File size: 11,930 Bytes
859131c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
from config import LinearMappingConfig
from transformers import (
    GPT2TokenizerFast, GPT2LMHeadModel, AutoModel,
    CLIPVisionModel, AutoProcessor, BatchEncoding,
)
from transformers.models.gpt2.modeling_gpt2 import GPT2DoubleHeadsModelOutput
import torch
import torch.nn as nn
from typing import List, Optional, Union, Tuple, Dict
from torchvision.transforms import CenterCrop, ConvertImageDtype, Normalize, Resize
from torchvision.transforms.functional import InterpolationMode


class Transform(torch.nn.Module):
    def __init__(self, image_size, mean, std):
        super().__init__()
        self.transforms = torch.nn.Sequential(
            Resize([image_size], interpolation=InterpolationMode.BICUBIC, antialias=True),
            CenterCrop(image_size),
            ConvertImageDtype(torch.float32),
            Normalize(mean, std),
        )

    def forward(self, x) -> torch.Tensor:
        """`x` should be an instance of `PIL.Image.Image`"""
        with torch.no_grad():
            x = self.transforms(x)
        return x


class LinearMappingProcessor:
    """
    A combination of ImageProcessor and GPT2TokenizerFast
    """

    def __init__(self, config: LinearMappingConfig):
        self.image_processor = AutoProcessor.from_pretrained(config.image_model)
        self.tokenizer = GPT2TokenizerFast.from_pretrained("gpt2")
        self.add_image_token = config.add_image_token
        if config.add_image_token:
            self.tokenizer.add_special_tokens({"cls_token": "|<image>|"})
        self.tokenizer.pad_token_id = self.tokenizer.eos_token_id
        self.tokenizer.padding_side = "right"
        self.prefix_length = config.prefix_length

    def __call__(self, texts=None, images=None, return_tensors="pt", **kwargs):
        """
        The processor assumes that images and texts are of the same number
        """

        if len(texts) == 0:     # empty strings should be None
            texts = None

        if images is not None:
            image_features = self.image_processor(images=images, return_tensors=return_tensors, **kwargs)
            image_features["attention_mask"] = torch.ones(image_features.pixel_values.size(0),
                                                          self.prefix_length).to(dtype=torch.int64)
            if texts is None and self.add_image_token:
                texts = [self.tokenizer.cls_token for _ in range(image_features.pixel_values.size(0))]
            elif texts is not None and self.add_image_token:
                if isinstance(texts, str):
                    texts = [texts]
                texts = [self.tokenizer.cls_token + text for text in texts]

        elif texts is None:
            texts = self.tokenizer.bos_token

        if texts is not None:
            encoding = self.tokenizer(texts, return_tensors=return_tensors, **kwargs)

        if texts is not None and images is not None:
            encoding["pixel_values"] = image_features.pixel_values

            encoding["attention_mask"] = torch.cat([
                image_features["attention_mask"],
                encoding["attention_mask"]
            ], dim=1).to(dtype=torch.long)  # create attention mask for images
            return encoding

        elif texts is not None:
            return encoding
        else:
            return BatchEncoding(data=dict(**image_features), tensor_type=return_tensors)

    def batch_decode(self, *args, **kwargs):
        """
        This method forwards all its arguments to GPT2TokenizerFast's [`~PreTrainedTokenizer.batch_decode`]. Please
        refer to the docstring of this method for more information.
        """
        return self.tokenizer.batch_decode(*args, **kwargs)

    def decode(self, *args, **kwargs):
        """
        This method forwards all its arguments to GPT2TokenizerFast's [`~PreTrainedTokenizer.decode`]. Please refer to
        the docstring of this method for more information.
        """
        return self.tokenizer.decode(*args, **kwargs)


class ImagePrefix(nn.Module):
    """
    Converts pixel values to prefix image prompts that are later fed to a LLM
    """

    def __init__(self, config: LinearMappingConfig):
        super().__init__()
        self.encoder = AutoModel.from_pretrained(config.image_model)
        if "clip" in config.image_model:
            self.encoder = CLIPVisionModel.from_pretrained(config.image_model)

        if config.freeze_image_model:
            for param in self.encoder.parameters():
                param.requires_grad = False

        self.linear = nn.Linear(config.image_hidden_size, config.text_hidden_size)
        self.ln = nn.LayerNorm(config.text_hidden_size)

    def forward(
            self, pixel_values: torch.Tensor  # B x C x H x W
    ) -> torch.Tensor:
        prefixes = self.encoder(pixel_values).last_hidden_state  # B x N x D
        prefix_prompts = self.linear(prefixes)
        return self.ln(prefix_prompts)


class LinearMapping(nn.Module):

    def __init__(self, config: LinearMappingConfig):
        super().__init__()
        self.image_prefix = ImagePrefix(config)
        self.language_model = GPT2LMHeadModel.from_pretrained(config.text_model)
        self.processor = LinearMappingProcessor(config)
        self.tokenizer = self.processor.tokenizer
        self.image_processor = self.processor.image_processor
        self.add_image_token = config.add_image_token
        if config.add_image_token:
            self.language_model.resize_token_embeddings(len(self.tokenizer))

