File size: 19,758 Bytes
4638546
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
83aa27c
 
 
 
 
 
 
684dace
4638546
 
f7464a2
4638546
 
 
 
 
 
 
7803ab7
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
a118490
7803ab7
 
 
 
 
 
 
 
 
 
 
 
4638546
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5e623ce
 
 
4638546
 
5e623ce
4638546
 
 
3d347a9
4638546
 
 
7803ab7
 
 
 
 
45d5056
 
7803ab7
 
 
 
 
 
241c337
7803ab7
 
 
 
22e9be9
7803ab7
 
22e9be9
7803ab7
 
 
4638546
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
#    Copyright 2023 Haotian Liu & Qinghao Ye (Modified from LLaVA)
#
#    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.

from abc import ABC, abstractmethod
from typing import List, Optional, Tuple, Union

import torch
import torch.nn as nn
from torch.nn import CrossEntropyLoss

import copy
import os
import sys

dir_path = os.path.dirname(os.path.realpath(__file__))
sys.path.insert(0, dir_path)

from transformers import AutoConfig, AutoModelForCausalLM, AutoTokenizer, CLIPImageProcessor, LlamaConfig, LlamaModel, LlamaForCausalLM
from transformers.modeling_outputs import CausalLMOutputWithPast

from .configuration_mplug_owl2 import MPLUGOwl2Config, MplugOwlVisionConfig, MplugOwlVisualAbstractorConfig
from .visual_encoder import MplugOwlVisionModel, MplugOwlVisualAbstractorModel
from .modeling_llama2 import replace_llama_modality_adaptive
IGNORE_INDEX = -100
IMAGE_TOKEN_INDEX = -200
DEFAULT_IMAGE_TOKEN = "<|image|>"
from icecream import ic

def tokenizer_image_token(prompt, tokenizer, image_token_index=IMAGE_TOKEN_INDEX, return_tensors=None):
    prompt_chunks = [tokenizer(chunk).input_ids if len(chunk) > 0 else [] for chunk in prompt.split(DEFAULT_IMAGE_TOKEN)]

    def insert_separator(X, sep):
        return [ele for sublist in zip(X, [sep]*len(X)) for ele in sublist][:-1]

    input_ids = []
    offset = 0
    if len(prompt_chunks) > 0 and len(prompt_chunks[0]) > 0 and prompt_chunks[0][0] == tokenizer.bos_token_id:
        offset = 1
        input_ids.append(prompt_chunks[0][0])

    for x in insert_separator(prompt_chunks, [image_token_index] * (offset + 1)):
        input_ids.extend(x[offset:])

    if return_tensors is not None:
        if return_tensors == 'pt':
            return torch.tensor(input_ids, dtype=torch.long)
        raise ValueError(f'Unsupported tensor type: {return_tensors}')
    return input_ids

def expand2square(pil_img, background_color):
        from PIL import Image
        width, height = pil_img.size
        if width == height:
            return pil_img
        elif width > height:
            result = Image.new(pil_img.mode, (width, width), background_color)
            result.paste(pil_img, (0, (width - height) // 2))
            return result
        else:
            result = Image.new(pil_img.mode, (height, height), background_color)
            result.paste(pil_img, ((height - width) // 2, 0))
            return result

class MPLUGOwl2MetaModel:
    def __init__(self, config):
        super(MPLUGOwl2MetaModel, self).__init__(config)
        self.vision_model = MplugOwlVisionModel(
            MplugOwlVisionConfig(**config.visual_config["visual_model"])
        )
        self.visual_abstractor = MplugOwlVisualAbstractorModel(
            MplugOwlVisualAbstractorConfig(**config.visual_config["visual_abstractor"]), config.hidden_size
        )
    
    def get_vision_tower(self):
        vision_model = getattr(self, 'vision_model', None)
        if type(vision_model) is list:
            vision_model = vision_model[0]
        return vision_model

    def get_visual_abstractor(self):
        visual_abstractor = getattr(self, 'visual_abstractor', None)
        if type(visual_abstractor) is list:
            visual_abstractor = visual_abstractor[0]
        return visual_abstractor


class MPLUGOwl2MetaForCausalLM(ABC):
    @abstractmethod
    def get_model(self):
        pass

    def encode_images(self, images):
        image_features = self.get_model().vision_model(images).last_hidden_state
        image_features = self.get_model().visual_abstractor(encoder_hidden_states=image_features).last_hidden_state
        return image_features

