File size: 13,332 Bytes
032e687
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
from PIL import Image
import torch
from xtuner.model import InternVL_V1_5
from typing import List, Optional, Tuple, Union
from transformers.modeling_outputs import CausalLMOutputWithPast
from torch.nn import CrossEntropyLoss
from transformers import (AutoModel, GenerationConfig, LlamaForCausalLM,
                          LlamaTokenizer)

from xtuner.utils import PROMPT_TEMPLATE
from xtuner.tools.utils import get_stop_criteria, is_cn_string
from transformers import GenerationConfig

from projects.llava_sam2.models.preprocess.image_resize import DirectResize

from projects.lisa.datasets.sem_seg_dataset import dynamic_preprocess

import torchvision.transforms as T
from torchvision.transforms.functional import InterpolationMode


class InternVL_vlm(InternVL_V1_5):

    def forward(self, data, data_samples=None, mode='loss'):
        pixel_values = data['pixel_values']

        if type(pixel_values) is list or pixel_values.ndim == 5:
            if type(pixel_values) is list:
                pixel_values = [
                    x.unsqueeze(0) if x.ndim == 3 else x for x in pixel_values
                ]
            # b*n, c, h, w
            concat_images = torch.cat(
                [image.to(self.model.vision_model.dtype) for image in pixel_values], dim=0)
        else:
            raise NotImplementedError()

        input_ids = data['input_ids']
        position_ids = data['position_ids']
        attention_mask = data['attention_mask']
        # sum is 0 are text
        image_flags = torch.sum(concat_images, dim=(1, 2, 3)) != 0
        image_flags = image_flags.long()

        labels = data['labels']
        use_cache = False

        outputs = self._llm_forward(
            input_ids=input_ids,
            position_ids=position_ids,
            attention_mask=attention_mask,
            image_flags=image_flags,
            pixel_values=concat_images,
            labels=labels,
            use_cache=use_cache,
            output_hidden_states=True)
        if mode == 'loss':
            return {'llm_loss': outputs.loss,}
        else:
            return outputs

    def _llm_forward(
            self,
            pixel_values: torch.FloatTensor,
            input_ids: torch.LongTensor = None,
            attention_mask: Optional[torch.Tensor] = None,
            position_ids: Optional[torch.LongTensor] = None,
            image_flags: Optional[torch.LongTensor] = None,
            past_key_values: Optional[List[torch.FloatTensor]] = None,
            labels: Optional[torch.LongTensor] = None,
            use_cache: Optional[bool] = None,
            output_attentions: Optional[bool] = None,
            output_hidden_states: Optional[bool] = None,
            return_dict: Optional[bool] = None,
    ) -> Union[Tuple, CausalLMOutputWithPast]:
        return_dict = return_dict if return_dict is not None \
            else self.model.config.use_return_dict

        image_flags = image_flags.squeeze(-1)
        # We only added the clone code here to avoid the error.
        input_embeds = self.model.language_model.get_input_embeddings()(
            input_ids).clone()

        vit_embeds = self.model.extract_feature(pixel_values)
        vit_embeds = vit_embeds.to(input_embeds.dtype)  # FIXME: why vit_embeds is float16?
        vit_embeds = vit_embeds[image_flags == 1]
        vit_batch_size = pixel_values.shape[0]

        B, N, C = input_embeds.shape
        input_embeds = input_embeds.reshape(B * N, C)

        self._count += 1

        input_ids = input_ids.reshape(B * N)
        selected = (input_ids == self.model.img_context_token_id)
        try:
            input_embeds[selected] = vit_embeds.reshape(-1, C)
        except Exception as e:
            vit_embeds = vit_embeds.reshape(-1, C)
            print(f'warning: {e}, input_embeds[selected].shape='
                  f'{input_embeds[selected].shape}, '
                  f'vit_embeds.shape={vit_embeds.shape}')
            n_token = selected.sum()
            input_embeds[selected] = vit_embeds[:n_token]

        input_embeds = input_embeds.reshape(B, N, C)

        outputs = self.model.language_model(
            inputs_embeds=input_embeds,
            attention_mask=attention_mask,
            position_ids=position_ids,
            past_key_values=past_key_values,
            use_cache=use_cache,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
        )
        logits = outputs.logits

