + elif down: + self.h_upd = Downsample(channels, False, dims) + self.x_upd = Downsample(channels, False, dims) + else: + self.h_upd = self.x_upd = nn.Identity() + + self.emb_layers = nn.Sequential( + nn.SiLU(), + linear( + emb_channels, + 2 * self.out_channels if use_scale_shift_norm else self.out_channels, + ), + ) + self.out_layers = nn.Sequential( + normalization(self.out_channels), + nn.SiLU(), + nn.Dropout(p=dropout), + zero_module( + conv_nd(dims, self.out_channels, 4, 3, padding=1) + ), + ) + + if self.out_channels == channels: + self.skip_connection = nn.Identity() + elif use_conv: + self.skip_connection = conv_nd( + dims, channels, self.out_channels, 3, padding=1 + ) + else: + self.skip_connection = conv_nd(dims, channels, 4, 1) + + def forward(self, x, emb): + """ + Apply the block to a Tensor, conditioned on a timestep embedding. + :param x: an [N x C x ...] Tensor of features. + :param emb: an [N x emb_channels] Tensor of timestep embeddings. + :return: an [N x C x ...] Tensor of outputs. + """ + return checkpoint( + self._forward, (x, emb), self.parameters(), self.use_checkpoint + ) + + + def _forward(self, x, emb): + if self.updown: + in_rest, in_conv = self.in_layers[:-1], self.in_layers[-1] + h = in_rest(x) + h = self.h_upd(h) + x = self.x_upd(x) + h = in_conv(h) + else: + h = self.in_layers(x) + emb_out = self.emb_layers(emb).type(h.dtype) + while len(emb_out.shape) < len(h.shape): + emb_out = emb_out[..., None] + if self.use_scale_shift_norm: + out_norm, out_rest = self.out_layers[0], self.out_layers[1:] + scale, shift = th.chunk(emb_out, 2, dim=1) + h = out_norm(h) * (1 + scale) + shift + h = out_rest(h) + else: + h = h + emb_out + h = self.out_layers(h) + return h + + +class AttentionBlock(nn.Module): + """ + An attention block that allows spatial positions to attend to each other. + Originally ported from here, but adapted to the N-d case. + https://github.com/hojonathanho/diffusion/blob/1e0dceb3b3495bbe19116a5e1b3596cd0706c543/diffusion_tf/models/unet.py#L66. + """ + + def __init__( + self, + channels, + num_heads=1, + num_head_channels=-1, + use_checkpoint=False, + use_new_attention_order=False, + ): + super().__init__() + self.channels = channels + if num_head_channels == -1: + self.num_heads = num_heads + else: + assert ( + channels % num_head_channels == 0 + ), f"q,k,v channels {channels} is not divisible by num_head_channels {num_head_channels}" + self.num_heads = channels // num_head_channels + self.use_checkpoint = use_checkpoint + self.norm = normalization(channels) + self.qkv = conv_nd(1, channels, channels * 3, 1) + if use_new_attention_order: + # split qkv before split heads + self.attention = QKVAttention(self.num_heads) + else: + # split heads before split qkv + self.attention = QKVAttentionLegacy(self.num_heads) + + self.proj_out = zero_module(conv_nd(1, channels, channels, 1)) + + def forward(self, x): + return checkpoint(self._forward, (x,), self.parameters(), True) # TODO: check checkpoint usage, is True # TODO: fix the .half call!!! + #return pt_checkpoint(self._forward, x) # pytorch + + def _forward(self, x): + b, c, *spatial = x.shape + x = x.reshape(b, c, -1) + qkv = self.qkv(self.norm(x)) + h = self.attention(qkv) + h = self.proj_out(h) + return (x + h).reshape(b, c, *spatial) + + +def count_flops_attn(model, _x, y): + """ + A counter for the `thop` package to count the operations in an + attention operation. + Meant to be used like: + macs, params = thop.profile( + model, + inputs=(inputs, timestamps), + custom_ops={QKVAttention: QKVAttention.count_flops}, + ) + """ + b, c, *spatial = y[0].shape + num_spatial = int(np.prod(spatial)) + # We perform two matmuls with the same number of ops. + # The first computes the weight matrix, the second computes + # the combination of the value vectors. + matmul_ops = 2 * b * (num_spatial ** 2) * c + model.total_ops += th.DoubleTensor([matmul_ops]) + + +class QKVAttentionLegacy(nn.Module): + """ + A module which performs QKV attention. Matches legacy QKVAttention + input/ouput heads shaping + """ + + def __init__(self, n_heads): + super().__init__() + self.n_heads = n_heads + + def forward(self, qkv): + """ + Apply QKV attention. + :param qkv: an [N x (H * 3 * C) x T] tensor of Qs, Ks, and Vs. + :return: an [N x (H * C) x T] tensor after attention. + """ + bs, width, length = qkv.shape + assert width % (3 * self.n_heads) == 0 + ch = width // (3 * self.n_heads) + q, k, v = qkv.reshape(bs * self.n_heads, ch * 3, length).split(ch, dim=1) + scale = 1 / math.sqrt(math.sqrt(ch)) + weight = th.einsum( + "bct,bcs->bts", q * scale, k * scale + ) # More stable with f16 than dividing afterwards + weight = th.softmax(weight.float(), dim=-1).type(weight.dtype) + a = th.einsum("bts,bcs->bct", weight, v) + return a.reshape(bs, -1, length) + + @staticmethod + def count_flops(model, _x, y): + return count_flops_attn(model, _x, y) + + +class QKVAttention(nn.Module): + """ + A module which performs QKV attention and splits in a different order. + """ + + def __init__(self, n_heads): + super().__init__() + self.n_heads = n_heads + + def forward(self, qkv): + """ + Apply QKV attention. + :param qkv: an [N x (3 * H * C) x T] tensor of Qs, Ks, and Vs. + :return: an [N x (H * C) x T] tensor after attention. + """ + bs, width, length = qkv.shape + assert width % (3 * self.n_heads) == 0 + ch = width // (3 * self.n_heads) + q, k, v = qkv.chunk(3, dim=1) + scale = 1 / math.sqrt(math.sqrt(ch)) + weight = th.einsum( + "bct,bcs->bts", + (q * scale).view(bs * self.n_heads, ch, length), + (k * scale).view(bs * self.n_heads, ch, length), + ) # More stable with f16 than dividing afterwards + weight = th.softmax(weight.float(), dim=-1).type(weight.dtype) + a = th.einsum("bts,bcs->bct", weight, v.reshape(bs * self.n_heads, ch, length)) + return a.reshape(bs, -1, length) + + @staticmethod + def count_flops(model, _x, y): + return count_flops_attn(model, _x, y) + + +class UNetModel(nn.Module): + def __init__(self, + image_size, # 32 + in_channels, # 9 + out_channels, # 4 + model_channels, # 320 + attention_resolutions, # [ 4, 2, 1 ] + num_res_blocks, # 2 + channel_mult=(1, 2, 4, 8), # [ 1, 2, 4, 4 ] + num_heads=-1, # 8 + use_spatial_transformer=False, # True + transformer_depth=1, # 1 + context_dim=None, # 768 + use_checkpoint=False, # True + legacy=True, # False + add_conv_in_front_of_unet=False, # False + dropout=0, + conv_resample=True, + dims=2, + num_classes=None, + num_head_channels=-1, + num_heads_upsample=-1, + use_scale_shift_norm=False): + + super().__init__() + + self.image_size = image_size # 32 + self.in_channels = in_channels # 9 + self.out_channels = out_channels # 4 + self.model_channels = model_channels # 320 + self.attention_resolutions = attention_resolutions # [4,2,1] + self.num_res_blocks = num_res_blocks # 2 + self.channel_mult = channel_mult # [1,2,4,4] + num_heads_upsample = num_heads # 8 + self.use_checkpoint = use_checkpoint # True + self.add_conv_in_front_of_unet=add_conv_in_front_of_unet # False + self.dropout = dropout # 0 + self.conv_resample = conv_resample # True + self.num_classes = num_classes # None + self.num_heads = num_heads # 8 + self.num_head_channels = num_head_channels # -1 + self.num_heads_upsample = num_heads_upsample # -1 + self.dtype = th.float32 + + + # 时间编码器 320 -> 320*4 -> 320*4 + time_embed_dim = model_channels * 4 + self.time_embed = nn.Sequential( + linear(model_channels, time_embed_dim), + nn.SiLU(), + linear(time_embed_dim, time_embed_dim), + ) + + + # 一阶段 self.input_blocks + self.input_blocks = nn.ModuleList( + [ + TimestepEmbedSequential( + conv_nd( + dims, # 2 + in_channels, # 9 + model_channels, # 320 + kernel_size=3, # 3 + padding=1 # 1 + ) + ) + ] + ) + + input_block_chans = [model_channels] # [320] + ch = model_channels # 320 + ds = 1 # 1 + + for level, mult in enumerate(channel_mult): # [0,1,2,3], [1,2,4,4] + for _ in range(num_res_blocks): # 2 + layers = [ + ResBlock( + ch, # [1,1,1,2,2,4,4,4]*320 + time_embed_dim, # 320*4 + dropout, # 0 + out_channels=mult * model_channels, # [1,1,2,2,4,4,4,4]*320 + dims=dims, # 2 + use_checkpoint=use_checkpoint, # True + use_scale_shift_norm=use_scale_shift_norm, # False + ) + ] + ch = mult * model_channels # [1,1,2,2,4,4,4,4]*320 + if ds in attention_resolutions: # 前6次小循环 + dim_head = ch // num_heads # [1,1,2,2,4,4]*40 + layers.append( + SpatialTransformer( + ch, # [1,1,2,2,4,4]*320 + num_heads, # 8 + dim_head, # [1,1,2,2,4,4]*40 + depth=transformer_depth, # 1 + context_dim=context_dim # 768 + ) + ) + self.input_blocks.append(TimestepEmbedSequential(*layers)) + input_block_chans.append(ch) # [1,1,2,2,4,4,4,4]*320 + if level != len(channel_mult) - 1: # 前3次大循环 + out_ch = ch # [1,2,4]*320 + self.input_blocks.append( + TimestepEmbedSequential( + Downsample( + ch, # [1,2,4]*320 + conv_resample, # True + dims=dims, # 2 + out_channels=out_ch # [1,2,4]*320 + ) + ) + ) + ch = out_ch # [1,2,4]*320 + input_block_chans.append(ch) # [1,2,4]*320 + ds *= 2 # 1 -> 2 -> 4 -> 8 + + + # 二阶段 self.middle_block + dim_head = ch // num_heads # 1280 // 8 + + self.middle_block = TimestepEmbedSequential( + ResBlock( + ch, # 4*320 + time_embed_dim, # 320 + dropout, # 0 + dims=dims, # 2 + use_checkpoint=use_checkpoint, # True + use_scale_shift_norm=use_scale_shift_norm, # False + ), + SpatialTransformer( + ch, # 4*320 + num_heads, # 8 + dim_head, # 160 + depth=transformer_depth, # 1 + context_dim=context_dim # 768 + ), + ResBlock( + ch, # 4*320 + time_embed_dim, # 320 + dropout, # 0 + dims=dims, # 2 + use_checkpoint=use_checkpoint, # True + use_scale_shift_norm=use_scale_shift_norm, # False + ), + ) + + + # 三阶段 self.output_blocks + self.output_blocks = nn.ModuleList([]) + for level, mult in list(enumerate(channel_mult))[::-1]: # [3,2,1,0], [4,4,2,1] + for i in range(num_res_blocks + 1): # 3 + ich = input_block_chans.pop() # [4,4, 4,4,4, 2,2,2, 1,1,1, 1]*320 + layers = [ + ResBlock( + ch + ich, # [4,4,4,4,4,4,4,2,2,2,1,1]*320+ich + time_embed_dim, # 320 + dropout, # 0 + out_channels=model_channels*mult, # [4,4,4,4,4,4,2,2,2,1,1,1]*320 + dims=dims, # 2 + use_checkpoint=use_checkpoint, # True + use_scale_shift_norm=use_scale_shift_norm, # False + ) + ] + ch = model_channels * mult # [4,4,4,4,4,4,2,2,2,1,1,1]*320 + if ds in attention_resolutions: # 后三次大循环 + dim_head = ch // num_heads + layers.append( + SpatialTransformer( + ch, num_heads, dim_head, depth=transformer_depth, context_dim=context_dim + ) + ) + if level and i == num_res_blocks: # 前三次大循环中每次最后的小循环 + out_ch = ch + layers.append( + Upsample(ch, conv_resample, dims=dims, out_channels=out_ch) + ) + ds //= 2 + self.output_blocks.append(TimestepEmbedSequential(*layers)) + + # 四阶段 self.out + self.out = nn.Sequential( + normalization(ch), + nn.SiLU(), + zero_module(conv_nd(dims, model_channels, out_channels, 3, padding=1)), + ) + + def forward(self, x, timesteps=None, context=None): + hs = [] + + t_emb = timestep_embedding(timesteps, self.model_channels) # [N, 320] + emb = self.time_embed(t_emb) # [N, 320*4] + h = x.type(self.dtype) # 将 x 转换为 torch.float32 + + for module in self.input_blocks: + h = module(h, emb, context) + hs.append(h) + h = self.middle_block(h, emb, context) + for module in self.output_blocks: + h = th.cat([h, hs.pop()], dim=1) + h = module(h, emb, context) + h = h.type(x.dtype) + + return self.out(h) # [N,4,H,W] + diff --git a/ldm/modules/diffusionmodules/util.py b/ldm/modules/diffusionmodules/util.py new file mode 100644 index 0000000000000000000000000000000000000000..8ba85e65946d9e48a1d15f9fd8d0f851cb45d426 --- /dev/null +++ b/ldm/modules/diffusionmodules/util.py @@ -0,0 +1,255 @@ +# adopted from +# https://github.com/openai/improved-diffusion/blob/main/improved_diffusion/gaussian_diffusion.py +# and +# https://github.com/lucidrains/denoising-diffusion-pytorch/blob/7706bdfc6f527f58d33f84b7b522e61e6e3164b3/denoising_diffusion_pytorch/denoising_diffusion_pytorch.py +# and +# https://github.com/openai/guided-diffusion/blob/0ba878e517b276c45d1195eb29f6f5f72659a05b/guided_diffusion/nn.py +# +# thanks! + + +import os +import math +import torch +import torch.nn as nn +import numpy as np +from einops import repeat + +from ldm.util import instantiate_from_config + + +def make_beta_schedule(schedule, n_timestep, linear_start=1e-4, linear_end=2e-2, cosine_s=8e-3): + if schedule == "linear": + betas = ( + torch.linspace(linear_start ** 0.5, linear_end ** 0.5, n_timestep, dtype=torch.float64) ** 2 + ) + + elif schedule == "cosine": + timesteps = ( + torch.arange(n_timestep + 1, dtype=torch.float64) / n_timestep + cosine_s + ) + alphas = timesteps / (1 + cosine_s) * np.pi / 2 + alphas = torch.cos(alphas).pow(2) + alphas = alphas / alphas[0] + betas = 1 - alphas[1:] / alphas[:-1] + betas = np.clip(betas, a_min=0, a_max=0.999) + + elif schedule == "sqrt_linear": + betas = torch.linspace(linear_start, linear_end, n_timestep, dtype=torch.float64) + elif schedule == "sqrt": + betas = torch.linspace(linear_start, linear_end, n_timestep, dtype=torch.float64) ** 0.5 + else: + raise ValueError(f"schedule '{schedule}' unknown.") + return betas.numpy() + + +def make_ddim_timesteps(ddim_discr_method, num_ddim_timesteps, num_ddpm_timesteps, verbose=True): + if ddim_discr_method == 'uniform': + c = num_ddpm_timesteps // num_ddim_timesteps + ddim_timesteps = np.asarray(list(range(0, num_ddpm_timesteps, c))) + elif ddim_discr_method == 'quad': + ddim_timesteps = ((np.linspace(0, np.sqrt(num_ddpm_timesteps * .8), num_ddim_timesteps)) ** 2).astype(int) + else: + raise NotImplementedError(f'There is no ddim discretization method called "{ddim_discr_method}"') + + # assert ddim_timesteps.shape[0] == num_ddim_timesteps + # add one to get the final alpha values right (the ones from first scale to data during sampling) + steps_out = ddim_timesteps + 1 + if verbose: + print(f'Selected timesteps for ddim sampler: {steps_out}') + return steps_out + + +def make_ddim_sampling_parameters(alphacums, ddim_timesteps, eta, verbose=True): + # select alphas for computing the variance schedule + alphas = alphacums[ddim_timesteps] + alphas_prev = np.asarray([alphacums[0]] + alphacums[ddim_timesteps[:-1]].tolist()) + + # according the the formula provided in https://arxiv.org/abs/2010.02502 + sigmas = eta * np.sqrt((1 - alphas_prev) / (1 - alphas) * (1 - alphas / alphas_prev)) + if verbose: + print(f'Selected alphas for ddim sampler: a_t: {alphas}; a_(t-1): {alphas_prev}') + print(f'For the chosen value of eta, which is {eta}, ' + f'this results in the following sigma_t schedule for ddim sampler {sigmas}') + return sigmas, alphas, alphas_prev + + +def betas_for_alpha_bar(num_diffusion_timesteps, alpha_bar, max_beta=0.999): + """ + Create a beta schedule that discretizes the given alpha_t_bar function, + which defines the cumulative product of (1-beta) over time from t = [0,1]. + :param num_diffusion_timesteps: the number of betas to produce. + :param alpha_bar: a lambda that takes an argument t from 0 to 1 and + produces the cumulative product of (1-beta) up to that + part of the diffusion process. + :param max_beta: the maximum beta to use; use values lower than 1 to + prevent singularities. + """ + betas = [] + for i in range(num_diffusion_timesteps): + t1 = i / num_diffusion_timesteps + t2 = (i + 1) / num_diffusion_timesteps + betas.append(min(1 - alpha_bar(t2) / alpha_bar(t1), max_beta)) + return np.array(betas) + + +def extract_into_tensor(a, t, x_shape): + b, *_ = t.shape + out = a.gather(-1, t) + return out.reshape(b, *((1,) * (len(x_shape) - 1))) + + +def checkpoint(func, inputs, params, flag): + """ + Evaluate a function without caching intermediate activations, allowing for + reduced memory at the expense of extra compute in the backward pass. + :param func: the function to evaluate. + :param inputs: the argument sequence to pass to `func`. + :param params: a sequence of parameters `func` depends on but does not + explicitly take as arguments. + :param flag: if False, disable gradient checkpointing. + """ + if flag: + args = tuple(inputs) + tuple(params) + return CheckpointFunction.apply(func, len(inputs), *args) + else: + return func(*inputs) + + +class CheckpointFunction(torch.autograd.Function): + @staticmethod + def forward(ctx, run_function, length, *args): + ctx.run_function = run_function + ctx.input_tensors = list(args[:length]) + ctx.input_params = list(args[length:]) + + with torch.no_grad(): + output_tensors = ctx.run_function(*ctx.input_tensors) + return output_tensors + + @staticmethod + def backward(ctx, *output_grads): + ctx.input_tensors = [x.detach().requires_grad_(True) for x in ctx.input_tensors] + with torch.enable_grad(): + # Fixes a bug where the first op in run_function modifies the + # Tensor storage in place, which is not allowed for detach()'d + # Tensors. + shallow_copies = [x.view_as(x) for x in ctx.input_tensors] + output_tensors = ctx.run_function(*shallow_copies) + input_grads = torch.autograd.grad( + output_tensors, + ctx.input_tensors + ctx.input_params, + output_grads, + allow_unused=True, + ) + del ctx.input_tensors + del ctx.input_params + del output_tensors + return (None, None) + input_grads + + +def timestep_embedding(timesteps, dim, max_period=10000): + half = dim // 2 # 160 + # 生成一个 shape 为 [160,] 的 torch，从 1 到 1/10000 指数衰减 + freqs = torch.exp(-math.log(max_period) * torch.arange(start=0, end=half, dtype=torch.float32) / half).to(device=timesteps.device) + # [1,2] * [1, 160] -> [2, 160] + args = timesteps[:, None].float() * freqs[None] + # [2, 320] + embedding = torch.cat([torch.cos(args), torch.sin(args)], dim=-1) + return embedding + + +def zero_module(module): + """ + Zero out the parameters of a module and return it. + """ + for p in module.parameters(): + p.detach().zero_() + return module + + +def scale_module(module, scale): + """ + Scale the parameters of a module and return it. + """ + for p in module.parameters(): + p.detach().mul_(scale) + return module + + +def mean_flat(tensor): + """ + Take the mean over all non-batch dimensions. + """ + return tensor.mean(dim=list(range(1, len(tensor.shape)))) + + +def normalization(channels): + """ + Make a standard normalization layer. + :param channels: number of input channels. + :return: an nn.Module for normalization. + """ + return GroupNorm32(32, channels) + + +# PyTorch 1.7 has SiLU, but we support PyTorch 1.5. +class SiLU(nn.Module): + def forward(self, x): + return x * torch.sigmoid(x) + + +class GroupNorm32(nn.GroupNorm): + def forward(self, x): + return super().forward(x.float()).type(x.dtype) + +def conv_nd(dims, *args, **kwargs): + """ + Create a 1D, 2D, or 3D convolution module. + """ + if dims == 1: + return nn.Conv1d(*args, **kwargs) + elif dims == 2: + return nn.Conv2d(*args, **kwargs) + elif dims == 3: + return nn.Conv3d(*args, **kwargs) + raise ValueError(f"unsupported dimensions: {dims}") + + +def linear(*args, **kwargs): + """ + Create a linear module. + """ + return nn.Linear(*args, **kwargs) + + +def avg_pool_nd(dims, *args, **kwargs): + """ + Create a 1D, 2D, or 3D average pooling module. + """ + if dims == 1: + return nn.AvgPool1d(*args, **kwargs) + elif dims == 2: + return nn.AvgPool2d(*args, **kwargs) + elif dims == 3: + return nn.AvgPool3d(*args, **kwargs) + raise ValueError(f"unsupported dimensions: {dims}") + + +class HybridConditioner(nn.Module): + + def __init__(self, c_concat_config, c_crossattn_config): + super().