Add student_mobilenet.py

student_mobilenet.py

@@ -0,0 +1,78 @@
+"""
+student_mobilenet.py
+MobileNetV2-based lightweight student model for DLoc knowledge distillation.
+
+Architecture (from professor's Stage 1 spec):
+  Encoder: MobileNetV2 backbone (first conv modified for 4-channel input)
+  Decoder: Upsample + Conv (320->128->64->32->1) + Sigmoid
+  Output: [B, 1, 161, 361] location heatmap
+
+~2.3M params vs DLoc's ~15M params.
+"""
+import torch
+import torch.nn as nn
+from torchvision.models import mobilenet_v2, MobileNet_V2_Weights
+
+
+class MobileNetStudent(nn.Module):
+    # Target output spatial size
+    OUTPUT_SIZE = (161, 361)
+
+    def __init__(self, input_nc=4):
+        super().__init__()
+
+        # Load MobileNetV2 backbone (features[0:18] = 320-channel output)
+        mobilenet = mobilenet_v2(weights=MobileNet_V2_Weights.DEFAULT)
+        self.encoder = mobilenet.features[:-1]  # drop final 320->1280 conv
+
+        # Replace first conv: 3-channel RGB -> 4-channel AoA
+        old_conv = self.encoder[0][0]
+        self.encoder[0][0] = nn.Conv2d(
+            input_nc, old_conv.out_channels,
+            kernel_size=old_conv.kernel_size,
+            stride=old_conv.stride,
+            padding=old_conv.padding,
+            bias=False,
+        )
+
+        # Decoder: 320 -> 128 -> 64 -> 32 -> 1
+        self.decoder = nn.Sequential(
+            nn.Upsample(scale_factor=2, mode='bilinear', align_corners=False),
+            nn.Conv2d(320, 128, kernel_size=3, padding=1, bias=False),
+            nn.BatchNorm2d(128),
+            nn.ReLU(inplace=True),
+
+            nn.Upsample(scale_factor=2, mode='bilinear', align_corners=False),
+            nn.Conv2d(128, 64, kernel_size=3, padding=1, bias=False),
+            nn.BatchNorm2d(64),
+            nn.ReLU(inplace=True),
+
+            nn.Upsample(scale_factor=2, mode='bilinear', align_corners=False),
+            nn.Conv2d(64, 32, kernel_size=3, padding=1, bias=False),
+            nn.BatchNorm2d(32),
+            nn.ReLU(inplace=True),
+
+            nn.Conv2d(32, 1, kernel_size=1),
+            nn.Sigmoid(),
+        )
+
+    def forward(self, x):
+        feat = self.encoder(x)
+        out = self.decoder(feat)
+        # Ensure exact output size (MobileNetV2 downsamples aggressively)
+        if out.shape[2:] != self.OUTPUT_SIZE:
+            out = nn.functional.interpolate(out, size=self.OUTPUT_SIZE, mode='bilinear', align_corners=False)
+        return out
+
+
+if __name__ == "__main__":
+    model = MobileNetStudent(input_nc=4)
+    n = sum(p.numel() for p in model.parameters())
+    print(f"Parameters: {n:,} ({n/1e6:.2f}M)")
+    x = torch.randn(2, 4, 161, 361)
+    with torch.no_grad():
+        y = model(x)
+    print(f"Input:  {x.shape}")
+    print(f"Output: {y.shape}")
+    assert y.shape == (2, 1, 161, 361), f"Wrong shape: {y.shape}"
+    print("OK")