File size: 10,990 Bytes
56fe6da |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 |
# coding=utf-8
#
# Code mainly copied from:
# https://github.com/huggingface/transformers/blob/main/src/transformers/models/clip/modeling_clip.py
# and adjusted for Jina CLIP
from functools import partial
from typing import Optional, Tuple, Union
import torch
import torch.nn.functional as f
import torch.utils.checkpoint
from torch import nn
from transformers import BatchEncoding, BatchFeature, PreTrainedModel, logging
from transformers.models.clip.modeling_clip import (
CLIPOutput,
CLIPTextModelOutput,
CLIPVisionModelOutput,
clip_loss,
)
from .configuration_clip import JinaCLIPConfig, JinaCLIPTextConfig, JinaCLIPVisionConfig
from .eva_model import EVAVisionTransformer
from .hf_model import HFTextEncoder
logger = logging.get_logger(__name__)
""" Jina CLIP model implementation """
class LayerNorm(nn.LayerNorm):
"""Subclass torch's LayerNorm (with cast back to input dtype)."""
def forward(self, x: torch.Tensor):
origtype = x.dtype
x = f.layer_norm(x, self.normalized_shape, self.weight, self.bias, self.eps)
return x.to(origtype)
def _build_text_tower(config: JinaCLIPTextConfig) -> HFTextEncoder:
return HFTextEncoder(
model_name_or_path=config.hf_model_name_or_path,
output_dim=config.embed_dim,
pooler_type=config.pooler_type,
proj_type=config.proj_type,
proj_bias=config.proj_bias,
pretrained=False,
output_tokens=False,
trust_remote_code=True,
revision=None,
model_config_kwargs=config.hf_model_config_kwargs,
)
def _build_vision_tower(config: JinaCLIPVisionConfig) -> EVAVisionTransformer:
norm_layer = partial(LayerNorm, eps=1e-6)
if config.fused_layer_norm:
try:
from apex.normalization import FusedLayerNorm
norm_layer = partial(FusedLayerNorm, eps=1e-6)
except (ModuleNotFoundError, ImportError):
logger.warning('Please install apex to use fused layer norm, ignoring')
return EVAVisionTransformer(
img_size=config.image_size,
patch_size=config.patch_size,
num_classes=config.embed_dim,
use_mean_pooling=False,
init_values=config.ls_init_value,
patch_dropout=config.patch_dropout,
embed_dim=config.width,
depth=config.layers,
num_heads=config.width // config.head_width,
mlp_ratio=config.mlp_ratio,
qkv_bias=config.qkv_bias,
drop_path_rate=config.drop_path_rate,
norm_layer=norm_layer,
xattn=config.x_attention,
rope=config.rope_embeddings,
postnorm=config.post_norm,
pt_hw_seq_len=config.pt_hw_seq_len,
intp_freq=config.intp_freq,
naiveswiglu=config.naive_swiglu,
subln=config.subln,
proj_type=config.proj_type,
)
class JinaCLIPPreTrainedModel(PreTrainedModel):
"""
An abstract class to handle weights initialization and a simple interface for
downloading and loading pretrained models.
"""
config_class = JinaCLIPConfig
base_model_prefix = 'clip'
supports_gradient_checkpointing = True
def _init_weights(self, module):
"""Initialize the weights"""
if isinstance(module, JinaCLIPModel):
if isinstance(module.text_projection, nn.Linear):
nn.init.normal_(
module.text_projection.weight,
std=module.text_embed_dim**-0.5 * self.config.initializer_factor,
)
if isinstance(module.text_projection, nn.Linear):
nn.init.normal_(
module.visual_projection.weight,
std=module.vision_embed_dim**-0.5 * self.config.initializer_factor,
)
if isinstance(module, nn.LayerNorm):
module.bias.data.zero_()
module.weight.data.fill_(1.0)
if isinstance(module, nn.Linear) and module.bias is not None:
module.bias.data.zero_()
class JinaCLIPTextModel(JinaCLIPPreTrainedModel):
config_class = JinaCLIPTextConfig
def __init__(self, config: JinaCLIPTextConfig):
super().__init__(config)
self.text_model = _build_text_tower(config)
self.post_init()
def forward(
self,
input_ids: Union[None, torch.Tensor, BatchEncoding] = None,
return_dict: Optional[bool] = None,
*_,
**__,
) -> Union[Tuple[Optional[torch.FloatTensor], ...], CLIPTextModelOutput]:
return_dict = (
return_dict if return_dict is not None else self.config.use_return_dict
)
x = input_ids.input_ids if isinstance(input_ids, BatchEncoding) else input_ids
feats = self.text_model(x=x)
out = CLIPTextModelOutput(text_embeds=feats)
return out if return_dict else out.to_tuple()
class JinaCLIPVisionModel(JinaCLIPPreTrainedModel):
config_class = JinaCLIPVisionConfig
main_input_name = 'pixel_values'
def __init__(self, config: JinaCLIPVisionConfig):
super().__init__(config)
self.vision_model = _build_vision_tower(config)
self.post_init()
def forward(
self,
pixel_values: Union[None, torch.FloatTensor, BatchFeature] = None,
return_dict: Optional[bool] = None,
*_,
**__,
) -> Union[Tuple[Optional[torch.FloatTensor], ...], CLIPVisionModelOutput]:
return_dict = (
return_dict if return_dict is not None else self.config.use_return_dict
)
x = (
pixel_values.pixel_values
if isinstance(pixel_values, BatchFeature)
else pixel_values
)
feats = self.vision_model(x=x)
out = CLIPVisionModelOutput(image_embeds=feats)
return out if return_dict else out.to_tuple()
class JinaCLIPModel(JinaCLIPPreTrainedModel):
config_class = JinaCLIPConfig
def __init__(self, config: JinaCLIPConfig):
super().__init__(config)
if not isinstance(config.text_config, JinaCLIPTextConfig):
raise ValueError(
'Attribute config.text_config is expected to be of type '
f'JinaCLIPTextConfig but is of type {type(config.text_config)}.'
