File size: 11,537 Bytes
8cf6c16
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import logging
import torch
import torch.nn as nn
from contextlib import suppress
from einops import rearrange
from transformers import LlamaForCausalLM, LlamaTokenizer, PreTrainedModel
from torchvision import transforms
from torchvision.transforms.functional import InterpolationMode

from .eva_vit import create_eva_vit_g
from .pooler import Pooler


def get_autocast(precision, cache_enabled=True):
    if precision == "amp":
        return lambda: torch.cuda.amp.autocast(cache_enabled=cache_enabled)
    elif precision == "amp_bfloat16" or precision == "amp_bf16" or precision == 'bf16':
        return lambda: torch.cuda.amp.autocast(dtype=torch.bfloat16, cache_enabled=cache_enabled)
    elif precision == 'fp16':
        return lambda: torch.cuda.amp.autocast(dtype=torch.float16, cache_enabled=cache_enabled)
    elif precision == 'fp32':
        return suppress
    else:
        raise ValueError('not supported precision: {}'.format(precision))
    
class LayerNorm(nn.LayerNorm):
    """Subclass torch's LayerNorm to handle fp16."""
    def forward(self, x: torch.Tensor):
        orig_type = x.dtype
        ret = super().forward(x.type(torch.float32))
        return ret.type(orig_type)

def init_vision_encoder(model_name, 
                        img_size, 
                        drop_path_rate, 
                        use_grad_checkpoint):
    if model_name == "eva_clip_g":
        visual_encoder = create_eva_vit_g(
            img_size, drop_path_rate, use_grad_checkpoint)
    else:
        raise ValueError()
    
    ln_vision = LayerNorm(visual_encoder.num_features)
    return visual_encoder, ln_vision

class ImageProcessor:
    def __init__(self, image_size=364, mean=None, std=None):
        if mean is None:
            self.mean = mean = (0.48145466, 0.4578275, 0.40821073)
        if std is None:
            self.std = std = (0.26862954, 0.26130258, 0.27577711)

        self.normalize = transforms.Normalize(mean, std)
        self.transform = transforms.Compose(
            [
                transforms.Resize(
                    (image_size, image_size), interpolation=InterpolationMode.BICUBIC
                ),
                transforms.ToTensor(),
                self.normalize,
            ]
        )

    def __call__(self, item):
        return self.transform(item)
    
class InfMLLM(PreTrainedModel):
    def __init__(self, config):
        super().__init__(config)
        vit_model = config.vit_model
        img_size = config.image_size
        lm_model = config.lm_model 
        lm_tokenizer = config.lm_tokenizer
        precision = config.precision
        pool_out_size = config.pool_out_size
        self.img_processor = ImageProcessor(image_size=img_size)

        self.visual_encoder, self.ln_vision = init_vision_encoder(
            vit_model, img_size, drop_path_rate=0.0, use_grad_checkpoint=False)

        self.lm_tokenizer = LlamaTokenizer.from_pretrained(lm_tokenizer, use_fast=False, trust_remote_code=True)
        self.lm_tokenizer.pad_token = self.lm_tokenizer.unk_token
        self.lm_model = LlamaForCausalLM.from_pretrained(lm_model, trust_remote_code=True, torch_dtype='auto')
        
        self.pooler = Pooler(dim_in=self.visual_encoder.num_features,
                                dim_out=self.lm_model.config.hidden_size,
                                pool_out_size=pool_out_size)
        self.llama_proj = nn.Identity()
        
        self.precision = precision
        self._apply_lemmatizer = config.apply_lemmatizer if hasattr(config, 'apply_lemmatizer') else False
        self._lemmatizer = None  
    
    def prompt_wrap(self, img_embeds, atts_img, prompts):
        assert len(img_embeds) == len(atts_img) == len(prompts)

        bos = torch.ones([1, 1], dtype=torch.long, device=img_embeds.device) * self.lm_tokenizer.bos_token_id
        bos_embeds = self.lm_model.get_input_embeddings()(bos)

        emb_lists = []
        image_mask = []
        for each_img_embed, each_prompt in zip(img_embeds, prompts):
            assert '<ImageHere>' in each_prompt
            p_before, p_after = each_prompt.split('<ImageHere>')

            p_before_tokens = self.lm_tokenizer(
                p_before, return_tensors="pt", add_special_tokens=False).to(img_embeds.device)
            p_after_tokens = self.lm_tokenizer(
                p_after, return_tensors="pt", add_special_tokens=False).to(img_embeds.device)
            
            p_before_embed = self.lm_model.get_input_embeddings()(p_before_tokens.input_ids.long())                                     # [1, 6, 4096]
            p_after_embed = self.lm_model.get_input_embeddings()(p_after_tokens.input_ids.long())                                       # [1, 17, 4096]
            # add 1 bos
            wrapped_emb = torch.cat([bos_embeds, p_before_embed, each_img_embed[None], p_after_embed], dim=1)                           # [1, 87, 4096]
            emb_lists.append(wrapped_emb)

            image_mask.append( torch.tensor([0] * wrapped_emb.size(1)) )
            image_mask[-1][range(bos_embeds.size(1) + p_before_embed.size(1), 
                                     bos_embeds.size(1) + p_before_embed.size(1) + len(each_img_embed))] = 1
            assert image_mask[-1].sum() == each_img_embed.size(0)

        emb_lens = [emb.shape[1] for emb in emb_lists]
        pad_emb = self.lm_model.get_input_embeddings()(torch.tensor(self.lm_tokenizer.pad_token_id, device=img_embeds.device))          # [4096]
        
