File size: 5,138 Bytes
b20701a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
'''

Models definition

'''

import math
import torch
import torch.nn as nn
import torch.nn.functional as F
from torchvision import models


class CRNN(nn.Module):
    def __init__(self, num_classes=100, hidden_size=256, resnet_pretrained_weights=None):
        super(CRNN, self).__init__()
        resnet = models.resnet18(weights=resnet_pretrained_weights)
        self.cnn = nn.Sequential(*list(resnet.children())[:-2])
        self.feature_dim = 512
        self.pool = nn.AdaptiveAvgPool2d((1, 1))
        self.rnn = nn.LSTM(self.feature_dim, hidden_size, batch_first=True)
        self.dropout = nn.Dropout(0.3)
        self.fc = nn.Linear(hidden_size, num_classes)

    def forward(self, x):
        B, T, C, H, W = x.size()
        x = x.view(B * T, C, H, W)
        features = self.cnn(x)
        pooled = self.pool(features).squeeze(-1).squeeze(-1)
        seq = pooled.view(B, T, self.feature_dim)
        rnn_out, _ = self.rnn(seq)
        final = self.dropout(rnn_out[:, -1, :])
        return self.fc(final)
    
    
class PositionalEncoding(nn.Module):
    '''Positional encoding for temporal sequence'''
    def __init__(self, d_model, max_len=64, dropout=0.1):
        super().__init__()
        self.dropout = nn.Dropout(p=dropout)

        pe = torch.zeros(max_len, d_model)
        position = torch.arange(0, max_len, dtype=torch.float).unsqueeze(1)
        div_term = torch.exp(torch.arange(0, d_model, 2).float() * (-math.log(10000.0) / d_model))

        pe[:, 0::2] = torch.sin(position * div_term)
        pe[:, 1::2] = torch.cos(position * div_term)
        pe = pe.unsqueeze(0)  # (1, max_len, d_model)

        self.register_buffer("pe", pe)

    def forward(self, x):
        # x: (B, T, d_model)
        x = x + self.pe[:, :x.size(1), :]
        return self.dropout(x)


class AttentionPooling(nn.Module):
    '''Attention pooling layer'''
    def __init__(self, dim):
        super().__init__()
        self.attention = nn.Sequential(
            nn.Linear(dim, dim // 4),
            nn.Tanh(),
            nn.Linear(dim // 4, 1)
        )

    def forward(self, x):
        # x: (B, T, dim)
        attn_weights = self.attention(x)  # (B, T, 1)
        attn_weights = F.softmax(attn_weights, dim=1)
        pooled = torch.sum(attn_weights * x, dim=1)  # (B, dim)
        return pooled
    

class ConvNeXtTransformer(nn.Module):
    '''

    ConvNeXt-Tiny + Transformer



    Input:  (B, T, C, H, W) = (B, 16, 3, 224, 224)

    Output: (B, num_classes) = (B, 100)

    '''
    def __init__(self, num_classes=100, hidden_size=256, convnext_pretrained_weights=None):
        super().__init__()

        # 1. ConvNeXt-Tiny Backbone
        convnext = models.convnext_tiny(weights=convnext_pretrained_weights)
        self.cnn = convnext.features
        self.pool = nn.AdaptiveAvgPool2d((1, 1))

        # ConvNeXt-Tiny output = 768
        self.feature_dim = 768

        # 2. Positional Encoding
        self.pos_encoder = PositionalEncoding(
            d_model=self.feature_dim,
            max_len=64,
            dropout=0.1
        )

        # 3. Transformer Encoder
        encoder_layer = nn.TransformerEncoderLayer(
            d_model=self.feature_dim,
            nhead=8,
            dim_feedforward=self.feature_dim * 4,
            dropout=0.3,
            activation="gelu",
            batch_first=True,
            norm_first=True
        )
        self.transformer = nn.TransformerEncoder(encoder_layer, num_layers=2)

        # 4. Attention Pooling
        self.attention_pool = AttentionPooling(self.feature_dim)

        # 5. Classifier
        self.fc = nn.Sequential(
            nn.LayerNorm(self.feature_dim),
            nn.Dropout(0.4),
            nn.Linear(self.feature_dim, num_classes)
        )

        self._init_weights()

    def _init_weights(self):
        for m in self.transformer.modules():
            if isinstance(m, nn.Linear):
                nn.init.trunc_normal_(m.weight, std=0.02)
                if m.bias is not None:
                    nn.init.zeros_(m.bias)

        for m in self.attention_pool.modules():
            if isinstance(m, nn.Linear):
                nn.init.trunc_normal_(m.weight, std=0.02)
                if m.bias is not None:
                    nn.init.zeros_(m.bias)

    def freeze_convnext_features(self, freeze_until=3):
        for i in range(freeze_until + 1):
            for p in self.cnn[i].parameters():
                p.requires_grad = False

    def forward(self, x):
        B, T, C, H, W = x.shape

        # CNN: (B, T, C, H, W) → (B, T, 768)
        x = x.view(B * T, C, H, W)
        x = self.cnn(x)
        x = self.pool(x)
        x = x.view(B, T, self.feature_dim)

        # Transformer: (B, T, 768) → (B, T, 768)
        x = self.pos_encoder(x)
        x = self.transformer(x)

        # Pooling: (B, T, 768) → (B, 768)
        x = self.attention_pool(x)

        # Classifier: (B, 768) → (B, num_classes)
        x = self.fc(x)

        return x