Spaces:
Sleeping
Sleeping
File size: 19,751 Bytes
c3f3b0b |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 |
# Copyright 2023 Haotian Liu
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# ------------------------------------------------------------------------
# Modified from LLaVA (https://github.com/haotian-liu/LLaVA)
# Copyright 2024 Jiachen Li
# ------------------------------------------------------------------------
from abc import ABC, abstractmethod
import torch
import torch.nn as nn
from .multimodal_encoder.builder import build_vision_tower
from .multimodal_projector.builder import build_vision_projector
from cumo.constants import IGNORE_INDEX, IMAGE_TOKEN_INDEX, DEFAULT_IMAGE_PATCH_TOKEN, DEFAULT_IM_START_TOKEN, DEFAULT_IM_END_TOKEN
from cumo.mm_utils import get_anyres_image_grid_shape
class LlavaMetaModel:
def __init__(self, config):
super(LlavaMetaModel, self).__init__(config)
if hasattr(config, "mm_vision_tower"):
self.vision_tower = build_vision_tower(config, delay_load=True)
self.mm_projector = build_vision_projector(config)
if 'unpad' in getattr(config, 'mm_patch_merge_type', ''):
self.image_newline = nn.Parameter(
torch.empty(config.hidden_size, dtype=self.dtype)
)
def get_vision_tower(self):
vision_tower = getattr(self, 'vision_tower', None)
if type(vision_tower) is list:
vision_tower = vision_tower[0]
return vision_tower
def initialize_vision_modules(self, model_args, fsdp=None):
vision_tower = model_args.vision_tower
mm_vision_select_layer = model_args.mm_vision_select_layer
mm_vision_select_feature = model_args.mm_vision_select_feature
pretrain_mm_mlp_adapter = model_args.pretrain_mm_mlp_adapter
vision_tower_dir = model_args.vision_tower_dir
mm_patch_merge_type = model_args.mm_patch_merge_type
self.config.mm_vision_tower = vision_tower
self.config.scales = model_args.scales
vision_tower = build_vision_tower(model_args)
if fsdp is not None and len(fsdp) > 0:
self.vision_tower = [vision_tower]
else:
self.vision_tower = vision_tower
self.config.use_mm_proj = True
self.config.mm_projector_type = getattr(model_args, 'mm_projector_type', 'linear')
self.config.mm_hidden_size = vision_tower.hidden_size
self.config.mm_vision_select_layer = mm_vision_select_layer
self.config.mm_vision_select_feature = mm_vision_select_feature
self.config.mm_patch_merge_type = mm_patch_merge_type
self.config.num_experts = model_args.num_experts
self.config.num_selected = model_args.num_selected
self.config.num_layers = model_args.num_layers
self.config.dropout = model_args.dropout
self.config.mlp_smoe = model_args.mlp_smoe
self.config.clip_smoe = model_args.clip_smoe
self.mm_projector = build_vision_projector(self.config)
if 'unpad' in mm_patch_merge_type:
embed_std = 1 / torch.sqrt(torch.tensor(self.config.hidden_size, dtype=self.dtype))
self.image_newline = nn.Parameter(
torch.randn(self.config.hidden_size, dtype=self.dtype) * embed_std
)
if pretrain_mm_mlp_adapter is not None:
mm_projector_weights = torch.load(pretrain_mm_mlp_adapter, map_location='cpu')
def get_w(weights, keyword):
return {k.split(keyword + '.')[1]: v for k, v in weights.items() if keyword in k}
if self.config.mlp_smoe:
for i in range(model_args.num_experts):
self.mm_projector.experts[i].load_state_dict(get_w(mm_projector_weights, 'mm_projector'))
else:
self.mm_projector.load_state_dict(get_w(mm_projector_weights, 'mm_projector'))
if vision_tower_dir is not None:
vision_tower_weights = torch.load(vision_tower_dir, map_location='cpu')
self.vision_tower.load_state_dict(vision_tower_weights, strict=False)
if self.config.clip_smoe:
current_staet_dict = self.vision_tower.state_dict()
for key, value in current_staet_dict.items():
if 'experts' in key:
key_splits = key.split('.')
new_key = [key_splits[0], key_splits[1], key_splits[2], key_splits[3], 'mlp', key_splits[6], key_splits[7]]
current_staet_dict[key] = vision_tower_weights['.'.join(new_key)]
self.vision_tower.load_state_dict(current_staet_dict, strict=True)
def unpad_image(tensor, original_size):
"""
Unpads a PyTorch tensor of a padded and resized image.
