File size: 7,510 Bytes
955949b 89d694e 955949b 89d694e 955949b 89d694e 1fe0679 89d694e c63ab68 955949b 89d694e 955949b 89d694e 955949b 1fe0679 89d694e 1fe0679 89d694e d91b3a3 89d694e 955949b 89d694e 955949b 2f07d8b 1fe0679 2f07d8b 955949b 89d694e 955949b 2f07d8b 955949b 89d694e 955949b 89d694e 955949b 2f07d8b 955949b 2f07d8b 955949b |
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 |
from typing import List, Union
import torch
import torch.nn as nn
import torch.nn.functional as F
from open_clip import create_model, get_tokenizer
from open_clip.transform import PreprocessCfg, image_transform_v2
from PIL import Image
from transformers import PretrainedConfig, PreTrainedModel
class MatryoshkaNllbClipConfig(PretrainedConfig):
def __init__(
self,
clip_model_name: str = "",
target_resolution: int = -1,
mrl_resolutions: List[int] = [],
preprocess_cfg: Union[dict, None] = None,
**kwargs,
):
super().__init__(**kwargs)
self.clip_model_name = clip_model_name
self.target_resolution = target_resolution
self.mrl_resolutions = mrl_resolutions
self.preprocess_cfg = preprocess_cfg
class MatryoshkaLayer(nn.Module):
def __init__(self, resolutions: List[int], target_resolution: int = 768):
super().__init__()
self.resolutions = resolutions
self.layers = nn.ModuleDict()
for resolution in resolutions:
self.layers[str(resolution)] = nn.Linear(target_resolution, resolution)
def forward(self, x, resolution: Union[int, None] = None):
if resolution is not None:
if resolution not in self.resolutions:
raise ValueError(f"Resolution {resolution} not in {self.resolutions}")
return self.layers[str(resolution)](x)
outputs = []
for resolution in self.resolutions:
outputs.append(self.layers[str(resolution)](x))
return outputs
class MatryoshkaNllbClip(PreTrainedModel):
config_class = MatryoshkaNllbClipConfig
def __init__(self, config: MatryoshkaNllbClipConfig, device):
super().__init__(config)
if isinstance(device, str):
device = torch.device(device)
self.config = config
self.model = create_model(config.clip_model_name, output_dict=True)
pp_cfg = PreprocessCfg(
size=config.preprocess_cfg["size"],
mean=config.preprocess_cfg["mean"],
std=config.preprocess_cfg["std"],
interpolation=config.preprocess_cfg["interpolation"],
resize_mode=config.preprocess_cfg["resize_mode"],
)
self.transform = image_transform_v2(
pp_cfg,
is_train=False,
)
self._device = device
self.model.to(device)
self.matryoshka_layer = MatryoshkaLayer(
config.mrl_resolutions, config.target_resolution
)
self.matryoshka_layer.to(device)
self.tokenizer = get_tokenizer(config.clip_model_name)
def forward(self, image_inputs, input_ids, resolution: Union[int, None] = None):
image_inputs = image_inputs.to(self._device)
input_ids = input_ids.to(self._device)
outputs = self.model(
image=image_inputs,
text=input_ids,
)
mrl_image_features = None
mrl_text_features = None
if resolution is not None:
mrl_image_features = self.matryoshka_layer.forward(
outputs["image_features"], resolution
)
mrl_text_features = self.matryoshka_layer.forward(
outputs["text_features"], resolution
)
return {
"image_features": outputs["image_features"],
"text_features": outputs["text_features"],
"mrl_image_features": mrl_image_features,
"mrl_text_features": mrl_text_features,
"logit_scale": outputs["logit_scale"],
"logit_bias": outputs["logit_bias"],
}
def encode_image(
self,
image,
normalize=False,
resolution: Union[int, None] = None,
):
with torch.inference_mode():
features = self.model.visual(image)
if resolution is not None:
if resolution not in self.matryoshka_layer.resolutions:
raise ValueError(
f"Resolution {resolution} not in {self.matryoshka_layer.resolutions}"
)
features = self.matryoshka_layer.layers[str(resolution)](features)
return F.normalize(features, dim=-1) if normalize else features
def encode_text(
self,
text,
normalize=False,
resolution: Union[int, None] = None,
):
with torch.inference_mode():
features = self.model.text(text)
if resolution is not None:
if resolution not in self.matryoshka_layer.resolutions:
raise ValueError(
f"Resolution {resolution} not in {self.matryoshka_layer.resolutions}"
)
features = self.matryoshka_layer.layers[str(resolution)](features)
return F.normalize(features, dim=-1) if normalize else features
def image_features(
self,
images: List[Image.Image],
normalize=False,
resolution: Union[int, None] = None,
):
image_inputs = [self.transform(image) for image in images]
image_inputs = torch.stack(image_inputs, dim=0).to(self._device)
with torch.inference_mode():
features = self.model.visual(image_inputs)
if resolution is not None:
if resolution not in self.matryoshka_layer.resolutions:
raise ValueError(
f"Resolution {resolution} not in {self.matryoshka_layer.resolutions}"
)
features = self.matryoshka_layer.layers[str(resolution)](features)
return F.normalize(features, dim=-1) if normalize else features
def text_features(
self,
texts: List[str],
langs: Union[List[str], None] = None,
normalize=False,
resolution: Union[int, None] = None,
):
if langs is None:
langs = ["eng_Latn"] * len(texts)
texts = [f"{lang}{text}" for lang, text in zip(langs, texts)]
input_ids = self.tokenizer.tokenizer.batch_encode_plus(
texts, return_tensors="pt", padding="longest", add_special_tokens=False
)["input_ids"].to(self._device)
with torch.inference_mode():
features = self.model.text(input_ids)
if resolution is not None:
if resolution not in self.matryoshka_layer.resolutions:
raise ValueError(
f"Resolution {resolution} not in {self.matryoshka_layer.resolutions}"
)
features = self.matryoshka_layer.layers[str(resolution)](features)
return F.normalize(features, dim=-1) if normalize else features
def get_logits(
self,
images: List[Image.Image],
texts: List[str],
langs: Union[List[str], None] = None,
resolution: Union[int, None] = None,
):
image_features = self.image_features(
images, normalize=True, resolution=resolution
)
text_features = self.text_features(
texts, langs, normalize=True, resolution=resolution
)
with torch.inference_mode():
image_logits = (
self.model.logit_scale.exp() * image_features @ text_features.T
)
if self.model.logit_bias is not None:
image_logits += self.model.logit_bias
text_logits = image_logits.T
return image_logits, text_logits
|