model_resnet3d.py

"""
3D ResNet for tornado detection + prediction.

Configs:
  ResNet3D-18:  BasicBlock, [2,2,2,2]  (~11M params)
  ResNet3D-34:  BasicBlock, [3,4,6,3]  (~21M params)
  ResNet3D-50:  Bottleneck, [3,4,6,3]  (~40M params)

Input: (B, 24, 8, 128, 128) — 24 dual-pol channels, 8 time frames, 128x128 grid
Output: (B, 4) — [det_neg, det_pos, pred_neg, pred_pos]
"""
import torch
import torch.nn as nn


class BasicBlock3D(nn.Module):
    expansion = 1

    def __init__(self, in_planes, planes, stride=1, downsample=None):
        super().__init__()
        self.conv1 = nn.Conv3d(in_planes, planes, kernel_size=3, stride=stride,
                               padding=1, bias=False)
        self.bn1 = nn.BatchNorm3d(planes)
        self.relu = nn.ReLU(inplace=True)
        self.conv2 = nn.Conv3d(planes, planes, kernel_size=3, stride=1,
                               padding=1, bias=False)
        self.bn2 = nn.BatchNorm3d(planes)
        self.downsample = downsample

    def forward(self, x):
        identity = x
        out = self.relu(self.bn1(self.conv1(x)))
        out = self.bn2(self.conv2(out))
        if self.downsample is not None:
            identity = self.downsample(x)
        out += identity
        return self.relu(out)


class Bottleneck3D(nn.Module):
    expansion = 4

    def __init__(self, in_planes, planes, stride=1, downsample=None):
        super().__init__()
        self.conv1 = nn.Conv3d(in_planes, planes, kernel_size=1, bias=False)
        self.bn1 = nn.BatchNorm3d(planes)
        self.conv2 = nn.Conv3d(planes, planes, kernel_size=3, stride=stride,
                               padding=1, bias=False)
        self.bn2 = nn.BatchNorm3d(planes)
        self.conv3 = nn.Conv3d(planes, planes * self.expansion, kernel_size=1, bias=False)
        self.bn3 = nn.BatchNorm3d(planes * self.expansion)
        self.relu = nn.ReLU(inplace=True)
        self.downsample = downsample

    def forward(self, x):
        identity = x
        out = self.relu(self.bn1(self.conv1(x)))
        out = self.relu(self.bn2(self.conv2(out)))
        out = self.bn3(self.conv3(out))
        if self.downsample is not None:
            identity = self.downsample(x)
        out += identity
        return self.relu(out)


class ResNet3D(nn.Module):
    """3D ResNet backbone. Returns feature vector of size 512 * block.expansion."""

    def __init__(self, block, layers, in_channels=24):
        super().__init__()
        self.in_planes = 64

        # Initial conv: don't downsample time aggressively
        self.conv1 = nn.Conv3d(in_channels, 64, kernel_size=(3, 7, 7),
                               stride=(1, 2, 2), padding=(1, 3, 3), bias=False)
        self.bn1 = nn.BatchNorm3d(64)
        self.relu = nn.ReLU(inplace=True)
        self.maxpool = nn.MaxPool3d(kernel_size=(1, 3, 3), stride=(1, 2, 2), padding=(0, 1, 1))

        # Residual layers — spatial downsampling in layers 2-4, temporal in layer 3
        self.layer1 = self._make_layer(block, 64, layers[0], stride=1)
        self.layer2 = self._make_layer(block, 128, layers[1], stride=2)           # spatial /2
        self.layer3 = self._make_layer(block, 256, layers[2], stride=(2, 2, 2))   # temporal /2, spatial /2
        self.layer4 = self._make_layer(block, 512, layers[3], stride=(2, 2, 2))   # temporal /2, spatial /2

        self.avgpool = nn.AdaptiveAvgPool3d(1)
        self.feat_dim = 512 * block.expansion