Args:
tensor (torch.Tensor): The image tensor, assumed to be in CxHxW format.
original_size (tuple): The original size of the image (height, width).
Returns:
torch.Tensor: The unpadded image tensor.
"""
original_width, original_height = original_size
current_height, current_width = tensor.shape[1:]
original_aspect_ratio = original_width / original_height
current_aspect_ratio = current_width / current_height
if original_aspect_ratio > current_aspect_ratio:
scale_factor = current_width / original_width
new_height = int(original_height * scale_factor)
padding = (current_height - new_height) // 2
unpadded_tensor = tensor[:, padding:current_height - padding, :]
else:
scale_factor = current_height / original_height
new_width = int(original_width * scale_factor)
padding = (current_width - new_width) // 2
unpadded_tensor = tensor[:, :, padding:current_width - padding]
return unpadded_tensor
class LlavaMetaForCausalLM(ABC):
@abstractmethod
def get_model(self):
pass
def get_vision_tower(self):
return self.get_model().get_vision_tower()
def prepare_inputs_labels_for_multimodal(
self, input_ids, position_ids, attention_mask, past_key_values, labels,
images, image_sizes=None
):
clip_balanced_loss = None
clip_router_z_loss = None
mlp_balanced_loss = None
mlp_router_z_loss = None
vision_tower = self.get_vision_tower()
if vision_tower is None or images is None or input_ids.shape[1] == 1:
return input_ids, position_ids, attention_mask, past_key_values, None, labels, clip_balanced_loss, clip_router_z_loss, mlp_balanced_loss, mlp_router_z_loss
if type(images) is list or images.ndim == 5:
if type(images) is list:
images = [x.unsqueeze(0) if x.ndim == 3 else x for x in images]
concat_images = torch.cat([image for image in images], dim=0)
image_features, clip_balanced_loss, clip_router_z_loss = self.get_model().get_vision_tower()(images)
image_features, mlp_balanced_loss, mlp_router_z_loss = self.get_model().mm_projector(image_features)
split_sizes = [image.shape[0] for image in images]
image_features = torch.split(image_features, split_sizes, dim=0)
mm_patch_merge_type = getattr(self.config, 'mm_patch_merge_type', 'flat')
image_aspect_ratio = getattr(self.config, 'image_aspect_ratio', 'square')
if mm_patch_merge_type == 'flat':
image_features = [x.flatten(0, 1) for x in image_features]
elif mm_patch_merge_type.startswith('spatial'):
new_image_features = []
for image_idx, image_feature in enumerate(image_features):
if image_feature.shape[0] > 1:
base_image_feature = image_feature[0]
image_feature = image_feature[1:]
height = width = self.get_vision_tower().num_patches_per_side
assert height * width == base_image_feature.shape[0]
if image_aspect_ratio == 'anyres':
num_patch_width, num_patch_height = get_anyres_image_grid_shape(image_sizes[image_idx], self.config.image_grid_pinpoints, self.get_vision_tower().config.image_size)
image_feature = image_feature.view(num_patch_height, num_patch_width, height, width, -1)
else:
raise NotImplementedError
if 'unpad' in mm_patch_merge_type:
image_feature = image_feature.permute(4, 0, 2, 1, 3).contiguous()
image_feature = image_feature.flatten(1, 2).flatten(2, 3)
image_feature = unpad_image(image_feature, image_sizes[image_idx])
image_feature = torch.cat((
image_feature,
self.model.image_newline[:, None, None].expand(*image_feature.shape[:-1], 1).to(image_feature.device)
), dim=-1)
image_feature = image_feature.flatten(1, 2).transpose(0, 1)
else:
image_feature = image_feature.permute(0, 2, 1, 3, 4).contiguous()
image_feature = image_feature.flatten(0, 3)
image_feature = torch.cat((base_image_feature, image_feature), dim=0)
else:
image_feature = image_feature[0]
if 'unpad' in mm_patch_merge_type:
image_feature = torch.cat((
image_feature,
self.model.image_newline[None].to(image_feature.device)
), dim=0)
new_image_features.append(image_feature)
image_features = new_image_features
else:
raise ValueError(f"Unexpected mm_patch_merge_type: {self.config.mm_patch_merge_type}")
else:
image_features, clip_balanced_loss, clip_router_z_loss = self.get_model().get_vision_tower()(images)
if self.config.mlp_smoe:
image_features, mlp_balanced_loss, mlp_router_z_loss = self.get_model().mm_projector(image_features)
else:
image_features = self.get_model().mm_projector(image_features)
if getattr(self.config, 'tune_mm_mlp_adapter', False) and getattr(self.config, 'mm_use_im_start_end', False):
