File size: 8,695 Bytes
c3c908f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import torch
import torch.nn as nn

class audiocnn(nn.Module):
    def __init__(self, num_classes=2):
        super(audiocnn, self).__init__()
        self.conv_block = nn.Sequential(
        nn.Conv2d(1, 16, kernel_size=3, padding=1),
        nn.ReLU(),
        nn.MaxPool2d(2),
        nn.Conv2d(16, 32, kernel_size=3, padding=1),
        nn.ReLU(),
        nn.MaxPool2d(2),
        nn.AdaptiveAvgPool2d((4,4))  # 최종 -> (B,32,4,4)
    )
        self.fc_block = nn.Sequential(
            nn.Linear(32*4*4, 128),
            nn.ReLU(),
            nn.Linear(128, num_classes)
        )

    def forward(self, x):
        x = self.conv_block(x)
        # x.shape: (B,32,new_freq,new_time)

        # 1) Flatten
        B, C, H, W = x.shape  # 동적 shape
        x = x.view(B, -1)     # (B, 32*H*W)

        # 2) FC
        x = self.fc_block(x)
        return x
    
class AudioCNN(nn.Module):
    def __init__(self, embed_dim=512):
        super(AudioCNN, self).__init__()
        self.conv_block = nn.Sequential(
            nn.Conv2d(1, 16, kernel_size=3, padding=1),
            nn.ReLU(),
            nn.MaxPool2d(2),
            nn.Conv2d(16, 32, kernel_size=3, padding=1),
            nn.ReLU(),
            nn.MaxPool2d(2),
            nn.AdaptiveAvgPool2d((4, 4))  # 최종 -> (B, 32, 4, 4)
        )
        self.projection = nn.Linear(32 * 4 * 4, embed_dim)

    def forward(self, x):
        x = self.conv_block(x)
        B, C, H, W = x.shape
        x = x.view(B, -1)  # Flatten (B, C * H * W)
        x = self.projection(x)  # Project to embed_dim
        return x

class ViTDecoder(nn.Module):
    def __init__(self, embed_dim=512, num_heads=8, num_layers=6, num_classes=2):
        super(ViTDecoder, self).__init__()

        # Transformer layers
        encoder_layer = nn.TransformerEncoderLayer(d_model=embed_dim, nhead=num_heads)
        self.transformer = nn.TransformerEncoder(encoder_layer, num_layers=num_layers)

        # Classification head
        self.classifier = nn.Sequential(
            nn.LayerNorm(embed_dim),
            nn.Linear(embed_dim, num_classes)
        )

    def forward(self, x):
        # Transformer expects input of shape (seq_len, batch, embed_dim)
        x = x.unsqueeze(1).permute(1, 0, 2)  # Add sequence dim (1, B, embed_dim)
        x = self.transformer(x)  # Pass through Transformer
        x = x.mean(dim=0)  # Take the mean over the sequence dimension (B, embed_dim)

        x = self.classifier(x)  # Classification head
        return x

class AudioCNNWithViTDecoder(nn.Module):
    def __init__(self, embed_dim=512, num_heads=8, num_layers=6, num_classes=2):
        super(AudioCNNWithViTDecoder, self).__init__()
        self.encoder = AudioCNN(embed_dim=embed_dim)
        self.decoder = ViTDecoder(embed_dim=embed_dim, num_heads=num_heads, num_layers=num_layers, num_classes=num_classes)

    def forward(self, x):
        x = self.encoder(x)  # Pass through AudioCNN encoder
        x = self.decoder(x)  # Pass through ViT decoder
        return x


# class AudioCNN(nn.Module):
#     def __init__(self, num_classes=2):
#         super(AudioCNN, self).__init__()
#         self.conv_block = nn.Sequential(
#         nn.Conv2d(1, 16, kernel_size=3, padding=1),
#         nn.ReLU(),
#         nn.MaxPool2d(2),
#         nn.Conv2d(16, 32, kernel_size=3, padding=1),
#         nn.ReLU(),
#         nn.MaxPool2d(2),
#         nn.AdaptiveAvgPool2d((4,4))  # 최종 -> (B,32,4,4)
#     )
#         self.fc_block = nn.Sequential(
#             nn.Linear(32*4*4, 128),
#             nn.ReLU(),
#             nn.Linear(128, num_classes)
#         )