__init__() + self.concat_conditioner = instantiate_from_config(c_concat_config) + self.crossattn_conditioner = instantiate_from_config(c_crossattn_config) + + def forward(self, c_concat, c_crossattn): + c_concat = self.concat_conditioner(c_concat) + c_crossattn = self.crossattn_conditioner(c_crossattn) + return {'c_concat': [c_concat], 'c_crossattn': [c_crossattn]} + + +def noise_like(shape, device, repeat=False): + repeat_noise = lambda: torch.randn((1, *shape[1:]), device=device).repeat(shape[0], *((1,) * (len(shape) - 1))) + noise = lambda: torch.randn(shape, device=device) + return repeat_noise() if repeat else noise() \ No newline at end of file diff --git a/ldm/modules/distributions/.DS_Store b/ldm/modules/distributions/.DS_Store new file mode 100644 index 0000000000000000000000000000000000000000..5008ddfcf53c02e82d7eee2e57c38e5672ef89f6 Binary files /dev/null and b/ldm/modules/distributions/.DS_Store differ diff --git a/ldm/modules/distributions/__init__.py b/ldm/modules/distributions/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/ldm/modules/distributions/__pycache__/__init__.cpython-38.pyc b/ldm/modules/distributions/__pycache__/__init__.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..75a4e669e41c7a7eec7958a7d9f455ce081c3f65 Binary files /dev/null and b/ldm/modules/distributions/__pycache__/__init__.cpython-38.pyc differ diff --git a/ldm/modules/distributions/__pycache__/distributions.cpython-38.pyc b/ldm/modules/distributions/__pycache__/distributions.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..10f7ed4eac2bb42471fc5d3e46c243196cea8f12 Binary files /dev/null and b/ldm/modules/distributions/__pycache__/distributions.cpython-38.pyc differ diff --git a/ldm/modules/distributions/distributions.py b/ldm/modules/distributions/distributions.py new file mode 100644 index 0000000000000000000000000000000000000000..507b82a4fd8d177727c964a4b0217bd6766ab106 --- /dev/null +++ b/ldm/modules/distributions/distributions.py @@ -0,0 +1,92 @@ +import torch +import numpy as np + + +class AbstractDistribution: + def sample(self): + raise NotImplementedError() + + def mode(self): + raise NotImplementedError() + + +class DiracDistribution(AbstractDistribution): + def __init__(self, value): + self.value = value + + def sample(self): + return self.value + + def mode(self): + return self.value + + +class DiagonalGaussianDistribution(object): + def __init__(self, parameters, deterministic=False): + self.parameters = parameters + self.mean, self.logvar = torch.chunk(parameters, 2, dim=1) + self.logvar = torch.clamp(self.logvar, -30.0, 20.0) + self.deterministic = deterministic + self.std = torch.exp(0.5 * self.logvar) + self.var = torch.exp(self.logvar) + if self.deterministic: + self.var = self.std = torch.zeros_like(self.mean).to(device=self.parameters.device) + + def sample(self): + x = self.mean + self.std * torch.randn(self.mean.shape).to(device=self.parameters.device) + return x + + def kl(self, other=None): + if self.deterministic: + return torch.Tensor([0.]) + else: + if other is None: + return 0.5 * torch.sum(torch.pow(self.mean, 2) + + self.var - 1.0 - self.logvar, + dim=[1, 2, 3]) + else: + return 0.5 * torch.sum( + torch.pow(self.mean - other.mean, 2) / other.var + + self.var / other.var - 1.0 - self.logvar + other.logvar, + dim=[1, 2, 3]) + + def nll(self, sample, dims=[1,2,3]): + if self.deterministic: + return torch.Tensor([0.]) + logtwopi = np.log(2.0 * np.pi) + return 0.5 * torch.sum( + logtwopi + self.logvar + torch.pow(sample - self.mean, 2) / self.var, + dim=dims) + + def mode(self): + return self.mean + + +def normal_kl(mean1, logvar1, mean2, logvar2): + """ + source: https://github.com/openai/guided-diffusion/blob/27c20a8fab9cb472df5d6bdd6c8d11c8f430b924/guided_diffusion/losses.py#L12 + Compute the KL divergence between two gaussians. + Shapes are automatically broadcasted, so batches can be compared to + scalars, among other use cases. + """ + tensor = None + for obj in (mean1, logvar1, mean2, logvar2): + if isinstance(obj, torch.Tensor): + tensor = obj + break + assert tensor is not None, "at least one argument must be a Tensor" + + # Force variances to be Tensors. Broadcasting helps convert scalars to + # Tensors, but it does not work for torch.exp(). + logvar1, logvar2 = [ + x if isinstance(x, torch.Tensor) else torch.tensor(x).to(tensor) + for x in (logvar1, logvar2) + ] + + return 0.5 * ( + -1.0 + + logvar2 + - logvar1 + + torch.exp(logvar1 - logvar2) + + ((mean1 - mean2) ** 2) * torch.exp(-logvar2) + ) diff --git a/ldm/modules/ema.py b/ldm/modules/ema.py new file mode 100644 index 0000000000000000000000000000000000000000..102a4ba7908a3dda4897e7273ef977c1ba5629ac --- /dev/null +++ b/ldm/modules/ema.py @@ -0,0 +1,76 @@ +import torch +from torch import nn + + +class LitEma(nn.Module): + def __init__(self, model, decay=0.9999, use_num_upates=True): + super().__init__() + if decay < 0.0 or decay > 1.0: + raise ValueError('Decay must be between 0 and 1') + + self.m_name2s_name = {} + self.register_buffer('decay', torch.tensor(decay, dtype=torch.float32)) + self.register_buffer('num_updates', torch.tensor(0,dtype=torch.int) if use_num_upates + else torch.tensor(-1,dtype=torch.int)) + + for name, p in model.named_parameters(): + if p.requires_grad: + #remove as '.'-character is not allowed in buffers + s_name = name.replace('.','') + self.m_name2s_name.update({name:s_name}) + self.register_buffer(s_name,p.clone().detach().data) + + self.collected_params = [] + + def forward(self,model): + decay = self.decay + + if self.num_updates >= 0: + self.num_updates += 1 + decay = min(self.decay,(1 + self.num_updates) / (10 + self.num_updates)) + + one_minus_decay = 1.0 - decay + + with torch.no_grad(): + m_param = dict(model.named_parameters()) + shadow_params = dict(self.named_buffers()) + + for key in m_param: + if m_param[key].requires_grad: + sname = self.m_name2s_name[key] + shadow_params[sname] = shadow_params[sname].type_as(m_param[key]) + shadow_params[sname].sub_(one_minus_decay * (shadow_params[sname] - m_param[key])) + else: + assert not key in self.m_name2s_name + + def copy_to(self, model): + m_param = dict(model.named_parameters()) + shadow_params = dict(self.named_buffers()) + for key in m_param: + if m_param[key].requires_grad: + m_param[key].data.copy_(shadow_params[self.m_name2s_name[key]].data) + else: + assert not key in self.m_name2s_name + + def store(self, parameters): + """ + Save the current parameters for restoring later. + Args: + parameters: Iterable of `torch.nn.Parameter`; the parameters to be + temporarily stored. + """ + self.collected_params = [param.clone() for param in parameters] + + def restore(self, parameters): + """ + Restore the parameters stored with the `store` method. + Useful to validate the model with EMA parameters without affecting the + original optimization process. Store the parameters before the + `copy_to` method. After validation (or model saving), use this to + restore the former parameters. + Args: + parameters: Iterable of `torch.nn.Parameter`; the parameters to be + updated with the stored parameters. + """ + for c_param, param in zip(self.collected_params, parameters): + param.data.copy_(c_param.data) diff --git a/ldm/modules/encoders/.DS_Store b/ldm/modules/encoders/.DS_Store new file mode 100644 index 0000000000000000000000000000000000000000..5008ddfcf53c02e82d7eee2e57c38e5672ef89f6 Binary files /dev/null and b/ldm/modules/encoders/.DS_Store differ diff --git a/ldm/modules/encoders/__init__.py b/ldm/modules/encoders/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/ldm/modules/encoders/__pycache__/__init__.cpython-38.pyc b/ldm/modules/encoders/__pycache__/__init__.