Upload modeling_internvl_chat.py with huggingface_hub
Browse files- modeling_internvl_chat.py +323 -0
modeling_internvl_chat.py
ADDED
@@ -0,0 +1,323 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# --------------------------------------------------------
|
2 |
+
# InternVL
|
3 |
+
# Copyright (c) 2024 OpenGVLab
|
4 |
+
# Licensed under The MIT License [see LICENSE for details]
|
5 |
+
# --------------------------------------------------------
|
6 |
+
import warnings
|
7 |
+
from typing import Any, List, Optional, Tuple, Union
|
8 |
+
|
9 |
+
import torch.utils.checkpoint
|
10 |
+
from torch import nn
|
11 |
+
from torch.nn import CrossEntropyLoss
|
12 |
+
from transformers import (AutoModel, GenerationConfig, LlamaForCausalLM,
|
13 |
+
LlamaTokenizer)
|
14 |
+
from transformers.modeling_outputs import CausalLMOutputWithPast
|
15 |
+
from transformers.modeling_utils import PreTrainedModel
|
16 |
+
from transformers.utils import ModelOutput, logging
|
17 |
+
|
18 |
+
from .configuration_internvl_chat import InternVLChatConfig
|
19 |
+
from .conversation import get_conv_template
|
20 |
+
from .modeling_intern_vit import InternVisionModel
|
21 |
+
from .modeling_phi3 import Phi3ForCausalLM
|
22 |
+
|
23 |
+
logger = logging.get_logger(__name__)
|
24 |
+
|
25 |
+
|
26 |
+
class InternVLChatModel(PreTrainedModel):
|
27 |
+
config_class = InternVLChatConfig
|
28 |
+
main_input_name = 'pixel_values'
|
29 |
+
_no_split_modules = ['InternVisionModel', 'LlamaDecoderLayer', 'Phi3DecoderLayer']
|
30 |
+
|
31 |
+
def __init__(self, config: InternVLChatConfig, vision_model=None, language_model=None):
|
32 |
+
super().__init__(config)
|
33 |
+
|
34 |
+
image_size = config.force_image_size or config.vision_config.image_size
|
35 |
+
patch_size = config.vision_config.patch_size
|
36 |
+
self.patch_size = patch_size
|
37 |
+
self.select_layer = config.select_layer
|
38 |
+
self.template = config.template
|
39 |
+
self.num_image_token = int((image_size // patch_size) ** 2 * (config.downsample_ratio ** 2))
|
40 |
+
self.downsample_ratio = config.downsample_ratio
|
41 |
+
self.ps_version = config.ps_version
|
42 |
+
|
43 |
+
logger.info(f'num_image_token: {self.num_image_token}')
|
44 |
+
logger.info(f'ps_version: {self.ps_version}')
|
45 |
+
if vision_model is not None:
|
46 |
+
self.vision_model = vision_model
|
47 |
+
else:
|
48 |
+
self.vision_model = InternVisionModel(config.vision_config)
|
49 |
+
if language_model is not None:
|
50 |
+
self.language_model = language_model
|
51 |
+
else:
|
52 |
+
if config.llm_config.architectures[0] == 'LlamaForCausalLM':
|
53 |
+
self.language_model = LlamaForCausalLM(config.llm_config)
|
54 |
+
elif config.llm_config.architectures[0] == 'Phi3ForCausalLM':
|
55 |
+
self.language_model = Phi3ForCausalLM(config.llm_config)
|
56 |
+
else:
|
57 |
+
raise NotImplementedError(f'{config.llm_config.architectures[0]} is not implemented.')