        assert not self.training 
        # during inference mode, padding on the left
        wrapped_embs = pad_emb.expand(len(emb_lens), max(emb_lens), -1).clone()                                                         # [12, 87, 4096]
        wrapped_atts = torch.zeros([len(emb_lens), max(emb_lens)], dtype=torch.int, device=img_embeds.device)                           # [12, 87]
        wrapped_image_masks = torch.zeros([len(emb_lens), max(emb_lens)], dtype=torch.int, device=img_embeds.device)                    # [12, 87]
        for i, emb in enumerate(emb_lists):
            wrapped_embs[i, -emb_lens[i]:] = emb
            wrapped_atts[i, -emb_lens[i]:] = 1
            wrapped_image_masks[i, -emb_lens[i]:] = image_mask[i]
        return wrapped_embs, wrapped_atts, wrapped_image_masks

    @torch.no_grad()
    def forward_image_feature(self, image):
        autocast = get_autocast(self.precision, cache_enabled=True)
        with autocast():
            if image.ndim == 4:
                image = image.unsqueeze(1).unsqueeze(1)                                             
            assert image.ndim == 6

            b, t, f = image.shape[:3]
            assert t == 1 and f == 1
            image = rearrange(image, "b t f c h w -> (b t f) c h w")

            image_embeds = self.ln_vision(self.visual_encoder(image))
            
            image_embeds = rearrange(image_embeds, "(b t f) L D -> b t f L D", t=t, f=f)
            query_output= self.pooler(image_embeds)
            query_output = query_output.squeeze(1)
            embeds_img = self.llama_proj(query_output)
            
            return embeds_img
        
    @torch.no_grad()
    def generate(
        self,
        samples,
        use_nucleus_sampling=False,
        num_beams=5,
        max_length=30,
        min_length=1,
        top_p=0.9,
        repetition_penalty=1.0,
        length_penalty=1.0,
        num_captions=1,
        temperature=1,
    ):
        autocast = get_autocast(self.precision, cache_enabled=True)
        with autocast():
            image = samples["image"]  
            embeds_img = self.forward_image_feature(image)
            atts_img = torch.ones(embeds_img.size()[:-1], dtype=torch.long).to(image.device)

            prompts = samples["prompts"]
            assert isinstance(prompts, (tuple, list))

            # Convert prompts to embeds and, repalce "<ImageHere>" with img_embeds
            inputs_embeds, attention_mask, masks_img = self.prompt_wrap(embeds_img, atts_img, prompts)
            
            model_args = dict(
                inputs_embeds=inputs_embeds, 
                attention_mask=attention_mask,
                do_sample=use_nucleus_sampling,
                top_p=top_p,
                temperature=temperature,
                num_beams=num_beams,
                max_length=max_length,
                min_length=min_length,
                eos_token_id=self.lm_tokenizer.eos_token_id,
                repetition_penalty=repetition_penalty,
                length_penalty=length_penalty,
                num_return_sequences=num_captions,
            )
            outputs = self.lm_model.generate(**model_args)

            output_text = self.lm_tokenizer.batch_decode(
                outputs, skip_special_tokens=True
            )
                     
            output_text = [text.strip() for text in output_text]

        return output_text
        
    @torch.no_grad()   
    def predict_answers(
        self,
        samples,
        num_beams=5,
        max_len=10,
        min_len=1,
        length_penalty=0,
    ):
        # VQA tasks
        autocast = get_autocast(self.precision, cache_enabled=True)
        with autocast():
            image = samples["image"]  
            embeds_img = self.forward_image_feature(image)
            atts_img = torch.ones(embeds_img.size()[:-1], dtype=torch.long).to(image.device)
            
            prompts = samples["prompts"]
            assert isinstance(prompts, (tuple, list))

            inputs_embeds, attention_mask, masks_img = self.prompt_wrap(embeds_img, atts_img, prompts)

            model_args = dict(
                inputs_embeds=inputs_embeds,
                attention_mask=attention_mask,
                do_sample=False,
                num_beams=num_beams,
                max_new_tokens=max_len,
                min_length=min_len,
                eos_token_id=self.lm_tokenizer.eos_token_id,
                length_penalty=length_penalty
            )

            outputs = self.lm_model.generate(**model_args)
            output_text = self.lm_tokenizer.batch_decode(
                outputs, skip_special_tokens=True
            )
            output_text = [text.strip() for text in output_text]

        if self._apply_lemmatizer or ("apply_lemmatizer" in samples.keys() and samples["apply_lemmatizer"]):
            output_text = self._lemmatize(output_text)

        return output_text
    
    def _lemmatize(self, answers):
        def apply(answer):
            doc = self.lemmatizer(answer)

            words = []
            for token in doc:
                if token.pos_ in ["NOUN", "VERB"]:
                    words.append(token.lemma_)
                else:
                    words.append(token.text)
            answer = " ".join(words)

            return answer

        return [apply(answer) for answer in answers]
    
    @property
    def lemmatizer(self):
        if self._lemmatizer is None:
            try:
                import spacy

                self._lemmatizer = spacy.load("en_core_web_sm")
            except ImportError:
                logging.error(
                    """
                    Please install spacy and en_core_web_sm model to apply lemmatization.
                    python -m spacy download en_core_web_sm
                    OR
                    import spacy.cli
                    spacy.cli.download("en_core_web_sm")
                    """
                )
                exit(1)

        return self._lemmatizer