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..af50b6c23ef3b34615cb211f158875194ef470ba Binary files /dev/null and b/ldm/modules/encoders/__pycache__/__init__.cpython-38.pyc differ diff --git a/ldm/modules/encoders/__pycache__/modules.cpython-38.pyc b/ldm/modules/encoders/__pycache__/modules.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..d9914b6dbd187849010733c5c5cee37a18f322c5 Binary files /dev/null and b/ldm/modules/encoders/__pycache__/modules.cpython-38.pyc differ diff --git a/ldm/modules/encoders/__pycache__/xf.cpython-38.pyc b/ldm/modules/encoders/__pycache__/xf.cpython-38.pyc new file mode 100644 index 0000000000000000000000000000000000000000..53d75d90e87f3583acb24d8f31c88340974d12a4 Binary files /dev/null and b/ldm/modules/encoders/__pycache__/xf.cpython-38.pyc differ diff --git a/ldm/modules/encoders/modules.py b/ldm/modules/encoders/modules.py new file mode 100644 index 0000000000000000000000000000000000000000..2e211aee1c8129d9a56e7a253c25a9c9db893d13 --- /dev/null +++ b/ldm/modules/encoders/modules.py @@ -0,0 +1,232 @@ +import torch +import torch.nn as nn +from functools import partial +import clip +from einops import rearrange, repeat +from transformers import CLIPTokenizer, CLIPTextModel,CLIPVisionModel,CLIPModel +import kornia +from ldm.modules.x_transformer import Encoder, TransformerWrapper # TODO: can we directly rely on lucidrains code and simply add this as a reuirement? --> test +from .xf import LayerNorm, Transformer +import math + +class AbstractEncoder(nn.Module): + def __init__(self): + super().__init__() + + def encode(self, *args, **kwargs): + raise NotImplementedError + + +class ClassEmbedder(nn.Module): + def __init__(self, embed_dim, n_classes=1000, key='class'): + super().__init__() + self.key = key + self.embedding = nn.Embedding(n_classes, embed_dim) + + def forward(self, batch, key=None): + if key is None: + key = self.key + # this is for use in crossattn + c = batch[key][:, None] + c = self.embedding(c) + return c + + +class TransformerEmbedder(AbstractEncoder): + """Some transformer encoder layers""" + def __init__(self, n_embed, n_layer, vocab_size, max_seq_len=77, device="cuda"): + super().__init__() + self.device = device + self.transformer = TransformerWrapper(num_tokens=vocab_size, max_seq_len=max_seq_len, + attn_layers=Encoder(dim=n_embed, depth=n_layer)) + + def forward(self, tokens): + tokens = tokens.to(self.device) # meh + z = self.transformer(tokens, return_embeddings=True) + return z + + def encode(self, x): + return self(x) + + +class BERTTokenizer(AbstractEncoder): + """ Uses a pretrained BERT tokenizer by huggingface. Vocab size: 30522 (?)""" + def __init__(self, device="cuda", vq_interface=True, max_length=77): + super().__init__() + from transformers import BertTokenizerFast # TODO: add to reuquirements + self.tokenizer = BertTokenizerFast.from_pretrained("bert-base-uncased") + self.device = device + self.vq_interface = vq_interface + self.max_length = max_length + + def forward(self, text): + batch_encoding = self.tokenizer(text, truncation=True, max_length=self.max_length, return_length=True, + return_overflowing_tokens=False, padding="max_length", return_tensors="pt") + tokens = batch_encoding["input_ids"].to(self.device) + return tokens + + @torch.no_grad() + def encode(self, text): + tokens = self(text) + if not self.vq_interface: + return tokens + return None, None, [None, None, tokens] + + def decode(self, text): + return text + + +class BERTEmbedder(AbstractEncoder): + """Uses the BERT tokenizr model and add some transformer encoder layers""" + def __init__(self, n_embed, n_layer, vocab_size=30522, max_seq_len=77, + device="cuda",use_tokenizer=True, embedding_dropout=0.0): + super().__init__() + self.use_tknz_fn = use_tokenizer + if self.use_tknz_fn: + self.tknz_fn = BERTTokenizer(vq_interface=False, max_length=max_seq_len) + self.device = device + self.transformer = TransformerWrapper(num_tokens=vocab_size, max_seq_len=max_seq_len, + attn_layers=Encoder(dim=n_embed, depth=n_layer), + emb_dropout=embedding_dropout) + + def forward(self, text): + if self.use_tknz_fn: + tokens = self.tknz_fn(text)#.to(self.device) + else: + tokens = text + z = self.transformer(tokens, return_embeddings=True) + return z + + def encode(self, text): + # output of length 77 + return self(text) + + +class SpatialRescaler(nn.Module): + def __init__(self, + n_stages=1, + method='bilinear', + multiplier=0.5, + in_channels=3, + out_channels=None, + bias=False): + super().__init__() + self.n_stages = n_stages + assert self.n_stages >= 0 + assert method in ['nearest','linear','bilinear','trilinear','bicubic','area'] + self.multiplier = multiplier + self.interpolator = partial(torch.nn.functional.interpolate, mode=method) + self.remap_output = out_channels is not None + if self.remap_output: + print(f'Spatial Rescaler mapping from {in_channels} to {out_channels} channels after resizing.') + self.channel_mapper = nn.Conv2d(in_channels,out_channels,1,bias=bias) + + def forward(self,x): + for stage in range(self.n_stages): + x = self.interpolator(x, scale_factor=self.multiplier) + + + if self.remap_output: + x = self.channel_mapper(x) + return x + + def encode(self, x): + return self(x) + + +class FrozenCLIPImageEmbedder(AbstractEncoder): + """Uses the CLIP transformer encoder for image (from Hugging Face)""" + def __init__(self, version="openai/clip-vit-large-patch14"): + super().__init__() + self.transformer = CLIPVisionModel.from_pretrained(version) + self.final_ln = LayerNorm(1024) + self.mapper = Transformer( + 1, + 1024, + 5, + 1, + ) + + self.freeze() + + def freeze(self): + self.transformer = self.transformer.eval() + for param in self.parameters(): + param.requires_grad = False + for param in self.mapper.parameters(): + param.requires_grad = True + for param in self.final_ln.parameters(): + param.requires_grad = True + + def forward(self, image): + image = image.to('cuda') + outputs = self.transformer(pixel_values=image) + z = outputs.pooler_output + z = z.unsqueeze(1) + z = self.mapper(z) + z = self.final_ln(z) + return z + + def encode(self, image): + return self(image) + + +class FrozenCLIPTextEmbedder(AbstractEncoder): + """ + Uses the CLIP transformer encoder for text (from Hugging Face) + """ + def __init__(self, version="openai/clip-vit-large-patch14"): + super().__init__() + + self.tokenizer = CLIPTokenizer.from_pretrained(version) + self.text_model = CLIPTextModel.from_pretrained(version) + + #up d_model 1024 to concat with ImageEmbedding + # self.linear_proj = nn.Linear(768, 1024) + self.final_ln = nn.LayerNorm(768) + encoder_layer = nn.TransformerEncoderLayer( + d_model=768, + nhead=8, + dim_feedforward=2048, + batch_first=True + ) + self.mapper = nn.TransformerEncoder(encoder_layer, num_layers=2) + self.device = torch.device("cuda") if torch.cuda.is_available() else torch.device("cpu") + + self.freeze() + + def freeze(self): + """ + Freezes the transformer weights while keeping + the mapper and final layer normalization trainable. + """ + self.text_model = self.text_model.eval() + for param in self.parameters(): + param.requires_grad = False + for param in self.mapper.parameters(): + param.requires_grad = True + for param in self.final_ln.parameters(): + param.requires_grad = True + # self.linear_proj.requires_grad = True + + def forward(self, text): + """ + Encodes text using the tokenizer and transformer. + Applies mapper and final layer normalization for processing. + """ + inputs = self.tokenizer(text, return_tensors="pt", padding=True, truncation=True) + inputs = {k: v.to('cuda') for k, v in inputs.items()} + outputs = self.text_model(**inputs) + + z = outputs.pooler_output + # z = self.linear_proj(z) #768 -> 1024 + z = z.unsqueeze(1) + z = self.mapper(z) + z = self.final_ln(z) + + return z + + def encode(self, text): + return self(text) + + diff --git a/ldm/modules/encoders/xf.py b/ldm/modules/encoders/xf.py new file mode 100644 index 0000000000000000000000000000000000000000..e7481451c0e78a6542bf65e3c842310b6f7c51e1 --- /dev/null +++ b/ldm/modules/encoders/xf.py @@ -0,0 +1,130 @@ +""" +Transformer implementation adapted from CLIP ViT: +https://github.com/openai/CLIP/blob/4c0275784d6d9da97ca1f47eaaee31de1867da91/clip/model.py +""" + +import math + +import torch as th +import torch.nn as nn + + +def convert_module_to_f16(l): + """ + Convert primitive modules to float16. + """ + if isinstance(l, (nn.Linear, nn.Conv2d, nn.ConvTranspose2d)): + l.weight.data = l.weight.data.half() + if l.bias is not None: + l.bias.data = l.bias.data.half() + + +class LayerNorm(nn.LayerNorm): + """ + Implementation that supports fp16 inputs but fp32 gains/biases. + """ + + def forward(self, x: th.Tensor): + return super().forward(x.float()).to(x.dtype) + + +class MultiheadAttention(nn.Module): + def __init__(self, n_ctx, width, heads): + super().__init__() + self.n_ctx = n_ctx + self.width = width + self.heads = heads + self.c_qkv = nn.Linear(width, width * 3) + self.c_proj = nn.Linear(width, width) + self.attention = QKVMultiheadAttention(heads, n_ctx) + + def forward(self, x): + x = self.c_qkv(x) + x = self.attention(x) + x = self.c_proj(x) + return x + + +class MLP(nn.Module): + def __init__(self, width): + super().