|
58 |
+
|
59 |
+
vit_hidden_size = config.vision_config.hidden_size
|
60 |
+
llm_hidden_size = config.llm_config.hidden_size
|
61 |
+
|
62 |
+
self.mlp1 = nn.Sequential(
|
63 |
+
nn.LayerNorm(vit_hidden_size * int(1 / self.downsample_ratio) ** 2),
|
64 |
+
nn.Linear(vit_hidden_size * int(1 / self.downsample_ratio) ** 2, llm_hidden_size),
|
65 |
+
nn.GELU(),
|
66 |
+
nn.Linear(llm_hidden_size, llm_hidden_size)
|
67 |
+
)
|
68 |
+
|
69 |
+
self.img_context_token_id = None
|
70 |
+
|
71 |
+
def forward(
|
72 |
+
self,
|
73 |
+
pixel_values: torch.FloatTensor,
|
74 |
+
input_ids: torch.LongTensor = None,
|
75 |
+
attention_mask: Optional[torch.Tensor] = None,
|
76 |
+
position_ids: Optional[torch.LongTensor] = None,
|
77 |
+
image_flags: Optional[torch.LongTensor] = None,
|
78 |
+
past_key_values: Optional[List[torch.FloatTensor]] = None,
|
79 |
+
labels: Optional[torch.LongTensor] = None,
|
80 |
+
use_cache: Optional[bool] = None,
|
81 |
+
output_attentions: Optional[bool] = None,
|
82 |
+
output_hidden_states: Optional[bool] = None,
|
83 |
+
return_dict: Optional[bool] = None,
|
84 |
+
) -> Union[Tuple, CausalLMOutputWithPast]:
|
85 |
+
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
86 |
+
|
87 |
+
image_flags = image_flags.squeeze(-1)
|
88 |
+
input_embeds = self.language_model.get_input_embeddings()(input_ids)
|
89 |
+
|
90 |
+
vit_embeds = self.extract_feature(pixel_values)
|
91 |
+
vit_embeds = vit_embeds[image_flags == 1]
|
92 |
+
vit_batch_size = pixel_values.shape[0]
|
93 |
+
|
94 |
+
B, N, C = input_embeds.shape
|
95 |
+
input_embeds = input_embeds.reshape(B * N, C)
|
96 |
+
|
97 |
+
if torch.distributed.get_rank() == 0:
|
98 |
+
print(f'dynamic ViT batch size: {vit_batch_size}, images per sample: {vit_batch_size / B}, dynamic token length: {N}')
|
99 |
+
|
100 |
+
input_ids = input_ids.reshape(B * N)
|
101 |
+
selected = (input_ids == self.img_context_token_id)
|
102 |
+
try:
|
103 |
+
input_embeds[selected] = input_embeds[selected] * 0.0 + vit_embeds.reshape(-1, C)
|
104 |
+
except Exception as e:
|
105 |
+
vit_embeds = vit_embeds.reshape(-1, C)
|
106 |
+
print(f'warning: {e}, input_embeds[selected].shape={input_embeds[selected].shape}, '
|
107 |
+
f'vit_embeds.shape={vit_embeds.shape}')
|
108 |
+
n_token = selected.sum()
|
109 |
+
input_embeds[selected] = input_embeds[selected] * 0.0 + vit_embeds[:n_token]
|
110 |
+
|
111 |
+
input_embeds = input_embeds.reshape(B, N, C)
|
112 |
+
|
113 |
+
outputs = self.language_model(
|
114 |
+
inputs_embeds=input_embeds,
|
115 |
+
attention_mask=attention_mask,
|
116 |
+
position_ids=position_ids,
|
117 |
+
past_key_values=past_key_values,
|
118 |
+
use_cache=use_cache,
|
119 |
+
output_attentions=output_attentions,
|
120 |
+
output_hidden_states=output_hidden_states,
|
121 |
+
return_dict=return_dict,
|
122 |
+
)
|
123 |
+
logits = outputs.logits
|
124 |
+
|
125 |
+
loss = None
|
126 |
+
if labels is not None:
|
127 |
+
# Shift so that tokens < n predict n
|