#     def forward(self, x):
#         x = self.conv_block(x)
#         # x.shape: (B,32,new_freq,new_time)

#         # 1) Flatten
#         B, C, H, W = x.shape  # 동적 shape
#         x = x.view(B, -1)     # (B, 32*H*W)

#         # 2) FC
#         x = self.fc_block(x)
#         return x



class audio_crossattn(nn.Module):
    def __init__(self, embed_dim=512):
        super(audio_crossattn, self).__init__()
        self.conv_block = nn.Sequential(
            nn.Conv2d(1, 16, kernel_size=3, padding=1),
            nn.ReLU(),
            nn.MaxPool2d(2),
            nn.Conv2d(16, 32, kernel_size=3, padding=1),
            nn.ReLU(),
            nn.MaxPool2d(2),
            nn.AdaptiveAvgPool2d((4, 4))  # 최종 출력 -> (B, 32, 4, 4)
        )
        self.projection = nn.Linear(32 * 4 * 4, embed_dim)

    def forward(self, x):
        x = self.conv_block(x)  # Convolutional feature extraction
        B, C, H, W = x.shape
        x = x.view(B, -1)  # Flatten (B, C * H * W)
        x = self.projection(x)  # Linear projection to embed_dim
        return x
    

class CrossAttentionLayer(nn.Module):
    def __init__(self, embed_dim, num_heads):
        super(CrossAttentionLayer, self).__init__()
        self.multihead_attn = nn.MultiheadAttention(embed_dim, num_heads, batch_first=True)
        self.layer_norm = nn.LayerNorm(embed_dim)
        self.feed_forward = nn.Sequential(
            nn.Linear(embed_dim, embed_dim * 4),
            nn.ReLU(),
            nn.Linear(embed_dim * 4, embed_dim)
        )

    def forward(self, x, cross_input):
        # Cross-attention between x and cross_input
        attn_output, _ = self.multihead_attn(query=x, key=cross_input, value=cross_input)
        x = self.layer_norm(x + attn_output)  # Add & Norm
        feed_forward_output = self.feed_forward(x)
        x = self.layer_norm(x + feed_forward_output)  # Add & Norm
        return x

class ViTDecoderWithCrossAttention(nn.Module):
    def __init__(self, embed_dim=512, num_heads=8, num_layers=6, num_classes=2):
        super(ViTDecoderWithCrossAttention, self).__init__()

        # Cross-Attention layers
        self.cross_attention_layers = nn.ModuleList([
            CrossAttentionLayer(embed_dim, num_heads) for _ in range(num_layers)
        ])

        # Transformer Encoder layers
        encoder_layer = nn.TransformerEncoderLayer(d_model=embed_dim, nhead=num_heads)
        self.transformer = nn.TransformerEncoder(encoder_layer, num_layers=num_layers)

        # Classification head
        self.classifier = nn.Sequential(
            nn.LayerNorm(embed_dim),
            nn.Linear(embed_dim, num_classes)
        )

    def forward(self, x, cross_attention_input):
        # Pass through Cross-Attention layers
        for layer in self.cross_attention_layers:
            x = layer(x, cross_attention_input)

        # Transformer expects input of shape (seq_len, batch, embed_dim)
        x = x.unsqueeze(1).permute(1, 0, 2)  # Add sequence dim (1, B, embed_dim)
        x = self.transformer(x)  # Pass through Transformer
        embedding = x.mean(dim=0)  # Take the mean over the sequence dimension (B, embed_dim)

        # Classification head
        x = self.classifier(embedding)
        return x, embedding

# class AudioCNNWithViTDecoderAndCrossAttention(nn.Module):
#     def __init__(self, embed_dim=512, num_heads=8, num_layers=6, num_classes=2):
#         super(AudioCNNWithViTDecoderAndCrossAttention, self).__init__()
#         self.encoder = audio_crossattn(embed_dim=embed_dim)
#         self.decoder = ViTDecoderWithCrossAttention(embed_dim=embed_dim, num_heads=num_heads, num_layers=num_layers, num_classes=num_classes)

#     def forward(self, x, cross_attention_input):
#         # Pass through AudioCNN encoder
#         x = self.encoder(x)

#         # Pass through ViTDecoder with Cross-Attention
#         x = self.decoder(x, cross_attention_input)
#         return x
class CCV(nn.Module):
    def __init__(self, embed_dim=512, num_heads=8, num_layers=6, num_classes=2, freeze_feature_extractor=True):
        super(CCV, self).__init__()
        self.encoder = AudioCNN(embed_dim=embed_dim)
        self.decoder = ViTDecoderWithCrossAttention(embed_dim=embed_dim, num_heads=num_heads, num_layers=num_layers, num_classes=num_classes)
        if freeze_feature_extractor:
            for param in self.encoder.parameters():
                param.requires_grad = False
            for param in self.decoder.parameters():
                param.requires_grad = False
    def forward(self, x, cross_attention_input=None):
        # Pass through AudioCNN encoder
        x = self.encoder(x)

        # If cross_attention_input is not provided, use the encoder output
        if cross_attention_input is None:
            cross_attention_input = x

        # Pass through ViTDecoder with Cross-Attention
        x, embedding = self.decoder(x, cross_attention_input)
        return x, embedding

#---------------------------------------------------------
'''
audiocnn weight frozen
crossatten decoder -lora tuning
'''