    def prepare_inputs_labels_for_multimodal(
        self, input_ids, attention_mask, past_key_values, labels, images
    ):
        if images is None or input_ids.shape[1] == 1:
            if past_key_values is not None and images is not None and input_ids.shape[1] == 1:
                attention_mask = torch.ones((attention_mask.shape[0], past_key_values[-1][-1].shape[-2] + 1), dtype=attention_mask.dtype, device=attention_mask.device)
            multiway_indices = torch.zeros_like(input_ids).long().to(self.device)
            return input_ids, multiway_indices, attention_mask, past_key_values, None, labels
        
        if type(images) is list or images.ndim == 5:
            concat_images = torch.cat([image for image in images], dim=0)
            image_features = self.encode_images(concat_images)
            split_sizes = [image.shape[0] for image in images]
            image_features = torch.split(image_features, split_sizes, dim=0)
            image_features = [x.flatten(0, 1) for x in image_features]
        else:
            image_features = self.encode_images(images)

        new_input_embeds = []
        new_modality_indicators = []
        new_labels = [] if labels is not None else None
        cur_image_idx = 0
        for batch_idx, cur_input_ids in enumerate(input_ids):
            if (cur_input_ids == IMAGE_TOKEN_INDEX).sum() == 0:
                # multimodal LLM, but the current sample is not multimodal
                # FIXME: this is a hacky fix, for deepspeed zero3 to work
                half_len = cur_input_ids.shape[0] // 2
                cur_image_features = image_features[cur_image_idx]
                cur_input_embeds_1 = self.get_model().embed_tokens(cur_input_ids[:half_len])
                cur_input_embeds_2 = self.get_model().embed_tokens(cur_input_ids[half_len:])
                cur_input_embeds = torch.cat([cur_input_embeds_1, cur_image_features[0:0], cur_input_embeds_2], dim=0)
                new_input_embeds.append(cur_input_embeds)
                
                cur_modality_indicators = torch.zeros(len(cur_input_embeds)).long().to(self.device)
                new_modality_indicators.append(cur_modality_indicators)
                if labels is not None:
                    new_labels.append(labels[batch_idx])
                cur_image_idx += 1
                continue
            image_token_indices = torch.where(cur_input_ids == IMAGE_TOKEN_INDEX)[0]
            cur_new_input_embeds = []
            cur_modality_indicators = []
            if labels is not None:
                cur_labels = labels[batch_idx]
                cur_new_labels = []
                assert cur_labels.shape == cur_input_ids.shape
            while image_token_indices.numel() > 0:
                cur_image_features = image_features[cur_image_idx]
                image_token_start = image_token_indices[0]
                cur_new_input_embeds.append(self.get_model().embed_tokens(cur_input_ids[:image_token_start]))
                cur_new_input_embeds.append(cur_image_features)
                
                # Add modality indicator
                assert image_token_start == len(cur_input_ids[:image_token_start])
                cur_modality_indicators.append(torch.zeros(len(cur_input_ids[:image_token_start])).long())
                cur_modality_indicators.append(torch.ones(len(cur_image_features)).long())
                
                if labels is not None:
                    cur_new_labels.append(cur_labels[:image_token_start])
                    cur_new_labels.append(torch.full((cur_image_features.shape[0],), IGNORE_INDEX, device=labels.device, dtype=labels.dtype))
                    cur_labels = cur_labels[image_token_start+1:]
                cur_image_idx += 1
                cur_input_ids = cur_input_ids[image_token_start+1:]
                image_token_indices = torch.where(cur_input_ids == IMAGE_TOKEN_INDEX)[0]
            if cur_input_ids.numel() > 0:
                cur_new_input_embeds.append(self.get_model().embed_tokens(cur_input_ids))
                cur_modality_indicators.append(torch.zeros(len(cur_input_ids)).long())
                if labels is not None:
                    cur_new_labels.append(cur_labels)
            cur_new_input_embeds = [x.to(device=self.device) for x in cur_new_input_embeds]
            cur_new_input_embeds = torch.cat(cur_new_input_embeds, dim=0)
            new_input_embeds.append(cur_new_input_embeds)
            
            # Modality
            cur_modality_indicators = [x.to(device=self.device) for x in cur_modality_indicators]
            cur_modality_indicators = torch.cat(cur_modality_indicators, dim=0)
            new_modality_indicators.append(cur_modality_indicators)
            
            
            if labels is not None:
                cur_new_labels = torch.cat(cur_new_labels, dim=0)
                new_labels.append(cur_new_labels)

        if any(x.shape != new_input_embeds[0].shape for x in new_input_embeds):
            max_len = max(x.shape[0] for x in new_input_embeds)
            
