File size: 11,511 Bytes
929a772
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import os
import torch
import torch.nn as nn

from transformers import CLIPVisionModel, CLIPImageProcessor, CLIPVisionConfig


class CLIPVisionTower(nn.Module):
    def __init__(self, vision_tower, args, delay_load=False):
        super().__init__()

        self.is_loaded = False

        self.vision_tower_name = vision_tower
        self.select_layer = args.mm_vision_select_layer
        self.select_feature = getattr(args, 'mm_vision_select_feature', 'patch')

        if not delay_load:
            self.load_model()
        elif getattr(args, 'unfreeze_mm_vision_tower', False):
            self.load_model()
        else:
            self.cfg_only = CLIPVisionConfig.from_pretrained(self.vision_tower_name)

    def load_model(self, device_map=None):
        if self.is_loaded:
            print('{} is already loaded, `load_model` called again, skipping.'.format(self.vision_tower_name))
            return

        self.image_processor = CLIPImageProcessor.from_pretrained(self.vision_tower_name)
        self.vision_tower = CLIPVisionModel.from_pretrained(self.vision_tower_name, device_map=device_map)
        self.vision_tower.requires_grad_(False)

        self.is_loaded = True

    # def feature_select(self, image_forward_outs):
    #     image_features = image_forward_outs.hidden_states[self.select_layer]
    #     if self.select_feature == 'patch':
    #         image_features = image_features[:, 1:]
    #     elif self.select_feature == 'cls_patch':
    #         image_features = image_features
    #     else:
    #         raise ValueError(f'Unexpected select feature: {self.select_feature}')
    #     return image_features
    
    def feature_select_withcls(self, image_forward_outs):
        image_features = image_forward_outs.hidden_states[self.select_layer]
        image_features = image_features
        return image_features
    @torch.no_grad()
    def forward(self, images):
        if type(images) is list:
            image_features = []
            for image in images:
                image_forward_out = self.vision_tower(image.to(device=self.device, dtype=self.dtype).unsqueeze(0), output_hidden_states=True)
                image_feature = self.feature_select_withcls(image_forward_out).to(image.dtype)
                image_features.append(image_feature)
        else:
            image_forward_outs = self.vision_tower(images.to(device=self.device, dtype=self.dtype), output_hidden_states=True)
            image_features = self.feature_select_withcls(image_forward_outs).to(images.dtype)

        return image_features
    def forward_select(self, images, token_num):
        if type(images) is list:
            image_features = []
            for image in images:
                image_forward_out = self.vision_tower(image.to(device=self.device, dtype=self.dtype).unsqueeze(0), output_hidden_states=True)
                image_feature = self.feature_select_withcls(image_forward_out).to(image.dtype)
                image_features.append(image_feature)
        else:
            image_forward_outs = self.vision_tower(images.to(device=self.device, dtype=self.dtype), output_hidden_states=True, output_attentions=True)
            attn_weights  = image_forward_outs.attentions[-2]
            hidden_states = image_forward_outs.hidden_states[-2]
            dominant_num =  token_num

            ## Dominant Visual Tokens
            cls_idx = 0
            cls_attention = attn_weights[:, :, cls_idx, cls_idx+1:]  
            cls_attention_sum = cls_attention.sum(dim=1)  
            topk_indices = cls_attention_sum.topk(dominant_num, dim=1).indices 
            
            topk_indices_sorted = torch.sort(topk_indices, dim=1).values

        return topk_indices_sorted

    def forward_select_scope(self, images, token_num, alpha):
        if type(images) is list:
            image_features = []
            for image in images:
                image_forward_out = self.vision_tower(image.to(device=self.device, dtype=self.dtype).unsqueeze(0), output_hidden_states=True)
                image_feature = self.feature_select_withcls(image_forward_out).to(image.dtype)
                image_features.append(image_feature)
        else:
            image_forward_outs = self.vision_tower(images.to(device=self.device, dtype=self.dtype), output_hidden_states=True, output_attentions=True)
            attn_weights  = image_forward_outs.attentions[-2]
            hidden_states = image_forward_outs.hidden_states[-2]
            dominant_num =  token_num

            ## Dominant Visual Tokens
            # cls_idx = 0
            # cls_attention = attn_weights[:, :, cls_idx, cls_idx+1:]  
            # cls_attention_sum = cls_attention.sum(dim=1)  
            # topk_indices = cls_attention_sum.topk(dominant_num, dim=1).indices 
            
            # topk_indices_sorted = torch.sort(topk_indices, dim=1).values

            cls_idx = 0
            cls_attention = attn_weights[:, :, cls_idx, cls_idx+1:]
            cls_attention_sum = cls_attention.sum(dim=1)

            image_features = hidden_states[:, cls_idx + 1:]
            bs = image_features.shape[0]
            dominant_num = int(dominant_num /bs)
            selected_idx, _ = SCOPE(image_features, dominant_num, cls_attention_sum, alpha)
            # selected_idx += 1

            all_indices = selected_idx
            topk_indices_sorted = torch.sort(all_indices, dim=1).values
            # hidden_states_save = dominant_tokens

        return topk_indices_sorted
    @property
    def dummy_feature(self):
        return torch.zeros(1, self.hidden_size, device=self.device, dtype=self.dtype)

