File size: 8,860 Bytes
c46712e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
# AUTOGENERATED! DO NOT EDIT! File to edit: ../notebooks/12_modelling.ipynb.

# %% auto 0
__all__ = ['VTDEConfig', 'VTDEModel', 'SamVisionPreTrainedModel', 'SamVisionModel']

# %% ../notebooks/12_modelling.ipynb 1
from transformers.models.clip.modeling_clip import CLIPOutput, clip_loss
from typing import Optional, Tuple, Union
from transformers import PreTrainedModel, VisionTextDualEncoderModel
import torch
from transformers import VisionTextDualEncoderConfig

class VTDEConfig(VisionTextDualEncoderConfig):
    model_type = "vtde"
    
    def __init__(self, projection_dim=512, logit_scale_init_value=2.6592, 
    text_pooling_mode='mean',
    vision_pooling_mode='max',
    **kwargs):
        """
        pooling_mode in ['mean', 'max', 'cls']
        https://arxiv.org/pdf/2210.09996.pdf
        https://github.com/kahnchana/clippy/blob/3c102c29c32f7c66c6e52e09b795fe9c061bbb03/src/open_clip/hf_model.py#L56
        also
        https://arxiv.org/pdf/2301.07836.pdf
        """
        self.text_pooling_mode = text_pooling_mode
        self.vision_pooling_mode = vision_pooling_mode
        super().__init__(projection_dim, logit_scale_init_value, **kwargs)

VTDEConfig.register_for_auto_class()


class VTDEModel(VisionTextDualEncoderModel):
    config_class = VTDEConfig
    base_model_prefix = "vtde"

    def __init__(
        self,
        config: Optional[VTDEConfig] = None,
        vision_model: Optional[PreTrainedModel] = None,
        text_model: Optional[PreTrainedModel] = None,
    ):
        # You can customize the constructor if needed
        super().__init__(config, vision_model, text_model)
        self.text_pooling_mode = config.text_pooling_mode
        self.vision_pooling_mode = config.vision_pooling_mode

    def get_text_features(
        self,
        input_ids=None,
        attention_mask=None,
        position_ids=None,
        token_type_ids=None,
        output_attentions=None,
        output_hidden_states=None,
        return_dict=None,
    ):
        text_outputs = self.text_model(
            input_ids=input_ids,
            attention_mask=attention_mask,
            token_type_ids=token_type_ids,
            position_ids=position_ids,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
        )
        if self.text_pooling_mode == 'cls':
            pooled_output = text_outputs[1]
        elif self.text_pooling_mode == 'mean':
            pooled_output = torch.mean(text_outputs[0], dim=1)
        elif self.text_pooling_mode == 'max':
            pooled_output = torch.max(text_outputs[0], dim=1)[0]
        elif self.text_pooling_mode == 'norm':
            """we select the patch with the largest norm"""
            last_hidden_states = text_outputs[0]
            patch_norms = torch.norm(last_hidden_states[:, 1:, :], dim=-1)
            max_norm_idx = torch.argmax(patch_norms, dim=1)
            pooled_output = last_hidden_states[:, max_norm_idx, :][:, 0, :]
        else:
            "We want to raise the name of the pooling mode"
            raise NotImplementedError

        text_features = self.text_projection(pooled_output)

        return text_features

    def get_image_features(
        self,
        pixel_values=None,
        output_attentions=None,
        output_hidden_states=None,
        return_dict=None,
    ):
        vision_outputs = self.vision_model(
            pixel_values=pixel_values,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
        )

        if self.vision_pooling_mode == 'cls':
            pooled_output = vision_outputs[1]
        elif self.vision_pooling_mode == 'mean':
            pooled_output = torch.mean(vision_outputs[0], dim=1)
        elif self.vision_pooling_mode == 'max':
            pooled_output = torch.max(vision_outputs[0], dim=1)[0]
        elif self.vision_pooling_mode == 'norm':
            """we select the patch with the largest norm"""
            last_hidden_states = vision_outputs[0]
            patch_norms = torch.norm(last_hidden_states[:, 1:, :], dim=-1)
            max_norm_idx = torch.argmax(patch_norms, dim=1)
            pooled_output = last_hidden_states[:, max_norm_idx, :][:, 0, :]
        else:
            raise NotImplementedError

