File size: 5,943 Bytes
dffe47c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
# SOURCE: https://github.com/Sulam-Group/IBYDMT/blob/main/ibydmt/multimodal.py

from abc import abstractmethod
from typing import Mapping, Optional

import clip
import open_clip
from transformers import (
    AlignModel,
    AlignProcessor,
    BlipForImageTextRetrieval,
    BlipProcessor,
    FlavaModel,
    FlavaProcessor,
)

from app_lib.config import Config
from app_lib.config import Constants as c


class VisionLanguageModel:
    def __init__(self, backbone: Optional[str] = None, device=c.DEVICE):
        pass

    @abstractmethod
    def encode_text(self, text):
        pass

    @abstractmethod
    def encode_image(self, image):
        pass


models: Mapping[str, VisionLanguageModel] = {}


def register_model(name):
    def register(cls: VisionLanguageModel):
        if name in models:
            raise ValueError(f"Model {name} is already registered")
        models[name] = cls

    return register


def get_model_name_and_backbone(config: Config):
    backbone = config.data.backbone.split(":")
    if len(backbone) == 1:
        backbone.append(None)
    return backbone


def get_model(config: Config, device=c.DEVICE) -> VisionLanguageModel:
    model_name, backbone = get_model_name_and_backbone(config)
    return models[model_name](backbone, device=device)


def get_text_encoder(config: Config, device=c.DEVICE):
    model = get_model(config, device=device)
    return model.encode_text


def get_image_encoder(config: Config, device=c.DEVICE):
    model = get_model(config, device=device)
    return model.encode_image


@register_model(name="clip")
class CLIPModel(VisionLanguageModel):
    def __init__(self, backbone: str, device=c.DEVICE):
        self.model, self.preprocess = clip.load(backbone, device=device)
        self.tokenize = clip.tokenize

        self.device = device

    def encode_text(self, text):
        text = self.tokenize(text).to(self.device)
        return self.model.encode_text(text)

    def encode_image(self, image):
        image = self.preprocess(image).unsqueeze(0).to(self.device)
        return self.model.encode_image(image)


@register_model(name="open_clip")
class OpenClipModel(VisionLanguageModel):
    OPENCLIP_WEIGHTS = {
        "ViT-B-32": "laion2b_s34b_b79k",
        "ViT-L-14": "laion2b_s32b_b82k",
    }

    def __init__(self, backbone: str, device=c.DEVICE):
        self.model, _, self.preprocess = open_clip.create_model_and_transforms(
            backbone, pretrained=self.OPENCLIP_WEIGHTS[backbone], device=device
        )
        self.tokenize = open_clip.get_tokenizer(backbone)

        self.device = device

    def encode_text(self, text):
        text = self.tokenize(text).to(self.device)
        return self.model.encode_text(text)

    def encode_image(self, image):
        image = self.preprocess(image).unsqueeze(0).to(self.device)
        return self.model.encode_image(image)


@register_model(name="flava")
class FLAVAModel(VisionLanguageModel):
    HF_MODEL = "facebook/flava-full"

    def __init__(self, backbone: Optional[str] = None, device=c.DEVICE):
        if backbone is None:
            backbone = self.HF_MODEL

        self.model = FlavaModel.from_pretrained(backbone).to(device)
        self.processor = FlavaProcessor.from_pretrained(backbone)

        self.device = device

    def encode_text(self, text):
        text_inputs = self.processor(
            text=text, return_tensors="pt", padding="max_length", max_length=77
        )
        text_inputs = {k: v.to(self.device) for k, v in text_inputs.items()}
        return self.model.get_text_features(**text_inputs)[:, 0, :]

    def encode_image(self, image):
        image_inputs = self.processor(images=image, return_tensors="pt")
        image_inputs = {k: v.to(self.device) for k, v in image_inputs.items()}
        return self.model.get_image_features(**image_inputs)[:, 0, :]


@register_model(name="align")
class ALIGNModel(VisionLanguageModel):
    HF_MODEL = "kakaobrain/align-base"

    def __init__(self, backbone: Optional[str] = None, device=c.DEVICE):
        if backbone is None:
            backbone = self.HF_MODEL

        self.model = AlignModel.from_pretrained(backbone).to(device)
        self.processor = AlignProcessor.from_pretrained(backbone)

        self.device = device

    def encode_text(self, text):
        text_inputs = self.processor(
            text=text, return_tensors="pt", padding="max_length", max_length=77
        )
        text_inputs = {k: v.to(self.device) for k, v in text_inputs.items()}
        return self.model.get_text_features(**text_inputs)

    def encode_image(self, image):
        image_inputs = self.processor(images=image, return_tensors="pt")
        image_inputs = {k: v.to(self.device) for k, v in image_inputs.items()}
        return self.model.get_image_features(**image_inputs)


@register_model(name="blip")
class BLIPModel(VisionLanguageModel):
    HF_MODEL = "Salesforce/blip-itm-base-coco"

    def __init__(self, backbone: Optional[str] = None, device=c.DEVICE):
        if backbone is None:
            backbone = self.HF_MODEL

        self.model = BlipForImageTextRetrieval.from_pretrained(backbone).to(device)
        self.processor = BlipProcessor.from_pretrained(backbone)

        self.device = device

    def encode_text(self, text):
        text_inputs = self.processor(
            text=text, return_tensors="pt", padding="max_length", max_length=77
        )
        text_inputs = {k: v.to(self.device) for k, v in text_inputs.items()}
        question_embeds = self.model.text_encoder(**text_inputs)[0]
        return self.model.text_proj(question_embeds[:, 0, :])

    def encode_image(self, image):
        image_inputs = self.processor(images=image, return_tensors="pt")
        image_inputs = {k: v.to(self.device) for k, v in image_inputs.items()}
        image_embeds = self.model.vision_model(**image_inputs)[0]
        return self.model.vision_proj(image_embeds[:, 0, :])