File size: 8,934 Bytes
4853e68
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""Transformer-based model for sign language recognition"""

import torch
import torch.nn as nn
import torch.nn.functional as F
from typing import Optional
import math


class PositionalEncoding(nn.Module):
    """Positional encoding for transformer"""
    
    def __init__(self, d_model: int, max_len: int = 5000, dropout: float = 0.1):
        super().__init__()
        self.dropout = nn.Dropout(p=dropout)
        
        # Create positional encoding matrix
        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)
        
        self.register_buffer('pe', pe)
    
    def forward(self, x: torch.Tensor) -> torch.Tensor:
        """

        Args:

            x: Tensor of shape (batch_size, seq_len, d_model)

        """
        x = x + self.pe[:, :x.size(1), :]
        return self.dropout(x)


class MediaPipeTransformer(nn.Module):
    """Transformer model for MediaPipe landmarks"""
    
    def __init__(self,

                 input_dim: int = 258,  # 33*4 + 21*3 + 21*3 (pose + hands, no face)

                 d_model: int = 256,

                 nhead: int = 8,

                 num_encoder_layers: int = 4,

                 dim_feedforward: int = 1024,

                 dropout: float = 0.3,

                 num_classes: int = 100,

                 max_seq_length: int = 64):
        """

        Args:

            input_dim: Dimension of input landmarks

            d_model: Dimension of model embeddings

            nhead: Number of attention heads

            num_encoder_layers: Number of transformer encoder layers

            dim_feedforward: Dimension of feedforward network

            dropout: Dropout rate

            num_classes: Number of output classes

            max_seq_length: Maximum sequence length

        """
        super().__init__()
        
        self.input_dim = input_dim
        self.d_model = d_model
        self.num_classes = num_classes
        
        # Input projection
        self.input_projection = nn.Sequential(
            nn.Linear(input_dim, d_model),
            nn.LayerNorm(d_model),
            nn.Dropout(dropout)
        )
        
        # Positional encoding
        self.pos_encoder = PositionalEncoding(d_model, max_seq_length, dropout)
        
        # Transformer encoder
        encoder_layer = nn.TransformerEncoderLayer(
            d_model=d_model,
            nhead=nhead,
            dim_feedforward=dim_feedforward,
            dropout=dropout,
            activation='gelu',
            batch_first=True
        )
        self.transformer_encoder = nn.TransformerEncoder(
            encoder_layer,
            num_layers=num_encoder_layers
        )
        
        # Classification head
        self.classifier = nn.Sequential(
            nn.Linear(d_model, d_model // 2),
            nn.LayerNorm(d_model // 2),
            nn.GELU(),
            nn.Dropout(dropout),
            nn.Linear(d_model // 2, num_classes)
        )
        
        # Initialize weights
        self._init_weights()
    
    def _init_weights(self):
        """Initialize weights"""
        for p in self.parameters():
            if p.dim() > 1:
                nn.init.xavier_uniform_(p)
    
    def forward(self, x: torch.Tensor, mask: Optional[torch.Tensor] = None) -> torch.Tensor:
        """

        Forward pass

        

        Args:

            x: Input tensor of shape (batch_size, seq_len, input_dim)

            mask: Optional padding mask of shape (batch_size, seq_len)

            

        Returns:

            Output logits of shape (batch_size, num_classes)

        """
        # Input projection
        x = self.input_projection(x)  # (batch, seq_len, d_model)
        
        # Add positional encoding
        x = self.pos_encoder(x)
        
        # Create attention mask for padding
        if mask is not None:
            # Convert padding mask to attention mask
            # True values are masked (padding), False values are not masked
            attn_mask = mask.unsqueeze(1).unsqueeze(2)  # (batch, 1, 1, seq_len)
            attn_mask = attn_mask.expand(-1, -1, x.size(1), -1)  # (batch, 1, seq_len, seq_len)
            attn_mask = attn_mask.squeeze(1)  # (batch, seq_len, seq_len)
        else:
            attn_mask = None
        