128 |
+
shift_logits = logits[..., :-1, :].contiguous()
|
129 |
+
shift_labels = labels[..., 1:].contiguous()
|
130 |
+
# Flatten the tokens
|
131 |
+
loss_fct = CrossEntropyLoss()
|
132 |
+
shift_logits = shift_logits.view(-1, self.language_model.config.vocab_size)
|
133 |
+
shift_labels = shift_labels.view(-1)
|
134 |
+
# Enable model parallelism
|
135 |
+
shift_labels = shift_labels.to(shift_logits.device)
|
136 |
+
loss = loss_fct(shift_logits, shift_labels)
|
137 |
+
|
138 |
+
if not return_dict:
|
139 |
+
output = (logits,) + outputs[1:]
|
140 |
+
return (loss,) + output if loss is not None else output
|
141 |
+
|
142 |
+
return CausalLMOutputWithPast(
|
143 |
+
loss=loss,
|
144 |
+
logits=logits,
|
145 |
+
past_key_values=outputs.past_key_values,
|
146 |
+
hidden_states=outputs.hidden_states,
|
147 |
+
attentions=outputs.attentions,
|
148 |
+
)
|
149 |
+
|
150 |
+
def pixel_shuffle(self, x, scale_factor=0.5):
|
151 |
+
n, w, h, c = x.size()
|
152 |
+
# N, W, H, C --> N, W, H * scale, C // scale
|
153 |
+
x = x.view(n, w, int(h * scale_factor), int(c / scale_factor))
|
154 |
+
# N, W, H * scale, C // scale --> N, H * scale, W, C // scale
|
155 |
+
x = x.permute(0, 2, 1, 3).contiguous()
|
156 |
+
# N, H * scale, W, C // scale --> N, H * scale, W * scale, C // (scale ** 2)
|
157 |
+
x = x.view(n, int(h * scale_factor), int(w * scale_factor),
|
158 |
+
int(c / (scale_factor * scale_factor)))
|
159 |
+
if self.ps_version == 'v1':
|
160 |
+
warnings.warn("In ps_version 'v1', the height and width have not been swapped back, "
|
161 |
+
'which results in a transposed image.')
|
162 |
+
else:
|
163 |
+
x = x.permute(0, 2, 1, 3).contiguous()
|
164 |
+
return x
|
165 |
+
|
166 |
+
def extract_feature(self, pixel_values):
|
167 |
+
if self.select_layer == -1:
|
168 |
+
vit_embeds = self.vision_model(
|
169 |
+
pixel_values=pixel_values,
|
170 |
+
output_hidden_states=False,
|
171 |
+
return_dict=True).last_hidden_state
|
172 |
+
else:
|
173 |
+
vit_embeds = self.vision_model(
|
174 |
+
pixel_values=pixel_values,
|
175 |
+
output_hidden_states=True,
|
176 |
+
return_dict=True).hidden_states[self.select_layer]
|
177 |
+
vit_embeds = vit_embeds[:, 1:, :]
|
178 |
+
|
179 |
+
h = w = int(vit_embeds.shape[1] ** 0.5)
|
180 |
+
vit_embeds = vit_embeds.reshape(vit_embeds.shape[0], h, w, -1)
|
181 |
+
vit_embeds = self.pixel_shuffle(vit_embeds, scale_factor=self.downsample_ratio)
|
182 |
+
vit_embeds = vit_embeds.reshape(vit_embeds.shape[0], -1, vit_embeds.shape[-1])
|
183 |
+
vit_embeds = self.mlp1(vit_embeds)
|
184 |
+
return vit_embeds
|
185 |
+
|
186 |
+
def batch_chat(self, tokenizer, pixel_values, num_patches_list, questions, generation_config, history=None,
|
187 |
+
return_history=False, IMG_START_TOKEN='<img>', IMG_END_TOKEN='</img>',
|
188 |
+
IMG_CONTEXT_TOKEN='<IMG_CONTEXT>', verbose=False):
|
189 |
+
if history is not None or return_history:
|
190 |
+
print('Now multi-turn chat is not supported in batch_chat.')