raise NotImplementedError
# Let's just add dummy tensors if they do not exist,
# it is a headache to deal with None all the time.
# But it is not ideal, and if you have a better idea,
# please open an issue / submit a PR, thanks.
_labels = labels
_position_ids = position_ids
_attention_mask = attention_mask
if attention_mask is None:
attention_mask = torch.ones_like(input_ids, dtype=torch.bool)
else:
attention_mask = attention_mask.bool()
if position_ids is None:
position_ids = torch.arange(0, input_ids.shape[1], dtype=torch.long, device=input_ids.device)
if labels is None:
labels = torch.full_like(input_ids, IGNORE_INDEX)
# remove the padding using attention_mask -- FIXME
_input_ids = input_ids
input_ids = [cur_input_ids[cur_attention_mask] for cur_input_ids, cur_attention_mask in zip(input_ids, attention_mask)]
labels = [cur_labels[cur_attention_mask] for cur_labels, cur_attention_mask in zip(labels, attention_mask)]
new_input_embeds = []
new_labels = []
cur_image_idx = 0
for batch_idx, cur_input_ids in enumerate(input_ids):
num_images = (cur_input_ids == IMAGE_TOKEN_INDEX).sum()
if num_images == 0:
cur_image_features = image_features[cur_image_idx]
cur_input_embeds_1 = self.get_model().embed_tokens(cur_input_ids)
cur_input_embeds = torch.cat([cur_input_embeds_1, cur_image_features[0:0]], dim=0)
new_input_embeds.append(cur_input_embeds)
new_labels.append(labels[batch_idx])
cur_image_idx += 1
continue
image_token_indices = [-1] + torch.where(cur_input_ids == IMAGE_TOKEN_INDEX)[0].tolist() + [cur_input_ids.shape[0]]
cur_input_ids_noim = []
cur_labels = labels[batch_idx]
cur_labels_noim = []
for i in range(len(image_token_indices) - 1):
cur_input_ids_noim.append(cur_input_ids[image_token_indices[i]+1:image_token_indices[i+1]])
cur_labels_noim.append(cur_labels[image_token_indices[i]+1:image_token_indices[i+1]])
split_sizes = [x.shape[0] for x in cur_labels_noim]
cur_input_embeds = self.get_model().embed_tokens(torch.cat(cur_input_ids_noim))
cur_input_embeds_no_im = torch.split(cur_input_embeds, split_sizes, dim=0)
cur_new_input_embeds = []
cur_new_labels = []
for i in range(num_images + 1):
cur_new_input_embeds.append(cur_input_embeds_no_im[i])
cur_new_labels.append(cur_labels_noim[i])
if i < num_images:
cur_image_features = image_features[cur_image_idx]
cur_image_idx += 1
cur_new_input_embeds.append(cur_image_features)
cur_new_labels.append(torch.full((cur_image_features.shape[0],), IGNORE_INDEX, device=cur_labels.device, dtype=cur_labels.dtype))
cur_new_input_embeds = [x.to(self.device) for x in cur_new_input_embeds]
cur_new_input_embeds = torch.cat(cur_new_input_embeds)
cur_new_labels = torch.cat(cur_new_labels)
new_input_embeds.append(cur_new_input_embeds)
new_labels.append(cur_new_labels)
# Truncate sequences to max length as image embeddings can make the sequence longer
tokenizer_model_max_length = getattr(self.config, 'tokenizer_model_max_length', None)
if tokenizer_model_max_length is not None:
new_input_embeds = [x[:tokenizer_model_max_length] for x in new_input_embeds]
new_labels = [x[:tokenizer_model_max_length] for x in new_labels]
# Combine them
max_len = max(x.shape[0] for x in new_input_embeds)
batch_size = len(new_input_embeds)
new_input_embeds_padded = []
new_labels_padded = torch.full((batch_size, max_len), IGNORE_INDEX, dtype=new_labels[0].dtype, device=new_labels[0].device)
attention_mask = torch.zeros((batch_size, max_len), dtype=attention_mask.dtype, device=attention_mask.device)
