File size: 9,199 Bytes
68324dd
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import torch
import torch.nn as nn
import numpy as np
from functools import partial
from lib.model_zoo.common.get_model import register

version = '0'
symbol = 'clip'

class AbstractEncoder(nn.Module):
    def __init__(self):
        super().__init__()

    def encode(self, *args, **kwargs):
        raise NotImplementedError

from transformers import CLIPTokenizer, CLIPTextModel

def disabled_train(self, mode=True):
    """Overwrite model.train with this function to make sure train/eval mode
    does not change anymore."""
    return self

@register('clip_text_frozen', version)
class FrozenCLIPTextEmbedder(AbstractEncoder):
    """Uses the CLIP transformer encoder for text (from huggingface)"""
    def __init__(self, version="openai/clip-vit-large-patch14", device="cuda", max_length=77):  # clip-vit-base-patch32
        super().__init__()
        self.tokenizer = CLIPTokenizer.from_pretrained(version)
        self.transformer = CLIPTextModel.from_pretrained(version)
        self.device = device
        self.max_length = max_length   # TODO: typical value?
        self.freeze()

    def freeze(self):
        self.transformer = self.transformer.eval()
        #self.train = disabled_train
        for param in self.parameters():
            param.requires_grad = False

    def forward(self, text):
        batch_encoding = self.tokenizer(text, truncation=True, max_length=self.max_length, return_length=True,
                                        return_overflowing_tokens=False, padding="max_length", return_tensors="pt")
        tokens = batch_encoding["input_ids"].to(self.device)
        outputs = self.transformer(input_ids=tokens)
        z = outputs.last_hidden_state
        return z

    def encode(self, text):
        return self(text)

from transformers import CLIPProcessor, CLIPVisionModel

@register('clip_vision_frozen', version)
class FrozenCLIPVisionEmbedder(AbstractEncoder):
    def __init__(self, version="openai/clip-vit-large-patch14", device="cuda", max_length=77):  # clip-vit-base-patch32
        super().__init__()
        self.processor = CLIPProcessor.from_pretrained(version)
        self.transformer = CLIPVisionModel.from_pretrained(version)
        self.device = device
        self.max_length = max_length   # TODO: typical value?
        self.freeze()

    def freeze(self):
        self.transformer = self.transformer.eval()
        #self.train = disabled_train
        for param in self.parameters():
            param.requires_grad = False

    def forward(self, images):
        inputs = self.processor(images=images, return_tensors="pt")
        pixels = inputs['pixel_values'].to(self.device)
        outputs = self.transformer(pixel_values=pixels)
        z = outputs.last_hidden_state
        return z

    def encode(self, image):
        return self(image)

from transformers import CLIPModel

@register('clip_frozen', version)
class FrozenCLIP(AbstractEncoder):
    def __init__(self, 
                 version="openai/clip-vit-large-patch14", 
                 max_length=77, 
                 encode_type='encode_text',):  # clip-vit-base-patch32
        super().__init__()
        self.tokenizer = CLIPTokenizer.from_pretrained(version)
        self.processor = CLIPProcessor.from_pretrained(version)
        self.model = CLIPModel.from_pretrained(version)
        self.max_length = max_length  # TODO: typical value?
        self.encode_type = encode_type
        self.pinv_text_projection = None
        self.freeze()

    def get_device(self):
        # A trick to get device
        return self.model.text_projection.weight.device

    def freeze(self):
        self.model = self.model.eval()
        #self.train = disabled_train
        for param in self.parameters():
            param.requires_grad = False

    def encode_text_pooled(self, text):
        batch_encoding = self.tokenizer(text, truncation=True, max_length=self.max_length, return_length=True,
                                        return_overflowing_tokens=False, padding="max_length", return_tensors="pt")
        tokens = batch_encoding["input_ids"].to(self.get_device())
        return self.model.get_text_features(input_ids=tokens)

    def encode_vision_pooled(self, images):
        inputs = self.processor(images=images, return_tensors="pt")
        pixels = inputs['pixel_values'].to(self.get_device())
        return self.model.get_image_features(pixel_values=pixels)

    def encode_text_noproj(self, text):
        batch_encoding = self.tokenizer(text, truncation=True, max_length=self.max_length, return_length=True,
                                        return_overflowing_tokens=False, padding="max_length", return_tensors="pt")
        tokens = batch_encoding["input_ids"].to(self.get_device())
        outputs = self.model.text_model(input_ids=tokens)
        return outputs.last_hidden_state
        