        # Weight initialization
        for m in self.modules():
            if isinstance(m, nn.Conv3d):
                nn.init.kaiming_normal_(m.weight, mode="fan_out", nonlinearity="relu")
            elif isinstance(m, nn.BatchNorm3d):
                nn.init.constant_(m.weight, 1)
                nn.init.constant_(m.bias, 0)

    def _make_layer(self, block, planes, num_blocks, stride=1):
        downsample = None
        if stride != 1 or self.in_planes != planes * block.expansion:
            if isinstance(stride, int):
                s = stride
            else:
                s = stride
            downsample = nn.Sequential(
                nn.Conv3d(self.in_planes, planes * block.expansion,
                          kernel_size=1, stride=s, bias=False),
                nn.BatchNorm3d(planes * block.expansion),
            )

        layers = [block(self.in_planes, planes, stride, downsample)]
        self.in_planes = planes * block.expansion
        for _ in range(1, num_blocks):
            layers.append(block(self.in_planes, planes))
        return nn.Sequential(*layers)

    def forward(self, x):
        # x: (B, C, T, H, W) = (B, 24, 8, 128, 128)
        x = self.relu(self.bn1(self.conv1(x)))   # (B, 64, 8, 64, 64)
        x = self.maxpool(x)                       # (B, 64, 8, 32, 32)
        x = self.layer1(x)                        # (B, 64, 8, 32, 32)
        x = self.layer2(x)                        # (B, 128, 8, 16, 16)
        x = self.layer3(x)                        # (B, 256, 4, 8, 8)
        x = self.layer4(x)                        # (B, 512, 2, 4, 4)
        x = self.avgpool(x)                       # (B, 512, 1, 1, 1)
        return x.flatten(1)                        # (B, 512)


class DualHeadResNet3D(nn.Module):
    """Dual-head wrapper: detection + prediction heads on shared ResNet3D backbone."""

    def __init__(self, block, layers, in_channels=24, drop_rate=0.3):
        super().__init__()
        self.backbone = ResNet3D(block, layers, in_channels)
        feat_dim = self.backbone.feat_dim

        self.dropout = nn.Dropout(drop_rate)
        self.detect_head = nn.Linear(feat_dim, 2)
        self.predict_head = nn.Linear(feat_dim, 2)

        # Init heads
        for head in [self.detect_head, self.predict_head]:
            nn.init.normal_(head.weight, std=0.01)
            nn.init.zeros_(head.bias)

    def forward(self, x):
        # x: (B, C, T, H, W)
        features = self.backbone(x)          # (B, feat_dim)
        # FP32 cast before heads to prevent Inf grads under AMP
        features = features.float()
        features = self.dropout(features)
        det = self.detect_head(features)     # (B, 2)
        pred = self.predict_head(features)   # (B, 2)
        return torch.cat([det, pred], dim=1) # (B, 4)


# --- Factory functions ---

CONFIGS = {
    "resnet18": {"block": BasicBlock3D, "layers": [2, 2, 2, 2]},
    "resnet34": {"block": BasicBlock3D, "layers": [3, 4, 6, 3]},
    "resnet50": {"block": Bottleneck3D, "layers": [3, 4, 6, 3]},
}


def build_resnet3d(config="resnet34", in_channels=24, drop_rate=0.3):
    """Build a DualHeadResNet3D from config name."""
    cfg = CONFIGS[config]
    return DualHeadResNet3D(cfg["block"], cfg["layers"], in_channels, drop_rate)


if __name__ == "__main__":
    print("=== ResNet3D Model Configs ===\n")

    for name in ["resnet18", "resnet34", "resnet50"]:
        model = build_resnet3d(name)
        n_params = sum(p.numel() for p in model.parameters() if p.requires_grad)
        n_backbone = sum(p.numel() for p in model.backbone.parameters() if p.requires_grad)
        print(f"{name}: {n_params:>12,} total params  ({n_backbone:,} backbone)")

    # Forward pass test
    print("\nForward pass test (resnet34)...")
    model = build_resnet3d("resnet34")
    x = torch.randn(2, 24, 8, 128, 128)
    with torch.no_grad():
        out = model(x)
    print(f"  Input:  {tuple(x.shape)}")
    print(f"  Output: {tuple(out.shape)}  (expected (2, 4))")
    assert out.shape == (2, 4), f"Expected (2, 4), got {out.shape}"
    print("  PASSED")