        loss = None
        if labels is not None:
            # Shift so that tokens < n predict n
            shift_logits = logits[..., :-1, :].contiguous()
            shift_labels = labels[..., 1:].contiguous()
            # Flatten the tokens
            loss_fct = CrossEntropyLoss()
            shift_logits = shift_logits.view(
                -1, self.model.language_model.config.vocab_size)
            shift_labels = shift_labels.view(-1)
            # Enable model parallelism
            shift_labels = shift_labels.to(shift_logits.device)
            loss = loss_fct(shift_logits, shift_labels)

        if not return_dict:
            output = (logits,) + outputs[1:]
            return (loss,) + output if loss is not None else output

        return CausalLMOutputWithPast(
            loss=loss,
            logits=logits,
            past_key_values=outputs.past_key_values,
            hidden_states=outputs.hidden_states,
            attentions=outputs.attentions,
        )

    @torch.no_grad()
    def generate(
            self,
            pixel_values: Optional[torch.FloatTensor] = None,
            input_ids: Optional[torch.FloatTensor] = None,
            attention_mask: Optional[torch.LongTensor] = None,
            visual_features: Optional[torch.FloatTensor] = None,
            generation_config: Optional[GenerationConfig] = None,
            output_hidden_states: Optional[bool] = None,
            return_dict: Optional[bool] = None,
            **generate_kwargs,
    ) -> torch.LongTensor:
        device = self.model.device
        assert self.model.img_context_token_id is not None
        if pixel_values is not None:
            if visual_features is not None:
                vit_embeds = visual_features
            else:
                if type(pixel_values) is list or pixel_values.ndim == 5:
                    if type(pixel_values) is list:
                        pixel_values = [
                            x.unsqueeze(0) if x.ndim == 3 else x for x in pixel_values
                        ]
                    # b*n, c, h, w
                    pixel_values = torch.cat(
                        [image.to(self.model.vision_model.dtype) for image in pixel_values], dim=0)
                vit_embeds = self.model.extract_feature(pixel_values.to(device))
            image_flags = torch.sum(pixel_values, dim=(1, 2, 3)) != 0
            image_flags = image_flags.long()
            vit_embeds = vit_embeds[image_flags == 1]

            input_embeds = self.model.language_model.get_input_embeddings()(input_ids.to(device))
            B, N, C = input_embeds.shape
            input_embeds = input_embeds.reshape(B * N, C)

            input_ids = input_ids.reshape(B * N)
            selected = (input_ids == self.model.img_context_token_id)
            assert selected.sum() != 0
            input_embeds[selected] = vit_embeds.reshape(-1, C).to(input_embeds.device)

            input_embeds = input_embeds.reshape(B, N, C)
        else:
            input_embeds = self.model.language_model.get_input_embeddings()(input_ids)

        outputs = self.model.language_model.generate(
            inputs_embeds=input_embeds,
            attention_mask=attention_mask.to(device),
            generation_config=generation_config,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
            use_cache=True,
            **generate_kwargs,
        )

        return outputs

    def preparing_for_generation(self, metainfo, **kwargs):
        # set stop criteria and generation configs for model
        self.torch_dtype = torch.bfloat16
        assert 'tokenizer' in metainfo
        tokenizer = metainfo['tokenizer']
        tokenizer_type = tokenizer['type']
        del tokenizer['type']
        self.tokenizer = tokenizer_type(**tokenizer)

        assert hasattr(self, 'tokenizer'), "The Model does not have the tokenizer!!!"
        self.bot_name = 'BOT'
        if 'template' in metainfo.keys():
            template = metainfo['template']
        else:
            template = PROMPT_TEMPLATE['phi3_chat']
        self.template = template
        stop_words = []
        stop_words += template.get('STOP_WORDS', [])
        stop_criteria = get_stop_criteria(
            tokenizer=self.tokenizer, stop_words=stop_words)
        self.stop_criteria = stop_criteria

        default_generation_kwargs = dict(
            max_new_tokens=512,
            do_sample=False,
            eos_token_id=self.tokenizer.eos_token_id,
            pad_token_id=(
                self.tokenizer.pad_token_id
                if self.tokenizer.pad_token_id is not None
                else self.tokenizer.eos_token_id
            ),
        )
        default_generation_kwargs.update(metainfo.get('generation_kwargs', {}))
        self.gen_config = GenerationConfig(**default_generation_kwargs)
        self.init_prediction_config = True

        self.to(self.torch_dtype)