__init__() + self.width = width + self.c_fc = nn.Linear(width, width * 4) + self.c_proj = nn.Linear(width * 4, width) + self.gelu = nn.GELU() + + def forward(self, x): + return self.c_proj(self.gelu(self.c_fc(x))) + + +class QKVMultiheadAttention(nn.Module): + def __init__(self, n_heads: int, n_ctx: int): + super().__init__() + self.n_heads = n_heads + self.n_ctx = n_ctx + + def forward(self, qkv): + bs, n_ctx, width = qkv.shape + attn_ch = width // self.n_heads // 3 + scale = 1 / math.sqrt(math.sqrt(attn_ch)) + qkv = qkv.view(bs, n_ctx, self.n_heads, -1) + q, k, v = th.split(qkv, attn_ch, dim=-1) + weight = th.einsum( + "bthc,bshc->bhts", q * scale, k * scale + ) # More stable with f16 than dividing afterwards + wdtype = weight.dtype + weight = th.softmax(weight.float(), dim=-1).type(wdtype) + return th.einsum("bhts,bshc->bthc", weight, v).reshape(bs, n_ctx, -1) + + +class ResidualAttentionBlock(nn.Module): + def __init__( + self, + n_ctx: int, + width: int, + heads: int, + ): + super().__init__() + + self.attn = MultiheadAttention( + n_ctx, + width, + heads, + ) + self.ln_1 = LayerNorm(width) + self.mlp = MLP(width) + self.ln_2 = LayerNorm(width) + + def forward(self, x: th.Tensor): + x = x + self.attn(self.ln_1(x)) + x = x + self.mlp(self.ln_2(x)) + return x + + +class Transformer(nn.Module): + def __init__( + self, + n_ctx: int, + width: int, + layers: int, + heads: int, + ): + super().__init__() + self.n_ctx = n_ctx + self.width = width + self.layers = layers + self.resblocks = nn.ModuleList( + [ + ResidualAttentionBlock( + n_ctx, + width, + heads, + ) + for _ in range(layers) + ] + ) + + def forward(self, x: th.Tensor): + for block in self.resblocks: + x = block(x) + return x diff --git a/ldm/modules/losses/contperceptual.py b/ldm/modules/losses/contperceptual.py new file mode 100644 index 0000000000000000000000000000000000000000..ad1b1abd93a3721605331d96e2010b65f5e853a2 --- /dev/null +++ b/ldm/modules/losses/contperceptual.py @@ -0,0 +1,111 @@ +import torch +import torch.nn as nn + +from taming.modules.losses.vqperceptual import * # TODO: taming dependency yes/no? + + +class LPIPSWithDiscriminator(nn.Module): + def __init__(self, disc_start, logvar_init=0.0, kl_weight=1.0, pixelloss_weight=1.0, + disc_num_layers=3, disc_in_channels=3, disc_factor=1.0, disc_weight=1.0, + perceptual_weight=1.0, use_actnorm=False, disc_conditional=False, + disc_loss="hinge"): + + super().__init__() + assert disc_loss in ["hinge", "vanilla"] + self.kl_weight = kl_weight + self.pixel_weight = pixelloss_weight + self.perceptual_loss = LPIPS().eval() + self.perceptual_weight = perceptual_weight + # output log variance + self.logvar = nn.Parameter(torch.ones(size=()) * logvar_init) + + self.discriminator = NLayerDiscriminator(input_nc=disc_in_channels, + n_layers=disc_num_layers, + use_actnorm=use_actnorm + ).apply(weights_init) + self.discriminator_iter_start = disc_start + self.disc_loss = hinge_d_loss if disc_loss == "hinge" else vanilla_d_loss + self.disc_factor = disc_factor + self.discriminator_weight = disc_weight + self.disc_conditional = disc_conditional + + def calculate_adaptive_weight(self, nll_loss, g_loss, last_layer=None): + if last_layer is not None: + nll_grads = torch.autograd.grad(nll_loss, last_layer, retain_graph=True)[0] + g_grads = torch.autograd.grad(g_loss, last_layer, retain_graph=True)[0] + else: + nll_grads = torch.autograd.grad(nll_loss, self.last_layer[0], retain_graph=True)[0] + g_grads = torch.autograd.grad(g_loss, self.last_layer[0], retain_graph=True)[0] + + d_weight = torch.norm(nll_grads) / (torch.norm(g_grads) + 1e-4) + d_weight = torch.clamp(d_weight, 0.0, 1e4).detach() + d_weight = d_weight * self.discriminator_weight + return d_weight + + def forward(self, inputs, reconstructions, posteriors, optimizer_idx, + global_step, last_layer=None, cond=None, split="train", + weights=None): + rec_loss = torch.abs(inputs.contiguous() - reconstructions.contiguous()) + if self.perceptual_weight > 0: + p_loss = self.perceptual_loss(inputs.contiguous(), reconstructions.contiguous()) + rec_loss = rec_loss + self.perceptual_weight * p_loss + + nll_loss = rec_loss / torch.exp(self.logvar) + self.logvar + weighted_nll_loss = nll_loss + if weights is not None: + weighted_nll_loss = weights*nll_loss + weighted_nll_loss = torch.sum(weighted_nll_loss) / weighted_nll_loss.shape[0] + nll_loss = torch.sum(nll_loss) / nll_loss.shape[0] + kl_loss = posteriors.kl() + kl_loss = torch.sum(kl_loss) / kl_loss.shape[0] + + # now the GAN part + if optimizer_idx == 0: + # generator update + if cond is None: + assert not self.disc_conditional + logits_fake = self.discriminator(reconstructions.contiguous()) + else: + assert self.disc_conditional + logits_fake = self.discriminator(torch.cat((reconstructions.contiguous(), cond), dim=1)) + g_loss = -torch.mean(logits_fake) + + if self.disc_factor > 0.0: + try: + d_weight = self.calculate_adaptive_weight(nll_loss, g_loss, last_layer=last_layer) + except RuntimeError: + assert not self.training + d_weight = torch.tensor(0.0) + else: + d_weight = torch.tensor(0.0) + + disc_factor = adopt_weight(self.disc_factor, global_step, threshold=self.discriminator_iter_start) + loss = weighted_nll_loss + self.kl_weight * kl_loss + d_weight * disc_factor * g_loss + + log = {"{}/total_loss".format(split): loss.clone().detach().mean(), "{}/logvar".format(split): self.logvar.detach(), + "{}/kl_loss".format(split): kl_loss.detach().mean(), "{}/nll_loss".format(split): nll_loss.detach().mean(), + "{}/rec_loss".format(split): rec_loss.detach().mean(), + "{}/d_weight".format(split): d_weight.detach(), + "{}/disc_factor".format(split): torch.tensor(disc_factor), + "{}/g_loss".format(split): g_loss.detach().mean(), + } + return loss, log + + if optimizer_idx == 1: + # second pass for discriminator update + if cond is None: + logits_real = self.discriminator(inputs.contiguous().detach()) + logits_fake = self.discriminator(reconstructions.contiguous().detach()) + else: + logits_real = self.discriminator(torch.cat((inputs.contiguous().detach(), cond), dim=1)) + logits_fake = self.discriminator(torch.cat((reconstructions.contiguous().detach(), cond), dim=1)) + + disc_factor = adopt_weight(self.disc_factor, global_step, threshold=self.discriminator_iter_start) + d_loss = disc_factor * self.disc_loss(logits_real, logits_fake) + + log = {"{}/disc_loss".format(split): d_loss.clone().detach().mean(), + "{}/logits_real".format(split): logits_real.detach().mean(), + "{}/logits_fake".format(split): logits_fake.detach().mean() + } + return d_loss, log + diff --git a/ldm/modules/losses/vqperceptual.py b/ldm/modules/losses/vqperceptual.py new file mode 100644 index 0000000000000000000000000000000000000000..aba4acddf455d162a4a0b091a7cc5f82a920789c --- /dev/null +++ b/ldm/modules/losses/vqperceptual.py @@ -0,0 +1,167 @@ +import torch +from torch import nn +import torch.nn.functional as F +from einops import repeat + +from taming.modules.discriminator.model import NLayerDiscriminator, weights_init +from taming.modules.losses.lpips import LPIPS +from taming.modules.losses.vqperceptual import hinge_d_loss, vanilla_d_loss + + +def hinge_d_loss_with_exemplar_weights(logits_real, logits_fake, weights): + assert weights.shape[0] == logits_real.shape[0] == logits_fake.shape[0] + loss_real = torch.mean(F.relu(1. - logits_real), dim=[1,2,3]) + loss_fake = torch.mean(F.relu(1. + logits_fake), dim=[1,2,3]) + loss_real = (weights * loss_real).sum() / weights.sum() + loss_fake = (weights * loss_fake).sum() / weights.sum() + d_loss = 0.5 * (loss_real + loss_fake) + return d_loss + +def adopt_weight(weight, global_step, threshold=0, value=0.): + if global_step < threshold: + weight = value + return weight + + +def measure_perplexity(predicted_indices, n_embed): + # src: https://github.com/karpathy/deep-vector-quantization/blob/main/model.py + # eval cluster perplexity. when perplexity == num_embeddings then all clusters are used exactly equally + encodings = F.one_hot(predicted_indices, n_embed).float().reshape(-1, n_embed) + avg_probs = encodings.mean(0) + perplexity = (-(avg_probs * torch.log(avg_probs + 1e-10)).sum()).exp() + cluster_use = torch.sum(avg_probs > 0) + return perplexity, cluster_use + +def l1(x, y): + return torch.abs(x-y) + + +def l2(x, y): + return torch.pow((x-y), 2) + + +class VQLPIPSWithDiscriminator(nn.Module): + def __init__(self, disc_start, codebook_weight=1.0, pixelloss_weight=1.0, + disc_num_layers=3, disc_in_channels=3, disc_factor=1.0, disc_weight=1.0, + perceptual_weight=1.0, use_actnorm=False, disc_conditional=False, + disc_ndf=64, disc_loss="hinge", n_classes=None, perceptual_loss="lpips", + pixel_loss="l1"): + super().__init__() + assert disc_loss in ["hinge", "vanilla"] + assert perceptual_loss in ["lpips", "clips", "dists"] + assert pixel_loss in ["l1", "l2"] + self.codebook_weight = codebook_weight + self.pixel_weight = pixelloss_weight + if perceptual_loss == "lpips": + print(f"{self.