|
191 |
+
raise NotImplementedError
|
192 |
+
img_context_token_id = tokenizer.convert_tokens_to_ids(IMG_CONTEXT_TOKEN)
|
193 |
+
self.img_context_token_id = img_context_token_id
|
194 |
+
|
195 |
+
from .conversation import get_conv_template
|
196 |
+
|
197 |
+
queries = []
|
198 |
+
if verbose:
|
199 |
+
image_bs = pixel_values.shape[0]
|
200 |
+
print(f'dynamic ViT batch size: {image_bs}, num_patches_list: {num_patches_list}')
|
201 |
+
for idx, num_patches in enumerate(num_patches_list):
|
202 |
+
image_token = IMG_START_TOKEN + IMG_CONTEXT_TOKEN * self.num_image_token * num_patches + IMG_END_TOKEN
|
203 |
+
question = image_token + '\n' + questions[idx]
|
204 |
+
template = get_conv_template(self.template)
|
205 |
+
template.append_message(template.roles[0], question)
|
206 |
+
template.append_message(template.roles[1], None)
|
207 |
+
query = template.get_prompt()
|
208 |
+
queries.append(query)
|
209 |
+
tokenizer.padding_side = 'left'
|
210 |
+
model_inputs = tokenizer(queries, return_tensors='pt', padding=True)
|
211 |
+
input_ids = model_inputs['input_ids'].cuda()
|
212 |
+
attention_mask = model_inputs['attention_mask'].cuda()
|
213 |
+
eos_token_id = tokenizer.convert_tokens_to_ids(template.sep)
|
214 |
+
generation_config['eos_token_id'] = eos_token_id
|
215 |
+
|
216 |
+
generation_output = self.generate(
|
217 |
+
pixel_values=pixel_values,
|
218 |
+
input_ids=input_ids,
|
219 |
+
attention_mask=attention_mask,
|
220 |
+
**generation_config
|
221 |
+
)
|
222 |
+
responses = tokenizer.batch_decode(generation_output, skip_special_tokens=True)
|
223 |
+
responses = [response.split(template.sep)[0].strip() for response in responses]
|
224 |
+
return responses
|
225 |
+
|
226 |
+
def chat(self, tokenizer, pixel_values, question, generation_config, history=None, return_history=False,
|
227 |
+
num_patches_list=None, IMG_START_TOKEN='<img>', IMG_END_TOKEN='</img>', IMG_CONTEXT_TOKEN='<IMG_CONTEXT>',
|
228 |
+
verbose=False):
|
229 |
+
|
230 |
+
if history is None and pixel_values is not None and '<image>' not in question:
|
231 |
+
question = '<image>\n' + question
|
232 |
+
|
233 |
+
if num_patches_list is None:
|
234 |
+
num_patches_list = [pixel_values.shape[0]] if pixel_values is not None else []
|
235 |
+
assert pixel_values is None or len(pixel_values) == sum(num_patches_list)
|
236 |
+
|
237 |
+
img_context_token_id = tokenizer.convert_tokens_to_ids(IMG_CONTEXT_TOKEN)
|
238 |
+
self.img_context_token_id = img_context_token_id
|
239 |
+
|
240 |
+
template = get_conv_template(self.template)
|
241 |
+
eos_token_id = tokenizer.convert_tokens_to_ids(template.sep)
|
242 |
+
|
243 |
+
history = [] if history is None else history
|
244 |
+
for (old_question, old_answer) in history:
|
245 |
+
template.append_message(template.roles[0], old_question)
|
246 |
+
template.append_message(template.roles[1], old_answer)
|
247 |
+
template.append_message(template.roles[0], question)
|
248 |
+
template.append_message(template.roles[1], None)
|
249 |
+
query = template.get_prompt()
|
250 |
+
|
251 |
+
if verbose and pixel_values is not None:
