File size: 6,050 Bytes
7c34c28
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
from typing import Dict, List, Tuple, Optional

import torch
import torch.nn as nn
from transformers import CLIPVisionModel, CLIPImageProcessor
from PIL import Image

from multi_token.modalities.base_modality import Modality
from multi_token.modalities.projectors import (
    build_patch_mlp_projector,
    build_mlp_vector_projector,
)
from multi_token.data_tools import load_image

PATCH_LAYER = -2
OUTPUT_LAYER = -1
OUTPUT_EMB_SIZE = 1024


class CLIPVisionModule(nn.Module):
    def __init__(self, model_name_or_path: str, feature_layer: int = PATCH_LAYER):
        super().__init__()
        self.feature_layer = feature_layer
        self.model_name_or_path = model_name_or_path
        self.image_processor = None
        self.image_model = None

        self.load_model()

    def load_model(self):
        self.image_processor = CLIPImageProcessor.from_pretrained(
            self.model_name_or_path
        )
        self.image_model = CLIPVisionModel.from_pretrained(self.model_name_or_path)
        self.image_model.requires_grad_(False)

    @torch.no_grad()
    def forward(self, images) -> torch.Tensor:
        if self.feature_layer == PATCH_LAYER:
            image_forward_outs = self.image_model(
                images.to(device=self.device, dtype=self.dtype),
                output_hidden_states=True,
            )
            image_features = image_forward_outs.hidden_states[self.feature_layer]
            image_features = image_features[:, 1:].to(images.dtype)
        else:
            image_forward_outs = self.image_model(
                images.to(device=self.device, dtype=self.dtype),
            )
            image_features = image_forward_outs.pooler_output.to(images.dtype).view(
                -1, 1, OUTPUT_EMB_SIZE
            )
        return image_features

    @property
    def dtype(self):
        return self.image_model.dtype

    @property
    def device(self):
        return self.image_model.device

    @property
    def config(self):
        return self.image_model.config

    @property
    def hidden_size(self):
        return self.config.hidden_size

    @property
    def num_patches(self):
        return (self.config.image_size // self.config.patch_size) ** 2


def _expand2square(pil_img: Image, background_color: Tuple) -> 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 CLIPVisionModality(Modality):
    def __init__(
        self,
        model_name_or_path: str = "openai/clip-vit-large-patch14-336",
        pad_non_square_images: bool = False,
        num_projector_layers: int = 2,
        feature_layer: int = PATCH_LAYER,
        num_tokens_output: Optional[int] = None,
    ):
        if feature_layer not in [PATCH_LAYER, OUTPUT_LAYER]:
            raise ValueError(
                f"feature_layer must be one of {PATCH_LAYER} or {OUTPUT_LAYER}"
            )
        if (feature_layer == PATCH_LAYER) != (num_tokens_output is None):
            raise ValueError(
                "num_tokens_output must be None if feature_layer is PATCH_LAYER"
            )
        self.model_name_or_path = model_name_or_path
        self.module = CLIPVisionModule(
            model_name_or_path=self.model_name_or_path, feature_layer=feature_layer
        )
        self.pad_non_square_images = pad_non_square_images
        self.num_projector_layers = num_projector_layers
        self.num_tokens_output = num_tokens_output

    def build_projector(self, lm_hidden_size: int) -> nn.Module:
        if self.module.feature_layer == PATCH_LAYER:
            return build_patch_mlp_projector(
                self.module.hidden_size,
                lm_hidden_size,
                num_layers=self.num_projector_layers,
            )
        else:
            return build_mlp_vector_projector(
                input_hidden_size=OUTPUT_EMB_SIZE,
                lm_hidden_size=lm_hidden_size,
                num_layers=self.num_projector_layers,
                num_tokens=self.num_tokens_output,
            )

    @property
    def name(self) -> str:
        return "vision_clip"

    @property
    def token(self) -> str:
        return "<image>"

    @property
    def data_key(self) -> str:
        return "images"

    @property
    def token_width(self) -> int:
        if self.module.feature_layer == PATCH_LAYER:
            return self.module.num_patches
        else:
            return self.num_tokens_output

    def to(self, dtype: torch.dtype, device: torch.device) -> "CLIPVisionModality":
        self.module.to(dtype=dtype, device=device)
        return self

    def preprocess_rows(self, rows: List[Dict]) -> List[Optional[torch.Tensor]]:
        row_values = []
        for row in rows:
            images = []
            for image_fn in row[self.data_key]:
                image_obj = load_image(image_fn)
                if self.pad_non_square_images:
                    image_obj = _expand2square(
                        image_obj,
                        tuple(
                            int(x * 255) for x in self.module.image_processor.image_mean
                        ),
                    )
                image = self.module.image_processor.preprocess(
                    image_obj, return_tensors="pt"
                )["pixel_values"][0]
                images.append(image)
            row_values.append(torch.stack(images) if len(images) > 0 else None)
        return row_values

    @torch.no_grad()
    def forward(self, encoded_values: List[torch.Tensor]) -> List[torch.Tensor]:
        image_features = []
        for image_batch in encoded_values:
            image_features.append(self.module.forward(image_batch))
        return image_features