File size: 6,610 Bytes
841bef5
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import torch
import torch.nn.functional as F
from torch import nn
from .backbone import build_backbone
import pdb
import numpy as np
from typing import Optional

class TokenOCR(nn.Module):
    def __init__(self, backbone):
        """ Initializes the model.
        Parameters:
            backbone: torch module of the backbone to be used. See backbone.py
            transformer: torch module of the transformer architecture. See transformer.py
            num_classes: number of object classes

        """
        super().__init__()
        self.language_embedding = nn.Embedding(92553, 2048, padding_idx=2)
        for p in self.parameters():
            p.requires_grad = False

        self.backbone = backbone
        init_tau=np.log(10)
        init_b=-2.71
        # self.t_prime = nn.Parameter(torch.ones([]) * init_tau)
        # self.b = nn.Parameter(torch.ones([]) * init_b)
        self.kb = True
        self.upsample = nn.Sequential(
                    nn.ConvTranspose2d(
            in_channels=2048,
            out_channels=512,
            kernel_size=4,
            stride=2,
            padding=1,
            bias=False
        ),
        nn.SyncBatchNorm(512),
        nn.ConvTranspose2d(
            in_channels=512,
            out_channels=512,
            kernel_size=4,
            stride=2,
            padding=1,
            bias=False
        ),
        nn.SyncBatchNorm(512),
        )
        self.ocr_mlp = nn.Sequential(
            nn.Linear(512, 2048),
            nn.GELU(),
            nn.Linear(2048, 2048)
        )

    def forward(self, 
            pixel_values: torch.FloatTensor,
            input_ids: torch.LongTensor = None,
            image_flags: Optional[torch.LongTensor] = None,
            mask_values: Optional[torch.LongTensor] = None,
            masks_flags: Optional[torch.LongTensor] = None,
            mask_nums: Optional[torch.LongTensor] = None,
            ):
        image_flags = image_flags.squeeze(-1)
        try:
            input_embeds = self.language_embedding(input_ids).clone()
        except:
            print('error'*1000)
            import pdb; pdb.set_trace()
        # import pdb; pdb.set_trace()
        vit_embeds, vit_embeds_shape = self.extract_feature_custom(pixel_values) #(vit_batch_size, 16*16, 2048)
        nb, nl, nd = vit_embeds.shape
        h, w = vit_embeds_shape
        vit_embeds = vit_embeds.reshape(nb, h, w, nd)
        vit_embeds = vit_embeds.split(list(image_flags)) #[(vit_batch_size / B, h, w, C)]*B
        vit_batch_size = pixel_values.shape[0]

        B, N, C = input_embeds.shape
        try:
            assert sum(image_flags) == mask_values.shape[0]
        except:
            print((mask_values.shape, image_flags, mask_nums))
        
        mask_values = torch.nn.functional.interpolate(mask_values.float(), size=(h, w), mode='bilinear', align_corners=False) #(128, 128)
        masks = mask_values.split(list(image_flags)) #[(vit_batch_size / B, N, 448, 448)]*B

        
        masks_flags = masks_flags.chunk(B)
        token_features = []
        input_embedings = []
        masked_input_ids = []
        masked_zero_bools = []
        for i, vit_embed in enumerate(vit_embeds):
            current_token = masks_flags[i].sum()
            mask = masks[i]
            limit_num = mask.shape[1]
            mask = mask.permute(1,0,2,3).reshape(limit_num, -1) > 0
            max_cluster_index = mask.sum(-1)
            zero_bool = max_cluster_index != 0
            # import pdb; pdb.set_trace()
            mask[~zero_bool] = 1 #for addressing bflost16 bug
            new_max_cluster_index = mask.sum(-1)
            mask = mask / new_max_cluster_index.unsqueeze(-1)
            token_feature = torch.matmul(mask.to(vit_embed), vit_embed.reshape(-1, vit_embed.shape[-1]))
            token_features.extend(token_feature)
            input_embedings.extend(input_embeds[i, :])
            masked_input_ids.extend(input_ids[i, zero_bool])
            masked_zero_bools.append(zero_bool)

        masked_zero_bools = torch.cat(masked_zero_bools)
        token_features = torch.stack(token_features)
        input_embedings= torch.stack(input_embedings)

        loss2 = F.mse_loss(token_features, input_embedings, reduction='none')[masked_zero_bools].sum(1).sqrt().mean()
        token_features = token_features / token_features.norm(dim=1, keepdim=True)
        input_embedings = input_embedings / input_embedings.norm(dim=1, keepdim=True)
        # cosine similarity as logits
        similarity = F.cosine_similarity(token_features, input_embedings, dim=1)
        loss1 = (1 - similarity[masked_zero_bools]).mean()
        # loss_d = loss1 + loss2
        # if rank == 0:
            # print(f'loss1:{loss_d}')

        ###siglip
        # masked_input_ids = torch.stack(masked_input_ids)
        # label_matrix = (masked_input_ids.unsqueeze(0) == masked_input_ids.unsqueeze(1)).int()
        # label_matrix = 2 * label_matrix - 1
        # if self.kb:
        #     logits = (input_embedings[masked_zero_bools] @ token_features[masked_zero_bools].t()) * self.t_prime.to(input_embedings.device).exp() + self.b.to(input_embedings.device)
        # else:
        #     logits = (input_embedings[masked_zero_bools] @ token_features[masked_zero_bools].t()) * self.t_prime.to(input_embedings.device).exp() - 8.9375
        # loss_s = -torch.sum(F.logsigmoid(label_matrix * logits)) / logits.shape[0]
        # if rank == 0:
        #     print(f'loss2:{loss_s}')
        return loss1, loss2

    def forward_tokenocr(self, pixel_values):
        vit_embeds = self.backbone(pixel_values)
        vit_embeds = vit_embeds['0']
        h, w = vit_embeds.shape[2], vit_embeds.shape[3]
        vit_embeds = self.upsample(vit_embeds)
        vit_embeds = vit_embeds.reshape(vit_embeds.shape[0], -1, vit_embeds.shape[-2] * vit_embeds.shape[-1])
        vit_embeds = self.ocr_mlp(vit_embeds.permute(0, 2, 1))
        return vit_embeds, (h*4, w*4)


class MLP(nn.Module):
    """ Very simple multi-layer perceptron (also called FFN)"""

    def __init__(self, input_dim, hidden_dim, output_dim, num_layers):
        super().__init__()
        self.num_layers = num_layers
        h = [hidden_dim] * (num_layers - 1)
        self.layers = nn.ModuleList(nn.Linear(n, k) for n, k in zip([input_dim] + h, h + [output_dim]))

    def forward(self, x):
        for i, layer in enumerate(self.layers):
            x = F.relu(layer(x)) if i < self.num_layers - 1 else layer(x)
        return x


def build(args):
    backbone = build_backbone(args)
    model = TokenOCR(backbone)
    return model