            # Embedding
            new_input_embeds_align = []
            for cur_new_embed in new_input_embeds:
                cur_new_embed = torch.cat((cur_new_embed, torch.zeros((max_len - cur_new_embed.shape[0], cur_new_embed.shape[1]), dtype=cur_new_embed.dtype, device=cur_new_embed.device)), dim=0)
                new_input_embeds_align.append(cur_new_embed)
            new_input_embeds = torch.stack(new_input_embeds_align, dim=0)
            
            # Modality
            new_modality_indicators_align = []
            for cur_modality_indicator in new_modality_indicators:
                cur_new_embed = torch.cat((cur_modality_indicator, torch.zeros(max_len - cur_modality_indicator.shape[0], dtype=cur_modality_indicator.dtype, device=cur_modality_indicator.device)), dim=0)
                new_modality_indicators_align.append(cur_new_embed)
            new_modality_indicators = torch.stack(new_modality_indicators_align, dim=0)
            
            # Label
            if labels is not None:
                new_labels_align = []
                _new_labels = new_labels
                for cur_new_label in new_labels:
                    cur_new_label = torch.cat((cur_new_label, torch.full((max_len - cur_new_label.shape[0],), IGNORE_INDEX, dtype=cur_new_label.dtype, device=cur_new_label.device)), dim=0)
                    new_labels_align.append(cur_new_label)
                new_labels = torch.stack(new_labels_align, dim=0)
            
            # Attention Mask
            if attention_mask is not None:
                new_attention_mask = []
                for cur_attention_mask, cur_new_labels, cur_new_labels_align in zip(attention_mask, _new_labels, new_labels):
                    new_attn_mask_pad_left = torch.full((cur_new_labels.shape[0] - labels.shape[1],), True, dtype=attention_mask.dtype, device=attention_mask.device)
                    new_attn_mask_pad_right = torch.full((cur_new_labels_align.shape[0] - cur_new_labels.shape[0],), False, dtype=attention_mask.dtype, device=attention_mask.device)
                    cur_new_attention_mask = torch.cat((new_attn_mask_pad_left, cur_attention_mask, new_attn_mask_pad_right), dim=0)
                    new_attention_mask.append(cur_new_attention_mask)
                attention_mask = torch.stack(new_attention_mask, dim=0)
                assert attention_mask.shape == new_labels.shape
        else:
            new_input_embeds = torch.stack(new_input_embeds, dim=0)
            new_modality_indicators = torch.stack(new_modality_indicators, dim=0)
            if labels is not None:
                new_labels = torch.stack(new_labels, dim=0)

            if attention_mask is not None:
                new_attn_mask_pad_left = torch.full((attention_mask.shape[0], new_input_embeds.shape[1] - input_ids.shape[1]), True, dtype=attention_mask.dtype, device=attention_mask.device)
                attention_mask = torch.cat((new_attn_mask_pad_left, attention_mask), dim=1)
                assert attention_mask.shape == new_input_embeds.shape[:2]
        return None, new_modality_indicators, attention_mask, past_key_values, new_input_embeds, new_labels



class MPLUGOwl2LlamaModel(MPLUGOwl2MetaModel, LlamaModel):
    config_class = MPLUGOwl2Config

    def __init__(self, config: MPLUGOwl2Config):
        super(MPLUGOwl2LlamaModel, self).__init__(config)


class MPLUGOwl2LlamaForCausalLM(LlamaForCausalLM, MPLUGOwl2MetaForCausalLM):
    config_class = MPLUGOwl2Config

    def __init__(self, config):
        super(LlamaForCausalLM, self).__init__(config)
        self.model = MPLUGOwl2LlamaModel(config)
        
        self.tokenizer = AutoTokenizer.from_pretrained("q-future/one-align")
        self.image_processor = CLIPImageProcessor.from_pretrained("q-future/one-align")

        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
        self.preferential_ids_ = [id_[1] for id_ in self.tokenizer(["excellent","good","fair","poor","bad"])["input_ids"]]