__class__.__name__}: Running with LPIPS.") + self.perceptual_loss = LPIPS().eval() + else: + raise ValueError(f"Unknown perceptual loss: >> {perceptual_loss} <<") + self.perceptual_weight = perceptual_weight + + if pixel_loss == "l1": + self.pixel_loss = l1 + else: + self.pixel_loss = l2 + + self.discriminator = NLayerDiscriminator(input_nc=disc_in_channels, + n_layers=disc_num_layers, + use_actnorm=use_actnorm, + ndf=disc_ndf + ).apply(weights_init) + self.discriminator_iter_start = disc_start + if disc_loss == "hinge": + self.disc_loss = hinge_d_loss + elif disc_loss == "vanilla": + self.disc_loss = vanilla_d_loss + else: + raise ValueError(f"Unknown GAN loss '{disc_loss}'.") + print(f"VQLPIPSWithDiscriminator running with {disc_loss} loss.") + self.disc_factor = disc_factor + self.discriminator_weight = disc_weight + self.disc_conditional = disc_conditional + self.n_classes = n_classes + + def calculate_adaptive_weight(self, nll_loss, g_loss, last_layer=None): + if last_layer is not None: + nll_grads = torch.autograd.grad(nll_loss, last_layer, retain_graph=True)[0] + g_grads = torch.autograd.grad(g_loss, last_layer, retain_graph=True)[0] + else: + nll_grads = torch.autograd.grad(nll_loss, self.last_layer[0], retain_graph=True)[0] + g_grads = torch.autograd.grad(g_loss, self.last_layer[0], retain_graph=True)[0] + + d_weight = torch.norm(nll_grads) / (torch.norm(g_grads) + 1e-4) + d_weight = torch.clamp(d_weight, 0.0, 1e4).detach() + d_weight = d_weight * self.discriminator_weight + return d_weight + + def forward(self, codebook_loss, inputs, reconstructions, optimizer_idx, + global_step, last_layer=None, cond=None, split="train", predicted_indices=None): + if not exists(codebook_loss): + codebook_loss = torch.tensor([0.]).to(inputs.device) + #rec_loss = torch.abs(inputs.contiguous() - reconstructions.contiguous()) + rec_loss = self.pixel_loss(inputs.contiguous(), reconstructions.contiguous()) + if self.perceptual_weight > 0: + p_loss = self.perceptual_loss(inputs.contiguous(), reconstructions.contiguous()) + rec_loss = rec_loss + self.perceptual_weight * p_loss + else: + p_loss = torch.tensor([0.0]) + + nll_loss = rec_loss + #nll_loss = torch.sum(nll_loss) / nll_loss.shape[0] + nll_loss = torch.mean(nll_loss) + + # now the GAN part + if optimizer_idx == 0: + # generator update + if cond is None: + assert not self.disc_conditional + logits_fake = self.discriminator(reconstructions.contiguous()) + else: + assert self.disc_conditional + logits_fake = self.discriminator(torch.cat((reconstructions.contiguous(), cond), dim=1)) + g_loss = -torch.mean(logits_fake) + + try: + d_weight = self.calculate_adaptive_weight(nll_loss, g_loss, last_layer=last_layer) + except RuntimeError: + assert not self.training + d_weight = torch.tensor(0.0) + + disc_factor = adopt_weight(self.disc_factor, global_step, threshold=self.discriminator_iter_start) + loss = nll_loss + d_weight * disc_factor * g_loss + self.codebook_weight * codebook_loss.mean() + + log = {"{}/total_loss".format(split): loss.clone().detach().mean(), + "{}/quant_loss".format(split): codebook_loss.detach().mean(), + "{}/nll_loss".format(split): nll_loss.detach().mean(), + "{}/rec_loss".format(split): rec_loss.detach().mean(), + "{}/p_loss".format(split): p_loss.detach().mean(), + "{}/d_weight".format(split): d_weight.detach(), + "{}/disc_factor".format(split): torch.tensor(disc_factor), + "{}/g_loss".format(split): g_loss.detach().mean(), + } + if predicted_indices is not None: + assert self.n_classes is not None + with torch.no_grad(): + perplexity, cluster_usage = measure_perplexity(predicted_indices, self.n_classes) + log[f"{split}/perplexity"] = perplexity + log[f"{split}/cluster_usage"] = cluster_usage + return loss, log + + if optimizer_idx == 1: + # second pass for discriminator update + if cond is None: + logits_real = self.discriminator(inputs.contiguous().detach()) + logits_fake = self.discriminator(reconstructions.contiguous().detach()) + else: + logits_real = self.discriminator(torch.cat((inputs.contiguous().detach(), cond), dim=1)) + logits_fake = self.discriminator(torch.cat((reconstructions.contiguous().detach(), cond), dim=1)) + + disc_factor = adopt_weight(self.disc_factor, global_step, threshold=self.discriminator_iter_start) + d_loss = disc_factor * self.disc_loss(logits_real, logits_fake) + + log = {"{}/disc_loss".format(split): d_loss.clone().detach().mean(), + "{}/logits_real".format(split): logits_real.detach().mean(), + "{}/logits_fake".format(split): logits_fake.detach().mean() + } + return d_loss, log diff --git a/ldm/modules/x_transformer.py b/ldm/modules/x_transformer.py new file mode 100644 index 0000000000000000000000000000000000000000..0c7a4d4f2bd89ce015bc7896e73b55772e508a66 --- /dev/null +++ b/ldm/modules/x_transformer.py @@ -0,0 +1,641 @@ +"""shout-out to https://github.com/lucidrains/x-transformers/tree/main/x_transformers""" +import torch +from torch import nn, einsum +import torch.nn.functional as F +from functools import partial +from inspect import isfunction +from collections import namedtuple +from einops import rearrange, repeat, reduce + +# constants + +DEFAULT_DIM_HEAD = 64 + +Intermediates = namedtuple('Intermediates', [ + 'pre_softmax_attn', + 'post_softmax_attn' +]) + +LayerIntermediates = namedtuple('Intermediates', [ + 'hiddens', + 'attn_intermediates' +]) + + +class AbsolutePositionalEmbedding(nn.Module): + def __init__(self, dim, max_seq_len): + super().__init__() + self.emb = nn.Embedding(max_seq_len, dim) + self.init_() + + def init_(self): + nn.init.normal_(self.emb.weight, std=0.02) + + def forward(self, x): + n = torch.arange(x.shape[1], device=x.device) + return self.emb(n)[None, :, :] + + +class FixedPositionalEmbedding(nn.Module): + def __init__(self, dim): + super().__init__() + inv_freq = 1. / (10000 ** (torch.arange(0, dim, 2).float() / dim)) + self.register_buffer('inv_freq', inv_freq) + + def forward(self, x, seq_dim=1, offset=0): + t = torch.arange(x.shape[seq_dim], device=x.device).type_as(self.inv_freq) + offset + sinusoid_inp = torch.einsum('i , j -> i j', t, self.inv_freq) + emb = torch.cat((sinusoid_inp.sin(), sinusoid_inp.cos()), dim=-1) + return emb[None, :, :] + + +# helpers + +def exists(val): + return val is not None + + +def default(val, d): + if exists(val): + return val + return d() if isfunction(d) else d + + +def always(val): + def inner(*args, **kwargs): + return val + return inner + + +def not_equals(val): + def inner(x): + return x != val + return inner + + +def equals(val): + def inner(x): + return x == val + return inner + + +def max_neg_value(tensor): + return -torch.finfo(tensor.dtype).max + + +# keyword argument helpers + +def pick_and_pop(keys, d): + values = list(map(lambda key: d.pop(key), keys)) + return dict(zip(keys, values)) + + +def group_dict_by_key(cond, d): + return_val = [dict(), dict()] + for key in d.keys(): + match = bool(cond(key)) + ind = int(not match) + return_val[ind][key] = d[key] + return (*return_val,) + + +def string_begins_with(prefix, str): + return str.startswith(prefix) + + +def group_by_key_prefix(prefix, d): + return group_dict_by_key(partial(string_begins_with, prefix), d) + + +def groupby_prefix_and_trim(prefix, d): + kwargs_with_prefix, kwargs = group_dict_by_key(partial(string_begins_with, prefix), d) + kwargs_without_prefix = dict(map(lambda x: (x[0][len(prefix):], x[1]), tuple(kwargs_with_prefix.items()))) + return kwargs_without_prefix, kwargs + + +# classes +class Scale(nn.Module): + def __init__(self, value, fn): + super().__init__() + self.value = value + self.fn = fn + + def forward(self, x, **kwargs): + x, *rest = self.fn(x, **kwargs) + return (x * self.value, *rest) + + +class Rezero(nn.Module): + def __init__(self, fn): + super().__init__() + self.fn = fn + self.g = nn.Parameter(torch.zeros(1)) + + def forward(self, x, **kwargs): + x, *rest = self.fn(x, **kwargs) + return (x * self.g, *rest) + + +class ScaleNorm(nn.Module): + def __init__(self, dim, eps=1e-5): + super().__init__() + self.scale = dim ** -0.5 + self.eps = eps + self.g = nn.Parameter(torch.ones(1)) + + def forward(self, x): + norm = torch.norm(x, dim=-1, keepdim=True) * self.scale + return x / norm.clamp(min=self.eps) * self.g + + +class RMSNorm(nn.Module): + def __init__(self, dim, eps=1e-8): + super().__init__() + self.scale = dim ** -0.5 + self.eps = eps + self.g = nn.Parameter(torch.ones(dim)) + + def forward(self, x): + norm = torch.norm(x, dim=-1, keepdim=True) * self.scale + return x / norm.clamp(min=self.eps) * self.g + + +class Residual(nn.Module): + def forward(self, x, residual): + return x + residual + + +class GRUGating(nn.Module): + def __init__(self, dim): + super().__init__() + self.gru = nn.GRUCell(dim, dim) + + def forward(self, x, residual): + gated_output = self.gru( + rearrange(x, 'b n d -> (b n) d'), + rearrange(residual, 'b n d -> (b n) d') + ) + + return gated_output.reshape_as(x) + + +# feedforward + +class GEGLU(nn.Module): + def __init__(self, dim_in, dim_out): + super().