File size: 9,475 Bytes
94953a3 cb9cf0f 94953a3 cb9cf0f 94953a3 cb9cf0f 94953a3 cb9cf0f 94953a3 cb9cf0f 94953a3 cb9cf0f 94953a3 cb9cf0f 94953a3 cb9cf0f 94953a3 cb9cf0f 94953a3 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 |
# -*- coding: utf-8 -*-
# ===========================================================================================
#
# Copyright (c) Beijing Academy of Artificial Intelligence (BAAI). All rights reserved.
#
# Author : Fan Zhang
# Email : [email protected]
# Institute : Beijing Academy of Artificial Intelligence (BAAI)
# Create On : 2023-12-19 10:45
# Last Modified : 2023-12-25 07:59
# File Name : pipeline_emu2_gen.py
# Description :
#
# ===========================================================================================
from dataclasses import dataclass
from typing import List, Optional
from PIL import Image
import numpy as np
import torch
from torchvision import transforms as TF
from tqdm import tqdm
from diffusers import DiffusionPipeline
from diffusers.utils import BaseOutput
from diffusers import UNet2DConditionModel, EulerDiscreteScheduler, AutoencoderKL
from diffusers.pipelines.stable_diffusion.safety_checker import StableDiffusionSafetyChecker
from transformers import CLIPImageProcessor
from transformers import AutoModelForCausalLM, AutoTokenizer
EVA_IMAGE_SIZE = 448
OPENAI_DATASET_MEAN = (0.48145466, 0.4578275, 0.40821073)
OPENAI_DATASET_STD = (0.26862954, 0.26130258, 0.27577711)
DEFAULT_IMG_PLACEHOLDER = "[<IMG_PLH>]"
@dataclass
class EmuVisualGenerationPipelineOutput(BaseOutput):
image: Image.Image
nsfw_content_detected: Optional[bool]
class EmuVisualGenerationPipeline(DiffusionPipeline):
def __init__(
self,
tokenizer: AutoTokenizer,
multimodal_encoder: AutoModelForCausalLM,
scheduler: EulerDiscreteScheduler,
unet: UNet2DConditionModel,
vae: AutoencoderKL,
feature_extractor: CLIPImageProcessor,
safety_checker: StableDiffusionSafetyChecker,
eva_size=EVA_IMAGE_SIZE,
eva_mean=OPENAI_DATASET_MEAN,
eva_std=OPENAI_DATASET_STD,
):
super().__init__()
self.register_modules(
tokenizer=tokenizer,
multimodal_encoder=multimodal_encoder,
scheduler=scheduler,
unet=unet,
vae=vae,
feature_extractor=feature_extractor,
safety_checker=safety_checker,
)
self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
self.transform = TF.Compose([
TF.Resize((eva_size, eva_size), interpolation=TF.InterpolationMode.BICUBIC),
TF.ToTensor(),
TF.Normalize(mean=eva_mean, std=eva_std),
])
self.negative_prompt = {}
def device(self, module):
return next(module.parameters()).device
def dtype(self, module):
return next(module.parameters()).dtype
@torch.no_grad()
def __call__(
self,
inputs: List[Image.Image | str] | str | Image.Image,
height: int = 1024,
width: int = 1024,
num_inference_steps: int = 50,
guidance_scale: float = 3.,
crop_info: List[int] = [0, 0],
original_size: List[int] = [1024, 1024],
):
if not isinstance(inputs, list):
inputs = [inputs]
# 0. Default height and width to unet
height = height or self.unet.config.sample_size * self.vae_scale_factor
width = width or self.unet.config.sample_size * self.vae_scale_factor
device = self.device(self.unet)
dtype = self.dtype(self.unet)
do_classifier_free_guidance = guidance_scale > 1.0
# 1. Encode input prompt
prompt_embeds = self._prepare_and_encode_inputs(
inputs,
do_classifier_free_guidance,
).to(dtype).to(device)
batch_size = prompt_embeds.shape[0] // 2 if do_classifier_free_guidance else prompt_embeds.shape[0]
unet_added_conditions = {}
time_ids = torch.LongTensor(original_size + crop_info + [height, width]).to(device)
if do_classifier_free_guidance:
unet_added_conditions["time_ids"] = torch.cat([time_ids, time_ids], dim=0)
else:
unet_added_conditions["time_ids"] = time_ids
unet_added_conditions["text_embeds"] = torch.mean(prompt_embeds, dim=1)
# 2. Prepare timesteps
self.scheduler.set_timesteps(num_inference_steps, device=device)
timesteps = self.scheduler.timesteps