        if config.freeze_text_model:
            for module in self.language_model.modules():
                if not isinstance(module, nn.LayerNorm) or config.freeze_ln:
                    for param in module.parameters():
                        param.requires_grad = False
            if config.add_image_token:
                # create a gradient mask for the lm_head weight and bias and hook it
                self.language_model.lm_head.weight.requires_grad = True
                self.weight_gradient_mask = nn.Parameter(torch.zeros_like(self.language_model.lm_head.weight),
                                                         requires_grad=False)
                self.weight_gradient_mask[-1, :] = 1.0
                self.language_model.lm_head.weight.register_hook(lambda grad: grad.mul_(self.weight_gradient_mask))

    def prepare_text_inputs(self, input_ids: torch.Tensor) -> torch.Tensor:
        return self.language_model.transformer.wte(input_ids.to(dtype=torch.int64))

    def prepare_inputs(
            self,
            input_ids: Optional[torch.Tensor],
            pixel_values: Optional[torch.Tensor]
    ) -> Dict:
        """
        Prepare captions and pixel values for training.
        It takes the captions' input ids and turn them into input embeddings
        and turns pixel values into prefix prompts.
        Then it concatenates them into one whole prompt batch.
        """
        if input_ids is not None and pixel_values is not None:

            text_embeddings = self.prepare_text_inputs(input_ids)  # B x T x D
            prefix_prompts = self.image_prefix(pixel_values)  # B x V x D
            inputs_embeddings = torch.cat([prefix_prompts, text_embeddings], dim=1)

            prefix_labels = torch.zeros(prefix_prompts.shape[:2], device=prefix_prompts.device) - 100
            labels = torch.cat([prefix_labels, input_ids], dim=1)  # B x (V + T)

            for label in labels:
                for k, token in enumerate(label):
                    if token == self.tokenizer.eos_token_id:
                        label[k + 1:] = -100
                        break
            return {"hidden_states": inputs_embeddings, "labels": labels.to(dtype=torch.int64)}

        elif pixel_values is not None:
            prefix_prompts = self.image_prefix(pixel_values)  # B x V x D
            prefix_labels = torch.zeros(prefix_prompts.shape[:2], device=prefix_prompts.device) - 100
            return {"hidden_states": prefix_prompts, "labels": prefix_labels.to(dtype=torch.int64)}

        elif input_ids is not None:
            text_embeddings = self.prepare_text_inputs(input_ids)
            labels = input_ids.clone()
            for label in labels:
                for k, token in enumerate(label):
                    if token == self.tokenizer.eos_token_id:
                        label[k + 1:] = -100
                        break
            return {"hidden_states": text_embeddings, "labels": labels.to(dtype=torch.int64)}
        else:
            return {"hidden_states": None, "labels": None}

    @torch.no_grad()
    def generate(
            self,
            input_ids: Optional[torch.Tensor] = None,
            pixel_values: Optional[torch.Tensor] = None,
            **kwargs
    ):
        if pixel_values is None:
            return self.language_model.generate(
                input_ids=input_ids,
                **kwargs
            )
        batch_size = pixel_values.size(0)
        past_input_ids = None
        if input_ids is None:
            if self.add_image_token:
                input_ids = torch.tensor([self.tokenizer.cls_token_id for _ in range(batch_size)]).view(batch_size, -1)
            else:
                input_ids = torch.tensor([self.tokenizer.bos_token_id for _ in range(batch_size)]).view(batch_size, -1)
        if input_ids.size(-1) <= 1:
            first_forward_outputs = self.forward(
                pixel_values=pixel_values
            )
        else:
            first_forward_outputs = self.forward(
                pixel_values=pixel_values,
                input_ids=input_ids[:, :-1]
            )
            past_input_ids = input_ids[:, :-1]
            input_ids = input_ids[:, -1].view(batch_size, -1)

        past_key_values = first_forward_outputs.past_key_values

        if kwargs.get("attention_mask", None) is None:
            attention_mask_size = (past_key_values[0][0].size(0), past_key_values[0][0].size(-2))

            attention_mask = torch.ones(attention_mask_size, dtype=torch.int64)
        else:
            attention_mask = kwargs.pop("attention_mask")

        generated_token_ids = self.language_model.generate(
            past_key_values=past_key_values,
            input_ids=input_ids,
            attention_mask=attention_mask,
            **kwargs
        )
        if past_input_ids is not None:
            generated_token_ids = torch.cat([past_input_ids, generated_token_ids], dim=-1)
        return generated_token_ids

    def forward(
            self,
            input_ids: Optional[torch.Tensor] = None,
            pixel_values: Optional[torch.Tensor] = None,
            labels: Optional[torch.Tensor] = None,
            inputs_embeds: Optional[torch.Tensor] = None,
            output_hidden_states: bool = True,
            output_attentions: bool = True,
            attention_mask: Optional[torch.Tensor] = None,
            return_dict: Optional[bool] = True,
            **kwargs
    ) -> Union[GPT2DoubleHeadsModelOutput, Tuple]:
        if (pixel_values is None and input_ids is None) and inputs_embeds is None:
            raise ValueError("You have to specify inputs")
        if inputs_embeds is not None and (pixel_values is not None or input_ids is not None):
            raise ValueError("Either inputs_embeds or (pixel_values and input_ids) should be specified, not both")

        inputs = self.prepare_inputs(input_ids, pixel_values)
        hidden_states = inputs.get('hidden_states', None) if inputs_embeds is None else inputs_embeds
        labels = inputs.get('labels', None) if labels is None else labels

        return self.language_model(
            inputs_embeds=hidden_states,
            labels=labels,
            output_hidden_states=output_hidden_states,
            output_attentions=output_attentions,
            attention_mask=attention_mask,
            return_dict=return_dict,
            **kwargs
        )