        image_features = self.visual_projection(pooled_output)

        return image_features

    def forward(
        self,
        input_ids: Optional[torch.LongTensor] = None,
        pixel_values: Optional[torch.FloatTensor] = None,
        attention_mask: Optional[torch.Tensor] = None,
        position_ids: Optional[torch.LongTensor] = None,
        return_loss: Optional[bool] = None,
        token_type_ids: Optional[torch.LongTensor] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
    ) -> Union[Tuple[torch.Tensor], CLIPOutput]:

        return_dict = return_dict if return_dict is not None else self.config.return_dict

        image_embeds = self.get_image_features(
            pixel_values=pixel_values,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
        )

        text_embeds = self.get_text_features(
            input_ids=input_ids,
            attention_mask=attention_mask,
            position_ids=position_ids,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
        )

        # normalized features
        image_embeds = image_embeds / image_embeds.norm(dim=-1, keepdim=True)
        text_embeds = text_embeds / text_embeds.norm(dim=-1, keepdim=True)

        # cosine similarity as logits
        logit_scale = self.logit_scale.exp()
        logits_per_text = torch.matmul(text_embeds, image_embeds.t()) * logit_scale
        logits_per_image = logits_per_text.T

        loss = None
        if return_loss:
            loss = clip_loss(logits_per_text)

        if not return_dict:
            output = (logits_per_image, logits_per_text, text_embeds, image_embeds, text_embeds, image_embeds)
            return ((loss,) + output) if loss is not None else output

        return CLIPOutput(
            loss=loss,
            logits_per_image=logits_per_image,
            logits_per_text=logits_per_text,
            text_embeds=text_embeds,
            image_embeds=image_embeds,
            text_model_output=text_embeds,
            vision_model_output=image_embeds,
        )
    

VTDEModel.register_for_auto_class("AutoModel")
VTDEModel.register_for_auto_class("AutoModelForZeroShotImageClassification")

# %% ../notebooks/12_modelling.ipynb 2
# we want to create a vision-text encoder model for SAM
from transformers import PreTrainedModel
from transformers.models.sam.modeling_sam import SamPositionalEmbedding, SamVisionEncoder, SamVisionEncoderOutput
from transformers.models.sam.configuration_sam import SamVisionConfig
from torch import nn

class SamVisionPreTrainedModel(PreTrainedModel):
    config_class = SamVisionConfig
    base_model_prefix = "sam_vision_encoder"
    main_input_name = "pixel_values"

    def _init_weights(self, module):
        std = self.config.initializer_range
        if isinstance(module, (nn.Linear, nn.Conv2d, nn.ConvTranspose2d)):
            module.weight.data.normal_(mean=0.0, std=std)
            if module.bias is not None:
                module.bias.data.zero_()
        elif isinstance(module, nn.Embedding):
            module.weight.data.normal_(mean=0.0, std=std)
            if module.padding_idx is not None:
                module.weight.data[module.padding_idx].zero_()

class SamVisionModel(SamVisionPreTrainedModel):

    def __init__(self, config):
        super().__init__(config)
        self.shared_image_embedding = SamPositionalEmbedding(config)
        self.vision_encoder = SamVisionEncoder(config)

    def forward(
        self,
        pixel_values=None,
        attention_mask=None,
        output_attentions=None,
        output_hidden_states=None,
        return_dict=None,
    ) -> SamVisionEncoderOutput:
        return_dict = return_dict if return_dict is not None else self.config.return_dict

        image_embeddings = self.shared_image_embedding(pixel_values)
        vision_encoder_outputs = self.vision_encoder(
            image_embeddings,
            attention_mask=attention_mask,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
        )

        return vision_encoder_outputs
    
SamVisionModel.register_for_auto_class("AutoModel")
# SamVisionConfig.register_for_auto_class()