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

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

    @property
    def config(self):
        if self.is_loaded:
            return self.vision_tower.config
        else:
            return self.cfg_only

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

    @property
    def num_patches_per_side(self):
        return self.config.image_size // self.config.patch_size

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



class CLIPVisionTowerS2(CLIPVisionTower):
    def __init__(self, vision_tower, args, delay_load=False):
        super().__init__(vision_tower, args, delay_load)

        self.s2_scales = getattr(args, 's2_scales', '336,672,1008')
        self.s2_scales = list(map(int, self.s2_scales.split(',')))
        self.s2_scales.sort()
        self.s2_split_size = self.s2_scales[0]
        self.s2_image_size = self.s2_scales[-1]

        try:
            from s2wrapper import forward as multiscale_forward
        except ImportError:
            raise ImportError('Package s2wrapper not found! Please install by running: \npip install git+https://github.com/bfshi/scaling_on_scales.git')
        self.multiscale_forward = multiscale_forward

        # change resize/crop size in preprocessing to the largest image size in s2_scale
        if not delay_load or getattr(args, 'unfreeze_mm_vision_tower', False):
            self.image_processor.size['shortest_edge'] = self.s2_image_size
            self.image_processor.crop_size['height'] = self.image_processor.crop_size['width'] = self.s2_image_size

    def load_model(self, device_map=None):
        if self.is_loaded:
            print('{} is already loaded, `load_model` called again, skipping.'.format(self.vision_tower_name))
            return

        self.image_processor = CLIPImageProcessor.from_pretrained(self.vision_tower_name)
        self.vision_tower = CLIPVisionModel.from_pretrained(self.vision_tower_name, device_map=device_map)
        self.vision_tower.requires_grad_(False)

        self.image_processor.size['shortest_edge'] = self.s2_image_size
        self.image_processor.crop_size['height'] = self.image_processor.crop_size['width'] = self.s2_image_size

        self.is_loaded = True

    @torch.no_grad()
    def forward_feature(self, images):
        image_forward_outs = self.vision_tower(images.to(device=self.device, dtype=self.dtype), output_hidden_states=True)
        image_features = self.feature_select(image_forward_outs).to(images.dtype)
        return image_features

    @torch.no_grad()
    def forward(self, images):
        if type(images) is list:
            image_features = []
            for image in images:
                image_feature = self.multiscale_forward(self.forward_feature, image.unsqueeze(0), img_sizes=self.s2_scales, max_split_size=self.s2_split_size)
                image_features.append(image_feature)
        else:
            image_features = self.multiscale_forward(self.forward_feature, images, img_sizes=self.s2_scales, max_split_size=self.s2_split_size)

        return image_features

    @property
    def hidden_size(self):
        return self.config.hidden_size * len(self.s2_scales)


def SCOPE(visual_feature_vectors, num_selected_token, cls_attn=None, alpha=1.0):
    """
    Batched version of SCOPE that processes all batch elements simultaneously.
    Args:
        visual_feature_vectors: [B, N, D] batch of feature vectors
        num_selected_token: Number of tokens to select per batch
        cls_attn: [B, N] batch of attention weights
    Returns:
        selected_idx: [B, K] selected token indices for each batch
        cosine_simi: [B, N, N] batch of cosine similarity matrices
    """
    # Calculate cosine similarity for all batches at once
    norm_vectors = visual_feature_vectors / visual_feature_vectors.norm(dim=-1, keepdim=True)
    cosine_simi = torch.bmm(norm_vectors, norm_vectors.transpose(1, 2))
    
    B, N = visual_feature_vectors.shape[:2]
    device = visual_feature_vectors.device
    dtype = visual_feature_vectors.dtype
    
    # Pre-allocate tensors for all batches
    selected = torch.zeros(B, N, dtype=torch.bool, device=device)
    selected_idx = torch.empty(B, num_selected_token, dtype=torch.long, device=device)
    cur_max = torch.zeros(B, N, dtype=dtype, device=device)
    
    # Precompute cls_attn ** alpha for all batches
    # alpha = float(os.environ.get('ALPHA', '1.0'))
    if cls_attn is not None:
        cls_attn_powered = cls_attn ** alpha
    else:
        cls_attn_powered = torch.ones(B, N, dtype=dtype, device=device)
    
    for i in range(num_selected_token):
        # Calculate gains for all batches simultaneously
        unselected_mask = ~selected
        gains = torch.maximum(
            torch.zeros(1, dtype=dtype, device=device),
            cosine_simi.masked_fill(~unselected_mask.unsqueeze(1), 0) - 
            cur_max.unsqueeze(2)
        ).sum(dim=1)
        
        # Apply attention weights
        combined = os.environ.get('COMBINED', 'multi')
        if combined == 'multi':
            gains = gains * cls_attn_powered
        elif combined == 'add':
            gains = gains + cls_attn_powered
        else:
            raise NotImplementedError
        # Mask out already selected tokens
        gains = gains.masked_fill(~unselected_mask, float('-inf'))
        
        # Find best elements for all batches
        best_idx = gains.argmax(dim=1)
        
        # Update states for all batches
        selected[torch.arange(B, device=device), best_idx] = True
        selected_idx[:, i] = best_idx
        cur_max = torch.maximum(cur_max, cosine_simi[torch.arange(B, device=device), best_idx])
    
    return selected_idx, cosine_simi