position_ids = torch.zeros((batch_size, max_len), dtype=position_ids.dtype, device=position_ids.device)
for i, (cur_new_embed, cur_new_labels) in enumerate(zip(new_input_embeds, new_labels)):
cur_len = cur_new_embed.shape[0]
if getattr(self.config, 'tokenizer_padding_side', 'right') == "left":
new_input_embeds_padded.append(torch.cat((
torch.zeros((max_len - cur_len, cur_new_embed.shape[1]), dtype=cur_new_embed.dtype, device=cur_new_embed.device),
cur_new_embed
), dim=0))
if cur_len > 0:
new_labels_padded[i, -cur_len:] = cur_new_labels
attention_mask[i, -cur_len:] = True
position_ids[i, -cur_len:] = torch.arange(0, cur_len, dtype=position_ids.dtype, device=position_ids.device)
else:
new_input_embeds_padded.append(torch.cat((
cur_new_embed,
torch.zeros((max_len - cur_len, cur_new_embed.shape[1]), dtype=cur_new_embed.dtype, device=cur_new_embed.device)
), dim=0))
if cur_len > 0:
new_labels_padded[i, :cur_len] = cur_new_labels
attention_mask[i, :cur_len] = True
position_ids[i, :cur_len] = torch.arange(0, cur_len, dtype=position_ids.dtype, device=position_ids.device)
new_input_embeds = torch.stack(new_input_embeds_padded, dim=0)
if _labels is None:
new_labels = None
else:
new_labels = new_labels_padded
if _attention_mask is None:
attention_mask = None
else:
attention_mask = attention_mask.to(dtype=_attention_mask.dtype)
if _position_ids is None:
position_ids = None
return None, position_ids, attention_mask, past_key_values, new_input_embeds, new_labels, clip_balanced_loss, clip_router_z_loss, mlp_balanced_loss, mlp_router_z_loss
def initialize_vision_tokenizer(self, model_args, tokenizer):
if model_args.mm_use_im_patch_token:
tokenizer.add_tokens([DEFAULT_IMAGE_PATCH_TOKEN], special_tokens=True)
self.resize_token_embeddings(len(tokenizer))
if model_args.mm_use_im_start_end:
num_new_tokens = tokenizer.add_tokens([DEFAULT_IM_START_TOKEN, DEFAULT_IM_END_TOKEN], special_tokens=True)
self.resize_token_embeddings(len(tokenizer))
if num_new_tokens > 0:
input_embeddings = self.get_input_embeddings().weight.data
output_embeddings = self.get_output_embeddings().weight.data
input_embeddings_avg = input_embeddings[:-num_new_tokens].mean(
dim=0, keepdim=True)
output_embeddings_avg = output_embeddings[:-num_new_tokens].mean(
dim=0, keepdim=True)
input_embeddings[-num_new_tokens:] = input_embeddings_avg
output_embeddings[-num_new_tokens:] = output_embeddings_avg
if model_args.tune_mm_mlp_adapter:
for p in self.get_input_embeddings().parameters():
p.requires_grad = True
for p in self.get_output_embeddings().parameters():
p.requires_grad = False
if model_args.pretrain_mm_mlp_adapter:
mm_projector_weights = torch.load(model_args.pretrain_mm_mlp_adapter, map_location='cpu')
embed_tokens_weight = mm_projector_weights['model.embed_tokens.weight']
assert num_new_tokens == 2
if input_embeddings.shape == embed_tokens_weight.shape:
input_embeddings[-num_new_tokens:] = embed_tokens_weight[-num_new_tokens:]
elif embed_tokens_weight.shape[0] == num_new_tokens:
input_embeddings[-num_new_tokens:] = embed_tokens_weight
else:
raise ValueError(f"Unexpected embed_tokens_weight shape. Pretrained: {embed_tokens_weight.shape}. Current: {input_embeddings.shape}. Numer of new tokens: {num_new_tokens}.")
elif model_args.mm_use_im_patch_token:
if model_args.tune_mm_mlp_adapter:
for p in self.get_input_embeddings().parameters():
p.requires_grad = False
for p in self.get_output_embeddings().parameters():
p.requires_grad = False
|