        # for multi image process
        self.min_dynamic_patch = 1
        self.max_dynamic_patch = 12
        self.downsample_ratio = 0.5
        self.image_size = 448
        self.use_thumbnail = True
        patch_size = 14
        self.patch_token = int((self.image_size // patch_size) ** 2 * (self.downsample_ratio ** 2))
        self.IMAGENET_MEAN = (0.485, 0.456, 0.406)
        self.IMAGENET_STD = (0.229, 0.224, 0.225)
        self.IMG_CONTEXT_TOKEN = '<IMG_CONTEXT>'
        self.IMG_START_TOKEN = '<img>'
        self.IMG_END_TOKEN = '</img>'

        self.transformer = T.Compose([
            T.Lambda(lambda img: img.convert('RGB') if img.mode != 'RGB' else img),
            T.Resize((self.image_size, self.image_size), interpolation=InterpolationMode.BICUBIC),
            T.ToTensor(),
            T.Normalize(mean=self.IMAGENET_MEAN, std=self.IMAGENET_STD)
        ])

        # change phi3 prepare for generation fuction
        # self.mllm.model.language_model.prepare_inputs_for_generation = MethodType(prepare_inputs_for_generation, self.mllm.model.language_model)
        return

    def predict_forward(self, question=None, image_path=None, **kwargs):

        assert self.init_prediction_config, "Please set prediction configs using self.preparing_for_generation()"

        input_dict = {}
        # prepare images
        assert image_path is not None, "InternVL2 only support process the image from scratch !!!"

        image = Image.open(image_path).convert('RGB')
        # for pixel segmentation tasks

        images = dynamic_preprocess(image, self.min_dynamic_patch,
                                    self.max_dynamic_patch,
                                    self.image_size, self.use_thumbnail)
        pixel_values = [self.transformer(image) for image in images]
        pixel_values = torch.stack(pixel_values).to(self.torch_dtype)
        input_dict['pixel_values'] = pixel_values

        num_image_tokens = pixel_values.shape[0] * self.patch_token
        image_token_str = f'{self.IMG_START_TOKEN}' \
                          f'{self.IMG_CONTEXT_TOKEN * num_image_tokens}' \
                          f'{self.IMG_END_TOKEN}'


        ret_predictions = []

        if isinstance(question, str):
            text_prompts = [question]
        for text_prompt in text_prompts:
            # add template for text
            text_prompt = text_prompt.replace('<image>', image_token_str)
            input_text = ''
            input_text += self.template['INSTRUCTION'].format(
                input=text_prompt, round=1, bot_name=self.bot_name)

            ids = self.tokenizer.encode(input_text)
            ids = torch.tensor(ids).cuda().unsqueeze(0)

            attention_mask = torch.ones_like(ids, dtype=torch.bool)

            mm_inputs = {
                'pixel_values': input_dict['pixel_values'],
                'input_ids': ids,
                'attention_mask': attention_mask,
                'position_ids': None,
                'past_key_values': None,
                'labels': None
            }

            generate_output = self.generate(
                **mm_inputs,
                generation_config=self.gen_config,
                streamer=None,
                bos_token_id=self.tokenizer.bos_token_id,
                stopping_criteria=self.stop_criteria,
                output_hidden_states=True,
                return_dict_in_generate=True
            )
            predict = self.tokenizer.decode(
                generate_output.sequences[0], skip_special_tokens=True).strip()
            # print(predict)
            ret_predictions.append(predict)

        if len(ret_predictions) == 1:
            ret_predictions = ret_predictions[0]
        print(ret_predictions)
        ret_dict = {'prediction': ret_predictions}
        ret_dict.update(kwargs)
        return ret_dict