models

two_stream_lipnet.py

@@ -0,0 +1,113 @@
+import torch
+import torch.nn as nn
+import torch.nn.init as init
+import math
+
+
+class TwoStreamLipNet(torch.nn.Module):
+    def __init__(self, dropout_p=0.5, coord_input_dim=40, coord_hidden_dim=128):
+        super(TwoStreamLipNet, self).__init__()
+        self.conv1 = nn.Conv3d(3, 32, (3, 5, 5), (1, 2, 2), (1, 2, 2))
+        self.pool1 = nn.MaxPool3d((1, 2, 2), (1, 2, 2))
+
+        self.conv2 = nn.Conv3d(32, 64, (3, 5, 5), (1, 1, 1), (1, 2, 2))
+        self.pool2 = nn.MaxPool3d((1, 2, 2), (1, 2, 2))
+
+        self.conv3 = nn.Conv3d(64, 96, (3, 3, 3), (1, 1, 1), (1, 1, 1))
+        self.pool3 = nn.MaxPool3d((1, 2, 2), (1, 2, 2))
+
+        self.gru1 = nn.GRU(96 * 4 * 8, 256, 1, bidirectional=True)
+        self.gru2 = nn.GRU(512, 256, 1, bidirectional=True)
+
+        self.FC = nn.Linear(512 + 2 * coord_hidden_dim, 27 + 1)
+        self.dropout_p = dropout_p
+
+        self.relu = nn.ReLU(inplace=True)
+        self.dropout = nn.Dropout(self.dropout_p)
+        self.dropout3d = nn.Dropout3d(self.dropout_p)
+
+        # New GRU layer for lip coordinates
+        self.coord_gru = nn.GRU(
+            coord_input_dim, coord_hidden_dim, 1, bidirectional=True
+        )
+
+        self._init()
+
+    def _init(self):
+        init.kaiming_normal_(self.conv1.weight, nonlinearity="relu")
+        init.constant_(self.conv1.bias, 0)
+
+        init.kaiming_normal_(self.conv2.weight, nonlinearity="relu")
+        init.constant_(self.conv2.bias, 0)
+
+        init.kaiming_normal_(self.conv3.weight, nonlinearity="relu")
+        init.constant_(self.conv3.bias, 0)
+
+        init.kaiming_normal_(self.FC.weight, nonlinearity="sigmoid")
+        init.constant_(self.FC.bias, 0)
+
+        for m in (self.gru1, self.gru2):
+            stdv = math.sqrt(2 / (96 * 3 * 6 + 256))
+            for i in range(0, 256 * 3, 256):
+                init.uniform_(
+                    m.weight_ih_l0[i : i + 256],
+                    -math.sqrt(3) * stdv,
+                    math.sqrt(3) * stdv,
+                )
+                init.orthogonal_(m.weight_hh_l0[i : i + 256])
+                init.constant_(m.bias_ih_l0[i : i + 256], 0)
+                init.uniform_(
+                    m.weight_ih_l0_reverse[i : i + 256],
+                    -math.sqrt(3) * stdv,
+                    math.sqrt(3) * stdv,
+                )
+                init.orthogonal_(m.weight_hh_l0_reverse[i : i + 256])
+                init.constant_(m.bias_ih_l0_reverse[i : i + 256], 0)
+
+    def forward(self, x, coords):
+        # branch 1
+        x = self.conv1(x)
+        x = self.relu(x)
+        x = self.dropout3d(x)
+        x = self.pool1(x)
+
+        x = self.conv2(x)
+        x = self.relu(x)
+        x = self.dropout3d(x)
+        x = self.pool2(x)
+
+        x = self.conv3(x)
+        x = self.relu(x)
+        x = self.dropout3d(x)
+        x = self.pool3(x)
+
+        # (B, C, T, H, W)->(T, B, C, H, W)
+        x = x.permute(2, 0, 1, 3, 4).contiguous()
+        # (B, C, T, H, W)->(T, B, C*H*W)
+        x = x.view(x.size(0), x.size(1), -1)
+
+        self.gru1.flatten_parameters()
+        self.gru2.flatten_parameters()
+
+        x, h = self.gru1(x)
+        x = self.dropout(x)
+        x, h = self.gru2(x)
+        x = self.dropout(x)
+
+        # branch 2
+        # Process lip coordinates through GRU
+        self.coord_gru.flatten_parameters()
+
+        # (B, T, N, C)->(T, B, C, N, C)
+        coords = coords.permute(1, 0, 2, 3).contiguous()
+        # (T, B, C, N, C)->(T, B, C, N*C)
+        coords = coords.view(coords.size(0), coords.size(1), -1)
+        coords, _ = self.coord_gru(coords)
+        coords = self.dropout(coords)
+
+        # combine the two branches
+        combined = torch.cat((x, coords), dim=2)
+
+        x = self.FC(combined)
+        x = x.permute(1, 0, 2).contiguous()
+        return x