        # Transformer encoding
        x = self.transformer_encoder(x, src_key_padding_mask=mask)  # (batch, seq_len, d_model)
        
        # Global average pooling over time dimension
        if mask is not None:
            # Mask out padding before pooling
            mask_expanded = (~mask).unsqueeze(-1).float()  # (batch, seq_len, 1)
            x = (x * mask_expanded).sum(dim=1) / mask_expanded.sum(dim=1)  # (batch, d_model)
        else:
            x = x.mean(dim=1)  # (batch, d_model)
        
        # Classification
        logits = self.classifier(x)  # (batch, num_classes)
        
        return logits
    
    def get_attention_weights(self, x: torch.Tensor, mask: Optional[torch.Tensor] = None):
        """Get attention weights for visualization"""
        # This is a simplified version - full implementation would require
        # modifying the transformer encoder to return attention weights
        with torch.no_grad():
            x = self.input_projection(x)
            x = self.pos_encoder(x)
            # Note: Standard PyTorch transformer doesn't return attention weights
            # You would need to implement a custom version for this
        return None


class LSTMModel(nn.Module):
    """LSTM baseline model for comparison"""
    
    def __init__(self,

                 input_dim: int = 258,

                 hidden_dim: int = 256,

                 num_layers: int = 2,

                 dropout: float = 0.3,

                 num_classes: int = 100,

                 bidirectional: bool = True):
        """

        Args:

            input_dim: Dimension of input landmarks

            hidden_dim: Hidden dimension of LSTM

            num_layers: Number of LSTM layers

            dropout: Dropout rate

            num_classes: Number of output classes

            bidirectional: Whether to use bidirectional LSTM

        """
        super().__init__()
        
        self.input_dim = input_dim
        self.hidden_dim = hidden_dim
        self.num_layers = num_layers
        self.bidirectional = bidirectional
        
        # LSTM
        self.lstm = nn.LSTM(
            input_size=input_dim,
            hidden_size=hidden_dim,
            num_layers=num_layers,
            dropout=dropout if num_layers > 1 else 0,
            bidirectional=bidirectional,
            batch_first=True
        )
        
        # Classification head
        lstm_output_dim = hidden_dim * 2 if bidirectional else hidden_dim
        self.classifier = nn.Sequential(
            nn.Linear(lstm_output_dim, hidden_dim),
            nn.LayerNorm(hidden_dim),
            nn.ReLU(),
            nn.Dropout(dropout),
            nn.Linear(hidden_dim, num_classes)
        )
    
    def forward(self, x: torch.Tensor, mask: Optional[torch.Tensor] = None) -> torch.Tensor:
        """

        Forward pass

        

        Args:

            x: Input tensor of shape (batch_size, seq_len, input_dim)

            mask: Optional padding mask

            

        Returns:

            Output logits of shape (batch_size, num_classes)

        """
        # LSTM
        lstm_out, (h_n, c_n) = self.lstm(x)  # lstm_out: (batch, seq_len, hidden_dim * num_directions)
        
        # Use last hidden state
        if self.bidirectional:
            # Concatenate forward and backward hidden states
            hidden = torch.cat([h_n[-2], h_n[-1]], dim=1)  # (batch, hidden_dim * 2)
        else:
            hidden = h_n[-1]  # (batch, hidden_dim)
        
        # Classification
        logits = self.classifier(hidden)  # (batch, num_classes)
        
        return logits


def create_model(model_name: str = "transformer", **kwargs) -> nn.Module:
    """

    Factory function to create models

    

    Args:

        model_name: Name of model ('transformer' or 'lstm')

        **kwargs: Model-specific arguments

        

    Returns:

        Model instance

    """
    if model_name.lower() == "transformer":
        return MediaPipeTransformer(**kwargs)
    elif model_name.lower() == "lstm":
        return LSTMModel(**kwargs)
    else:
        raise ValueError(f"Unknown model: {model_name}")