__init__() + self.proj = nn.Linear(dim_in, dim_out * 2) + + def forward(self, x): + x, gate = self.proj(x).chunk(2, dim=-1) + return x * F.gelu(gate) + + +class FeedForward(nn.Module): + def __init__(self, dim, dim_out=None, mult=4, glu=False, dropout=0.): + super().__init__() + inner_dim = int(dim * mult) + dim_out = default(dim_out, dim) + project_in = nn.Sequential( + nn.Linear(dim, inner_dim), + nn.GELU() + ) if not glu else GEGLU(dim, inner_dim) + + self.net = nn.Sequential( + project_in, + nn.Dropout(dropout), + nn.Linear(inner_dim, dim_out) + ) + + def forward(self, x): + return self.net(x) + + +# attention. +class Attention(nn.Module): + def __init__( + self, + dim, + dim_head=DEFAULT_DIM_HEAD, + heads=8, + causal=False, + mask=None, + talking_heads=False, + sparse_topk=None, + use_entmax15=False, + num_mem_kv=0, + dropout=0., + on_attn=False + ): + super().__init__() + if use_entmax15: + raise NotImplementedError("Check out entmax activation instead of softmax activation!") + self.scale = dim_head ** -0.5 + self.heads = heads + self.causal = causal + self.mask = mask + + inner_dim = dim_head * heads + + self.to_q = nn.Linear(dim, inner_dim, bias=False) + self.to_k = nn.Linear(dim, inner_dim, bias=False) + self.to_v = nn.Linear(dim, inner_dim, bias=False) + self.dropout = nn.Dropout(dropout) + + # talking heads + self.talking_heads = talking_heads + if talking_heads: + self.pre_softmax_proj = nn.Parameter(torch.randn(heads, heads)) + self.post_softmax_proj = nn.Parameter(torch.randn(heads, heads)) + + # explicit topk sparse attention + self.sparse_topk = sparse_topk + + # entmax + #self.attn_fn = entmax15 if use_entmax15 else F.softmax + self.attn_fn = F.softmax + + # add memory key / values + self.num_mem_kv = num_mem_kv + if num_mem_kv > 0: + self.mem_k = nn.Parameter(torch.randn(heads, num_mem_kv, dim_head)) + self.mem_v = nn.Parameter(torch.randn(heads, num_mem_kv, dim_head)) + + # attention on attention + self.attn_on_attn = on_attn + self.to_out = nn.Sequential(nn.Linear(inner_dim, dim * 2), nn.GLU()) if on_attn else nn.Linear(inner_dim, dim) + + def forward( + self, + x, + context=None, + mask=None, + context_mask=None, + rel_pos=None, + sinusoidal_emb=None, + prev_attn=None, + mem=None + ): + b, n, _, h, talking_heads, device = *x.shape, self.heads, self.talking_heads, x.device + kv_input = default(context, x) + + q_input = x + k_input = kv_input + v_input = kv_input + + if exists(mem): + k_input = torch.cat((mem, k_input), dim=-2) + v_input = torch.cat((mem, v_input), dim=-2) + + if exists(sinusoidal_emb): + # in shortformer, the query would start at a position offset depending on the past cached memory + offset = k_input.shape[-2] - q_input.shape[-2] + q_input = q_input + sinusoidal_emb(q_input, offset=offset) + k_input = k_input + sinusoidal_emb(k_input) + + q = self.to_q(q_input) + k = self.to_k(k_input) + v = self.to_v(v_input) + + q, k, v = map(lambda t: rearrange(t, 'b n (h d) -> b h n d', h=h), (q, k, v)) + + input_mask = None + if any(map(exists, (mask, context_mask))): + q_mask = default(mask, lambda: torch.ones((b, n), device=device).bool()) + k_mask = q_mask if not exists(context) else context_mask + k_mask = default(k_mask, lambda: torch.ones((b, k.shape[-2]), device=device).bool()) + q_mask = rearrange(q_mask, 'b i -> b () i ()') + k_mask = rearrange(k_mask, 'b j -> b () () j') + input_mask = q_mask * k_mask + + if self.num_mem_kv > 0: + mem_k, mem_v = map(lambda t: repeat(t, 'h n d -> b h n d', b=b), (self.mem_k, self.mem_v)) + k = torch.cat((mem_k, k), dim=-2) + v = torch.cat((mem_v, v), dim=-2) + if exists(input_mask): + input_mask = F.pad(input_mask, (self.num_mem_kv, 0), value=True) + + dots = einsum('b h i d, b h j d -> b h i j', q, k) * self.scale + mask_value = max_neg_value(dots) + + if exists(prev_attn): + dots = dots + prev_attn + + pre_softmax_attn = dots + + if talking_heads: + dots = einsum('b h i j, h k -> b k i j', dots, self.pre_softmax_proj).contiguous() + + if exists(rel_pos): + dots = rel_pos(dots) + + if exists(input_mask): + dots.masked_fill_(~input_mask, mask_value) + del input_mask + + if self.causal: + i, j = dots.shape[-2:] + r = torch.arange(i, device=device) + mask = rearrange(r, 'i -> () () i ()') < rearrange(r, 'j -> () () () j') + mask = F.pad(mask, (j - i, 0), value=False) + dots.masked_fill_(mask, mask_value) + del mask + + if exists(self.sparse_topk) and self.sparse_topk < dots.shape[-1]: + top, _ = dots.topk(self.sparse_topk, dim=-1) + vk = top[..., -1].unsqueeze(-1).expand_as(dots) + mask = dots < vk + dots.masked_fill_(mask, mask_value) + del mask + + attn = self.attn_fn(dots, dim=-1) + post_softmax_attn = attn + + attn = self.dropout(attn) + + if talking_heads: + attn = einsum('b h i j, h k -> b k i j', attn, self.post_softmax_proj).contiguous() + + out = einsum('b h i j, b h j d -> b h i d', attn, v) + out = rearrange(out, 'b h n d -> b n (h d)') + + intermediates = Intermediates( + pre_softmax_attn=pre_softmax_attn, + post_softmax_attn=post_softmax_attn + ) + + return self.to_out(out), intermediates + + +class AttentionLayers(nn.Module): + def __init__( + self, + dim, + depth, + heads=8, + causal=False, + cross_attend=False, + only_cross=False, + use_scalenorm=False, + use_rmsnorm=False, + use_rezero=False, + rel_pos_num_buckets=32, + rel_pos_max_distance=128, + position_infused_attn=False, + custom_layers=None, + sandwich_coef=None, + par_ratio=None, + residual_attn=False, + cross_residual_attn=False, + macaron=False, + pre_norm=True, + gate_residual=False, + **kwargs + ): + super().__init__() + ff_kwargs, kwargs = groupby_prefix_and_trim('ff_', kwargs) + attn_kwargs, _ = groupby_prefix_and_trim('attn_', kwargs) + + dim_head = attn_kwargs.get('dim_head', DEFAULT_DIM_HEAD) + + self.dim = dim + self.depth = depth + self.layers = nn.ModuleList([]) + + self.has_pos_emb = position_infused_attn + self.pia_pos_emb = FixedPositionalEmbedding(dim) if position_infused_attn else None + self.rotary_pos_emb = always(None) + + assert rel_pos_num_buckets <= rel_pos_max_distance, 'number of relative position buckets must be less than the relative position max distance' + self.rel_pos = None + + self.pre_norm = pre_norm + + self.residual_attn = residual_attn + self.cross_residual_attn = cross_residual_attn + + norm_class = ScaleNorm if use_scalenorm else nn.LayerNorm + norm_class = RMSNorm if use_rmsnorm else norm_class + norm_fn = partial(norm_class, dim) + + norm_fn = nn.Identity if use_rezero else norm_fn + branch_fn = Rezero if use_rezero else None + + if cross_attend and not only_cross: + default_block = ('a', 'c', 'f') + elif cross_attend and only_cross: + default_block = ('c', 'f') + else: + default_block = ('a', 'f') + + if macaron: + default_block = ('f',) + default_block + + if exists(custom_layers): + layer_types = custom_layers + elif exists(par_ratio): + par_depth = depth * len(default_block) + assert 1 < par_ratio <= par_depth, 'par ratio out of range' + default_block = tuple(filter(not_equals('f'), default_block)) + par_attn = par_depth // par_ratio + depth_cut = par_depth * 2 // 3 # 2 / 3 attention layer cutoff suggested by PAR paper + par_width = (depth_cut + depth_cut // par_attn) // par_attn + assert len(default_block) <= par_width, 'default block is too large for par_ratio' + par_block = default_block + ('f',) * (par_width - len(default_block)) + par_head = par_block * par_attn + layer_types = par_head + ('f',) * (par_depth - len(par_head)) + elif exists(sandwich_coef): + assert sandwich_coef > 0 and sandwich_coef <= depth, 'sandwich coefficient should be less than the depth' + layer_types = ('a',) * sandwich_coef + default_block * (depth - sandwich_coef) + ('f',) * sandwich_coef + else: + layer_types = default_block * depth + + self.layer_types = layer_types + self.num_attn_layers = len(list(filter(equals('a'), layer_types))) + + for layer_type in self.layer_types: + if layer_type == 'a': + layer = Attention(dim, heads=heads, causal=causal, **attn_kwargs) + elif layer_type == 'c': + layer = Attention(dim, heads=heads, **attn_kwargs) + elif layer_type == 'f': + layer = FeedForward(dim, **ff_kwargs) + layer = layer if not macaron else Scale(0.5, layer) + else: + raise Exception(f'invalid layer type {layer_type}') + + if isinstance(layer, Attention) and exists(branch_fn): + layer = branch_fn(layer) + + if gate_residual: + residual_fn = GRUGating(dim) + else: + residual_fn = Residual() + + self.layers.append(nn.ModuleList([ + norm_fn(), + layer, + residual_fn + ])) + + def forward( + self, + x, + context=None, + mask=None, + context_mask=None, + mems=None, + return_hiddens=False + ): + hiddens = [] + intermediates = [] + prev_attn = None + prev_cross_attn = None + + mems = mems.copy() if exists(mems) else [None] * self.num_attn_layers + + for ind, (layer_type, (norm, block, residual_fn)) in enumerate(zip(self.layer_types, self.layers)): + is_last = ind == (len(self.layers) - 1) + + if layer_type == 'a': + hiddens.append(x) + layer_mem = mems.pop(0) + + residual = x + + if self.pre_norm: + x = norm(x) + + if layer_type == 'a': + out, inter = block(x, mask=mask, sinusoidal_emb=self.pia_pos_emb, rel_pos=self.rel_pos, + prev_attn=prev_attn, mem=layer_mem) + elif layer_type == 'c': + out, inter = block(x, context=context, mask=mask, context_mask=context_mask, prev_attn=prev_cross_attn) + elif layer_type == 'f': + out = block(x) + + x = residual_fn(out, residual) + + if layer_type in ('a', 'c'): + intermediates.append(inter) + + if layer_type == 'a' and self.residual_attn: + prev_attn = inter.pre_softmax_attn + elif layer_type == 'c' and self.cross_residual_attn: + prev_cross_attn = inter.pre_softmax_attn + + if not self.pre_norm and not is_last: + x = norm(x) + + if return_hiddens: + intermediates = LayerIntermediates( + hiddens=hiddens, + attn_intermediates=intermediates + ) + + return x, intermediates + + return x + + +class Encoder(AttentionLayers): + def __init__(self, **kwargs): + assert 'causal' not in kwargs, 'cannot set causality on encoder' + super().