# 3. Prepare latent variables
shape = (
batch_size,
self.unet.config.in_channels,
height // self.vae_scale_factor,
width // self.vae_scale_factor,
)
latents = torch.randn(shape, device=device, dtype=dtype)
latents = latents * self.scheduler.init_noise_sigma
# 4. Denoising loop
for t in tqdm(timesteps):
# expand the latents if we are doing classifier free guidance
# 2B x 4 x H x W
latent_model_input = torch.cat([latents] * 2) if do_classifier_free_guidance else latents
latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
noise_pred = self.unet(
latent_model_input,
t,
encoder_hidden_states=prompt_embeds,
added_cond_kwargs=unet_added_conditions,
).sample
# perform guidance
if do_classifier_free_guidance:
noise_pred_cond, noise_pred_uncond = noise_pred.chunk(2)
noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_cond - noise_pred_uncond)
# compute the previous noisy sample x_t -> x_t-1
latents = self.scheduler.step(noise_pred, t, latents).prev_sample
# 5. Post-processing
images = self.decode_latents(latents)
# 6. Run safety checker
images, has_nsfw_concept = self.run_safety_checker(images)
# 7. Convert to PIL
images = self.numpy_to_pil(images)
return EmuVisualGenerationPipelineOutput(
image=images[0],
nsfw_content_detected=None if has_nsfw_concept is None else has_nsfw_concept[0],
)
def _prepare_and_encode_inputs(
self,
inputs: List[str | Image.Image],
do_classifier_free_guidance: bool = False,
placeholder: str = DEFAULT_IMG_PLACEHOLDER,
):
device = self.device(self.multimodal_encoder.model.visual)
dtype = self.dtype(self.multimodal_encoder.model.visual)
has_image, has_text = False, False
text_prompt, image_prompt = "", []
for x in inputs:
if isinstance(x, str):
has_text = True
text_prompt += x
else:
has_image = True
text_prompt += placeholder
image_prompt.append(self.transform(x))
if len(image_prompt) == 0:
image_prompt = None
else:
image_prompt = torch.stack(image_prompt)
image_prompt = image_prompt.type(dtype).to(device)
if has_image and not has_text:
prompt = self.multimodal_encoder.model.encode_image(image=image_prompt)
if do_classifier_free_guidance:
key = "[NULL_IMAGE]"
if key not in self.negative_prompt:
negative_image = torch.zeros_like(image_prompt)
self.negative_prompt[key] = self.multimodal_encoder.model.encode_image(image=negative_image)
prompt = torch.cat([prompt, self.negative_prompt[key]], dim=0)
else:
prompt = self.multimodal_encoder.generate_image(text=[text_prompt], image=image_prompt, tokenizer=self.tokenizer)
if do_classifier_free_guidance:
key = ""
if key not in self.negative_prompt:
self.negative_prompt[key] = self.multimodal_encoder.generate_image(text=[""], tokenizer=self.tokenizer)
prompt = torch.cat([prompt, self.negative_prompt[key]], dim=0)
return prompt
def decode_latents(self, latents: torch.Tensor) -> np.ndarray:
latents = 1 / self.vae.config.scaling_factor * latents
image = self.vae.decode(latents).sample
image = (image / 2 + 0.5).clamp(0, 1)
image = image.cpu().permute(0, 2, 3, 1).float().numpy()
return image
def numpy_to_pil(self, images: np.ndarray) -> List[Image.Image]:
"""
Convert a numpy image or a batch of images to a PIL image.
"""
if images.ndim == 3:
images = images[None, ...]
images = (images * 255).round().astype("uint8")
if images.shape[-1] == 1:
# special case for grayscale (single channel) images
pil_images = [Image.fromarray(image.squeeze(), mode="L") for image in images]
else:
pil_images = [Image.fromarray(image) for image in images]
return pil_images
def run_safety_checker(self, images: np.ndarray):
if self.safety_checker is not None:
device = self.device(self.safety_checker)
dtype = self.dtype(self.safety_checker)
safety_checker_input = self.feature_extractor(self.numpy_to_pil(images), return_tensors="pt").to(device)
images, has_nsfw_concept = self.safety_checker(
images=images, clip_input=safety_checker_input.pixel_values.to(dtype)
)
else:
has_nsfw_concept = None
return images, has_nsfw_concept
|