    def encode_vision_noproj(self, images):
        inputs = self.processor(images=images, return_tensors="pt")
        pixels = inputs['pixel_values'].to(self.get_device())
        outputs = self.model.vision_model(pixel_values=pixels)
        return outputs.last_hidden_state

    def encode_text_bug(self, text):
        batch_encoding = self.tokenizer(text, truncation=True, max_length=self.max_length, return_length=True,
                                        return_overflowing_tokens=False, padding="max_length", return_tensors="pt")
        tokens = batch_encoding["input_ids"].to(self.get_device())
        outputs = self.model.text_model(input_ids=tokens)
        z = outputs.last_hidden_state
        z_pooled = outputs.pooler_output
        z = z / torch.norm(z_pooled.unsqueeze(1), dim=-1, keepdim=True)
        return self.model.text_projection(z)

    def encode_text(self, text):
        batch_encoding = self.tokenizer(text, truncation=True, max_length=self.max_length, return_length=True,
                                        return_overflowing_tokens=False, padding="max_length", return_tensors="pt")
        tokens = batch_encoding["input_ids"].to(self.get_device())
        outputs = self.model.text_model(input_ids=tokens)
        z = self.model.text_projection(outputs.last_hidden_state)
        z_pooled = self.model.text_projection(outputs.pooler_output)
        z = z / torch.norm(z_pooled.unsqueeze(1), dim=-1, keepdim=True)
        return z

    def encode_vision(self, images):
        z = self.encode_vision_noproj(images)
        z = self.model.vision_model.post_layernorm(z)
        z = self.model.visual_projection(z)
        z_pooled = z[:, 0:1]
        # z_pooled_normed = z_pooled / z_pooled.norm(dim=-1, keepdim=True)
        z = z / torch.norm(z_pooled, dim=-1, keepdim=True)
        return z

    def encode_vision_pinvtext(self, images):
        blank_text_encode_norm_avg = 28.9096
        z = self.encode_vision(images)
        if self.pinv_text_projection is None:
            self.pinv_text_projection = torch.linalg.pinv(self.model.text_projection.weight).T
        z = torch.matmul(z, self.pinv_text_projection)
        # z = z / torch.norm(z[:, 0:1], dim=-1, keepdim=True)
        z = z / torch.norm(z, dim=-1, keepdim=True)
        z = z*blank_text_encode_norm_avg
        # return z[:, 1:2].repeat(1, 77, 1)
        z2 = self.encode_text_noproj('')
        # z2[:, 1:77] = z[:, 0:76]
        return torch.flip(z, dims=(1,))[:, 0:77]

    def encode(self, *args, **kwargs):
        return getattr(self, self.encode_type)(*args, **kwargs)

#############################
# copyed from justin's code #
#############################

@register('clip_vision_frozen_justin', version)
class FrozenCLIPVisionEmbedder_Justin(AbstractEncoder):
    """
        Uses the CLIP image encoder.
        """
    def __init__(
            self,
            model='ViT-L/14',
            jit=False,
            device='cuda' if torch.cuda.is_available() else 'cpu',
            antialias=False,
        ):
        super().__init__()
        from . import clip_justin
        self.model, _ = clip_justin.load(name=model, device=device, jit=jit)
        self.device = device
        self.antialias = antialias

        self.register_buffer('mean', torch.Tensor([0.48145466, 0.4578275, 0.40821073]), persistent=False)
        self.register_buffer('std', torch.Tensor([0.26862954, 0.26130258, 0.27577711]), persistent=False)

        # I didn't call this originally, but seems like it was frozen anyway
        self.freeze()

    def freeze(self):
        self.transformer = self.model.eval()
        for param in self.parameters():
            param.requires_grad = False

    def preprocess(self, x):
        import kornia
        # Expects inputs in the range -1, 1
        x = kornia.geometry.resize(x, (224, 224),
                                   interpolation='bicubic',align_corners=True,
                                   antialias=self.antialias)
        x = (x + 1.) / 2.
        # renormalize according to clip
        x = kornia.enhance.normalize(x, self.mean, self.std)
        return x

    def forward(self, x):
        # x is assumed to be in range [-1,1]
        return self.model.encode_image(self.preprocess(x)).float()

    def encode(self, im):
        return self(im).unsqueeze(1)