__init__(causal=False, **kwargs) + + + +class TransformerWrapper(nn.Module): + def __init__( + self, + *, + num_tokens, + max_seq_len, + attn_layers, + emb_dim=None, + max_mem_len=0., + emb_dropout=0., + num_memory_tokens=None, + tie_embedding=False, + use_pos_emb=True + ): + super().__init__() + assert isinstance(attn_layers, AttentionLayers), 'attention layers must be one of Encoder or Decoder' + + dim = attn_layers.dim + emb_dim = default(emb_dim, dim) + + self.max_seq_len = max_seq_len + self.max_mem_len = max_mem_len + self.num_tokens = num_tokens + + self.token_emb = nn.Embedding(num_tokens, emb_dim) + self.pos_emb = AbsolutePositionalEmbedding(emb_dim, max_seq_len) if ( + use_pos_emb and not attn_layers.has_pos_emb) else always(0) + self.emb_dropout = nn.Dropout(emb_dropout) + + self.project_emb = nn.Linear(emb_dim, dim) if emb_dim != dim else nn.Identity() + self.attn_layers = attn_layers + self.norm = nn.LayerNorm(dim) + + self.init_() + + self.to_logits = nn.Linear(dim, num_tokens) if not tie_embedding else lambda t: t @ self.token_emb.weight.t() + + # memory tokens (like [cls]) from Memory Transformers paper + num_memory_tokens = default(num_memory_tokens, 0) + self.num_memory_tokens = num_memory_tokens + if num_memory_tokens > 0: + self.memory_tokens = nn.Parameter(torch.randn(num_memory_tokens, dim)) + + # let funnel encoder know number of memory tokens, if specified + if hasattr(attn_layers, 'num_memory_tokens'): + attn_layers.num_memory_tokens = num_memory_tokens + + def init_(self): + nn.init.normal_(self.token_emb.weight, std=0.02) + + def forward( + self, + x, + return_embeddings=False, + mask=None, + return_mems=False, + return_attn=False, + mems=None, + **kwargs + ): + b, n, device, num_mem = *x.shape, x.device, self.num_memory_tokens + x = self.token_emb(x) + x += self.pos_emb(x) + x = self.emb_dropout(x) + + x = self.project_emb(x) + + if num_mem > 0: + mem = repeat(self.memory_tokens, 'n d -> b n d', b=b) + x = torch.cat((mem, x), dim=1) + + # auto-handle masking after appending memory tokens + if exists(mask): + mask = F.pad(mask, (num_mem, 0), value=True) + + x, intermediates = self.attn_layers(x, mask=mask, mems=mems, return_hiddens=True, **kwargs) + x = self.norm(x) + + mem, x = x[:, :num_mem], x[:, num_mem:] + + out = self.to_logits(x) if not return_embeddings else x + + if return_mems: + hiddens = intermediates.hiddens + new_mems = list(map(lambda pair: torch.cat(pair, dim=-2), zip(mems, hiddens))) if exists(mems) else hiddens + new_mems = list(map(lambda t: t[..., -self.max_mem_len:, :].detach(), new_mems)) + return out, new_mems + + if return_attn: + attn_maps = list(map(lambda t: t.post_softmax_attn, intermediates.attn_intermediates)) + return out, attn_maps + + return out + diff --git a/ldm/util.py b/ldm/util.py new file mode 100644 index 0000000000000000000000000000000000000000..e2b9a02351a22cd63f8280434705e711425aa7c4 --- /dev/null +++ b/ldm/util.py @@ -0,0 +1,197 @@ +import importlib + +import torch +import numpy as np +from collections import abc +from einops import rearrange +from functools import partial + +import multiprocessing as mp +from threading import Thread +from queue import Queue + +from inspect import isfunction +from PIL import Image, ImageDraw, ImageFont + + +def log_txt_as_img(wh, xc, size=10): + # wh a tuple of (width, height) + # xc a list of captions to plot + b = len(xc) + txts = list() + for bi in range(b): + txt = Image.new("RGB", wh, color="white") + draw = ImageDraw.Draw(txt) + font = ImageFont.truetype('data/DejaVuSans.ttf', size=size) + nc = int(40 * (wh[0] / 256)) + lines = "\n".join(xc[bi][start:start + nc] for start in range(0, len(xc[bi]), nc)) + + try: + draw.text((0, 0), lines, fill="black", font=font) + except UnicodeEncodeError: + print("Cant encode string for logging. Skipping.") + + txt = np.array(txt).transpose(2, 0, 1) / 127.5 - 1.0 + txts.append(txt) + txts = np.stack(txts) + txts = torch.tensor(txts) + return txts + + +def ismap(x): + if not isinstance(x, torch.Tensor): + return False + return (len(x.shape) == 4) and (x.shape[1] > 3) + + +def isimage(x): + if not isinstance(x, torch.Tensor): + return False + return (len(x.shape) == 4) and (x.shape[1] == 3 or x.shape[1] == 1) + + +def exists(x): + return x is not None + + +def default(val, d): + if exists(val): + return val + return d() if isfunction(d) else d + + +def mean_flat(tensor): + """ + https://github.com/openai/guided-diffusion/blob/27c20a8fab9cb472df5d6bdd6c8d11c8f430b924/guided_diffusion/nn.py#L86 + Take the mean over all non-batch dimensions. + """ + return tensor.mean(dim=list(range(1, len(tensor.shape)))) + + +def count_params(model, verbose=False): + total_params = sum(p.numel() for p in model.parameters()) + if verbose: + print(f"{model.__class__.__name__} has {total_params * 1.e-6:.2f} M params.") + return total_params + + +def instantiate_from_config(config): + # config["target"] = "ldm.models.diffusion.control.ControlLDM" + # config["params"] = {"linear_start": 0.00085...} + return get_obj_from_str(config["target"])(**config.get("params", dict())) + +def get_obj_from_str(string): + # module = "ldm.models.diffusion.control" + # cls = "ControlLDM" + module, cls = string.rsplit(".", 1) + return getattr(importlib.import_module(module, package=None), cls) + + +def _do_parallel_data_prefetch(func, Q, data, idx, idx_to_fn=False): + # create dummy dataset instance + + # run prefetching + if idx_to_fn: + res = func(data, worker_id=idx) + else: + res = func(data) + Q.put([idx, res]) + Q.put("Done") + + +def parallel_data_prefetch( + func: callable, data, n_proc, target_data_type="ndarray", cpu_intensive=True, use_worker_id=False +): + # if target_data_type not in ["ndarray", "list"]: + # raise ValueError( + # "Data, which is passed to parallel_data_prefetch has to be either of type list or ndarray." + # ) + if isinstance(data, np.ndarray) and target_data_type == "list": + raise ValueError("list expected but function got ndarray.") + elif isinstance(data, abc.Iterable): + if isinstance(data, dict): + print( + f'WARNING:"data" argument passed to parallel_data_prefetch is a dict: Using only its values and disregarding keys.' + ) + data = list(data.values()) + if target_data_type == "ndarray": + data = np.asarray(data) + else: + data = list(data) + else: + raise TypeError( + f"The data, that shall be processed parallel has to be either an np.ndarray or an Iterable, but is actually {type(data)}." + ) + + if cpu_intensive: + Q = mp.Queue(1000) + proc = mp.Process + else: + Q = Queue(1000) + proc = Thread + # spawn processes + if target_data_type == "ndarray": + arguments = [ + [func, Q, part, i, use_worker_id] + for i, part in enumerate(np.array_split(data, n_proc)) + ] + else: + step = ( + int(len(data) / n_proc + 1) + if len(data) % n_proc != 0 + else int(len(data) / n_proc) + ) + arguments = [ + [func, Q, part, i, use_worker_id] + for i, part in enumerate( + [data[i: i + step] for i in range(0, len(data), step)] + ) + ] + processes = [] + for i in range(n_proc): + p = proc(target=_do_parallel_data_prefetch, args=arguments[i]) + processes += [p] + + # start processes + print(f"Start prefetching...") + import time + + start = time.time() + gather_res = [[] for _ in range(n_proc)] + try: + for p in processes: + p.start() + + k = 0 + while k < n_proc: + # get result + res = Q.get() + if res == "Done": + k += 1 + else: + gather_res[res[0]] = res[1] + + except Exception as e: + print("Exception: ", e) + for p in processes: + p.terminate() + + raise e + finally: + for p in processes: + p.join() + print(f"Prefetching complete. [{time.time() - start} sec.]") + + if target_data_type == 'ndarray': + if not isinstance(gather_res[0], np.ndarray): + return np.concatenate([np.asarray(r) for r in gather_res], axis=0) + + # order outputs + return np.concatenate(gather_res, axis=0) + elif target_data_type == 'list': + out = [] + for r in gather_res: + out.extend(r) + return out + else: + return gather_res