        # Initialize weights and apply final processing
        self.post_init()
        

    def get_model(self):
        return self.model
    
    def score(self, images, 
              task_: str = "quality",
              input_: str = "image",
             ):
        if not hasattr(self, "weight_tensor"):
            self.weight_tensor = torch.Tensor([5.,4.,3.,2.,1.]).half().to(self.device)
        prompt = "USER: How would you rate the {} of this {}?\n<|image|>\nASSISTANT: The {} of the {} is".format(task_, input_, input_, task_)
        if input_ == "image":
            images = [expand2square(img, tuple(int(x*255) for x in self.image_processor.image_mean)) for img in images]
            input_ids = tokenizer_image_token(prompt, self.tokenizer, IMAGE_TOKEN_INDEX, return_tensors='pt').unsqueeze(0).to(self.device)
            with torch.inference_mode():
                image_tensor = self.image_processor.preprocess(images, return_tensors="pt")["pixel_values"].half().to(self.device)
                output_logits = self(input_ids.repeat(image_tensor.shape[0], 1),
                                images=image_tensor)["logits"][:,-1, self.preferential_ids_]
                return torch.softmax(output_logits, -1) @ self.weight_tensor
        else:
            video = [[expand2square(frame, tuple(int(x*255) for x in self.image_processor.image_mean)) for frame in vid] for vid in images]
            input_ids = tokenizer_image_token(prompt, self.tokenizer, IMAGE_TOKEN_INDEX, return_tensors='pt').unsqueeze(0).to(self.device)
            with torch.inference_mode():
                video_tensors = [self.image_processor.preprocess(vid, return_tensors="pt")["pixel_values"].half().to(self.model.device) for vid in video]
                output_logits = self(input_ids.repeat(len(video_tensors), 1),
                            images=video_tensors)["logits"][:,-1, self.preferential_ids_]
                return torch.softmax(output_logits, -1) @ self.weight_tensor
        
    def forward(
        self,
        input_ids: torch.LongTensor = None,
        # modality_indicators: torch.LongTensor = None,
        attention_mask: Optional[torch.Tensor] = None,
        past_key_values: Optional[List[torch.FloatTensor]] = None,
        inputs_embeds: Optional[torch.FloatTensor] = None,
        labels: Optional[torch.LongTensor] = None,
        use_cache: Optional[bool] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        images: Optional[torch.FloatTensor] = None,
        return_dict: Optional[bool] = None,
    ) -> Union[Tuple, CausalLMOutputWithPast]:
        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
        output_hidden_states = (
            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
        )
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
        input_ids, modality_indicators, attention_mask, past_key_values, inputs_embeds, labels = \
            self.prepare_inputs_labels_for_multimodal(input_ids, attention_mask, past_key_values, labels, images)

        # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
        outputs = self.model(
            input_ids=input_ids,
            modality_indicators=modality_indicators,
            attention_mask=attention_mask,
            past_key_values=past_key_values,
            inputs_embeds=inputs_embeds,
            use_cache=use_cache,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict
        )

        hidden_states = outputs[0]
        logits = self.lm_head(hidden_states)

        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.config.vocab_size)
            shift_labels = shift_labels.view(-1)
            # Enable model/pipeline 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,
        )

    def prepare_inputs_for_generation(
        self, input_ids, past_key_values=None, attention_mask=None, inputs_embeds=None, **kwargs
    ):
        if past_key_values:
            input_ids = input_ids[:, -1:]

        # if `inputs_embeds` are passed, we only want to use them in the 1st generation step
        if inputs_embeds is not None and past_key_values is None:
            model_inputs = {"inputs_embeds": inputs_embeds}
        else:
            model_inputs = {"input_ids": input_ids}

        model_inputs.update(
            {
                "past_key_values": past_key_values,
                "use_cache": kwargs.get("use_cache"),
                "attention_mask": attention_mask,
                "images": kwargs.get("images", None),
            }
        )
        return model_inputs

AutoConfig.register("mplug_owl2", MPLUGOwl2Config)
AutoModelForCausalLM.register(MPLUGOwl2Config, MPLUGOwl2LlamaForCausalLM)

replace_llama_modality_adaptive()

if __name__ == "__main__":
    config = MPLUGOwl2Config.from_pretrained('q-future/one-align')
    from icecream import ic
    # config = MPLUGOwl2Config()
    model =  AutoModelForCausalLM(config)
    
    images = torch.randn(2, 3, 448, 448)
    input_ids = torch.cat([
        torch.ones(8).long(), torch.tensor([-1]*1).long(), torch.ones(8).long(), torch.tensor([-1]*1).long(), torch.ones(8).long()
    ], dim=0).unsqueeze(0)
    labels = input_ids.clone()
    labels[labels < 0] = -100
    
    # image_feature = model.encode_images(images)
    # ic(image_feature.shape)
    
    output = model(images=images, input_ids=input_ids, labels=labels)
    ic(output.loss)
    ic(output.logits.shape)
    
    model.save_pretrained('/cpfs01/shared/public/test/tmp_owl')