)
if not isinstance(config.vision_config, JinaCLIPVisionConfig):
raise ValueError(
'Attribute config.vision_config is expected to be of type '
f'JinaCLIPVisionConfig but is of type {type(config.vision_config)}.'
)
text_config = config.text_config
vision_config = config.vision_config
self.add_projections = config.add_projections
self.projection_dim = config.projection_dim
self.text_embed_dim = text_config.embed_dim
self.vision_embed_dim = vision_config.embed_dim
self.text_model = _build_text_tower(text_config)
self.vision_model = _build_vision_tower(vision_config)
self.logit_scale = nn.Parameter(
torch.tensor(self.config.logit_scale_init_value)
)
if self.add_projections:
self.visual_projection = nn.Linear(
self.vision_embed_dim, self.projection_dim, bias=False
)
self.text_projection = nn.Linear(
self.text_embed_dim, self.projection_dim, bias=False
)
else:
self.visual_projection = nn.Identity()
self.text_projection = nn.Identity()
self.post_init()
def get_text_features(
self,
input_ids: Union[None, torch.Tensor, BatchEncoding] = None,
*_,
**__,
) -> torch.FloatTensor:
x = input_ids.input_ids if isinstance(input_ids, BatchEncoding) else input_ids
return self.text_projection(self.text_model(x=x))
def get_image_features(
self,
pixel_values: Union[None, torch.FloatTensor, BatchFeature] = None,
*_,
**__,
) -> torch.FloatTensor:
x = (
pixel_values.pixel_values
if isinstance(pixel_values, BatchFeature)
else pixel_values
)
return self.visual_projection(self.vision_model(x=x))
def encode_text(
self,
input_ids: Union[None, torch.Tensor, BatchEncoding] = None,
return_dict: Optional[bool] = None,
*_,
**__,
) -> Union[Tuple[Optional[torch.FloatTensor], ...], CLIPTextModelOutput]:
return_dict = (
return_dict if return_dict is not None else self.config.use_return_dict
)
feats = self.get_text_features(input_ids=input_ids)
out = CLIPTextModelOutput(text_embeds=feats)
return out if return_dict else out.to_tuple()
def encode_image(
self,
pixel_values: Union[None, torch.FloatTensor, BatchFeature] = None,
return_dict: Optional[bool] = None,
*_,
**__,
) -> Union[Tuple[Optional[torch.FloatTensor], ...], CLIPVisionModelOutput]:
return_dict = (
return_dict if return_dict is not None else self.config.use_return_dict
)
feats = self.get_image_features(pixel_values=pixel_values)
out = CLIPVisionModelOutput(image_embeds=feats)
return out if return_dict else out.to_tuple()
def forward(
self,
input_ids: Union[None, torch.Tensor, BatchEncoding] = None,
pixel_values: Union[None, torch.FloatTensor, BatchFeature] = None,
return_dict: Optional[bool] = None,
return_loss: Optional[bool] = None,
*_,
**__,
) -> Union[Tuple[Optional[torch.FloatTensor], ...], CLIPOutput]:
return_dict = (
return_dict if return_dict is not None else self.config.use_return_dict
)
image_embeds = self.get_image_features(pixel_values=pixel_values)
text_embeds = self.get_text_features(input_ids=input_ids)
# normalized features
image_embeds = image_embeds / image_embeds.norm(p=2, dim=-1, keepdim=True)
text_embeds = text_embeds / text_embeds.norm(p=2, dim=-1, keepdim=True)
# cosine similarity as logits
logit_scale = self.logit_scale.exp()
logits_per_text = torch.matmul(text_embeds, image_embeds.t()) * logit_scale
logits_per_image = logits_per_text.t()
loss = None
if return_loss:
loss = clip_loss(logits_per_text)
if not return_dict:
output = (
logits_per_image,
logits_per_text,
text_embeds,
image_embeds,
None,
None,
)
return ((loss,) + output) if loss is not None else output
return CLIPOutput(
loss=loss,
logits_per_image=logits_per_image,
logits_per_text=logits_per_text,
text_embeds=text_embeds,
image_embeds=image_embeds,
text_model_output=None,
vision_model_output=None,
)
|