|
252 |
+
image_bs = pixel_values.shape[0]
|
253 |
+
print(f'dynamic ViT batch size: {image_bs}')
|
254 |
+
|
255 |
+
for num_patches in num_patches_list:
|
256 |
+
image_tokens = IMG_START_TOKEN + IMG_CONTEXT_TOKEN * self.num_image_token * num_patches + IMG_END_TOKEN
|
257 |
+
query = query.replace('<image>', image_tokens, 1)
|
258 |
+
|
259 |
+
model_inputs = tokenizer(query, return_tensors='pt')
|
260 |
+
input_ids = model_inputs['input_ids'].cuda()
|
261 |
+
attention_mask = model_inputs['attention_mask'].cuda()
|
262 |
+
generation_config['eos_token_id'] = eos_token_id
|
263 |
+
generation_output = self.generate(
|
264 |
+
pixel_values=pixel_values,
|
265 |
+
input_ids=input_ids,
|
266 |
+
attention_mask=attention_mask,
|
267 |
+
**generation_config
|
268 |
+
)
|
269 |
+
response = tokenizer.batch_decode(generation_output, skip_special_tokens=True)[0]
|
270 |
+
response = response.split(template.sep)[0].strip()
|
271 |
+
history.append((question, response))
|
272 |
+
if return_history:
|
273 |
+
return response, history
|
274 |
+
else:
|
275 |
+
query_to_print = query.replace(IMG_CONTEXT_TOKEN, '')
|
276 |
+
query_to_print = query_to_print.replace(f'{IMG_START_TOKEN}{IMG_END_TOKEN}', '<image>')
|
277 |
+
if verbose:
|
278 |
+
print(query_to_print, response)
|
279 |
+
return response
|
280 |
+
|
281 |
+
@torch.no_grad()
|
282 |
+
def generate(
|
283 |
+
self,
|
284 |
+
pixel_values: Optional[torch.FloatTensor] = None,
|
285 |
+
input_ids: Optional[torch.FloatTensor] = None,
|
286 |
+
attention_mask: Optional[torch.LongTensor] = None,
|
287 |
+
visual_features: Optional[torch.FloatTensor] = None,
|
288 |
+
generation_config: Optional[GenerationConfig] = None,
|
289 |
+
output_hidden_states: Optional[bool] = None,
|
290 |
+
return_dict: Optional[bool] = None,
|
291 |
+
**generate_kwargs,
|
292 |
+
) -> torch.LongTensor:
|
293 |
+
|
294 |
+
assert self.img_context_token_id is not None
|
295 |
+
if pixel_values is not None:
|
296 |
+
if visual_features is not None:
|
297 |
+
vit_embeds = visual_features
|
298 |
+
else:
|
299 |
+
vit_embeds = self.extract_feature(pixel_values)
|
300 |
+
input_embeds = self.language_model.get_input_embeddings()(input_ids)
|
301 |
+
B, N, C = input_embeds.shape
|
302 |
+
input_embeds = input_embeds.reshape(B * N, C)
|
303 |
+
|
304 |
+
input_ids = input_ids.reshape(B * N)
|
305 |
+
selected = (input_ids == self.img_context_token_id)
|
306 |
+
assert selected.sum() != 0
|
307 |
+
input_embeds[selected] = vit_embeds.reshape(-1, C).to(input_embeds.device)
|
308 |
+
|
309 |
+
input_embeds = input_embeds.reshape(B, N, C)
|
310 |
+
else:
|
311 |
+
input_embeds = self.language_model.get_input_embeddings()(input_ids)
|
312 |
+
|
313 |
+
outputs = self.language_model.generate(
|
314 |
+
inputs_embeds=input_embeds,
|
315 |
+
attention_mask=attention_mask,
|
316 |
+
generation_config=generation_config,
|
317 |
+
output_hidden_states=output_hidden_states,
|
318 |
+
return_dict=return_dict,
|
319 |
+
use_cache=True,
|
320 |
+
**generate_kwargs,
|
321 |
+
)
|
322 |
+
|
323 |
+
return outputs
|