import torch import torch.nn as nn class PlainBlock(nn.Module): def __init__(self, channels, bn=True): super().__init__() self.conv1 = nn.Conv2d(channels, channels, 3, padding=1, bias=not bn) self.bn1 = nn.BatchNorm2d(channels) if bn else nn.Identity() self.conv2 = nn.Conv2d(channels, channels, 3, padding=1, bias=not bn) self.bn2 = nn.BatchNorm2d(channels) if bn else nn.Identity() def forward(self, x): out = torch.relu(self.bn1(self.conv1(x))) out = self.bn2(self.conv2(out)) return torch.relu(out) class ResBlock(nn.Module): def __init__(self, channels, bn=True): super().__init__() self.conv1 = nn.Conv2d(channels, channels, 3, padding=1, bias=not bn) self.bn1 = nn.BatchNorm2d(channels) if bn else nn.Identity() self.conv2 = nn.Conv2d(channels, channels, 3, padding=1, bias=not bn) self.bn2 = nn.BatchNorm2d(channels) if bn else nn.Identity() def forward(self, x): out = torch.relu(self.bn1(self.conv1(x))) out = self.bn2(self.conv2(out)) return torch.relu(out + x) class CNN(nn.Module): def __init__(self, depth, channels=64, num_classes=10, use_residual=False, bn=True): super().__init__() block = ResBlock if use_residual else PlainBlock self.stem = nn.Sequential( nn.Conv2d(1, channels, 3, stride=2, padding=1, bias=not bn), nn.BatchNorm2d(channels) if bn else nn.Identity(), nn.ReLU(inplace=True), ) self.blocks = nn.Sequential(*[block(channels, bn) for _ in range(depth)]) self.fc = nn.Linear(channels, num_classes) def forward(self, x): x = self.stem(x) x = self.blocks(x) x = torch.nn.functional.adaptive_avg_pool2d(x, 1).flatten(1) return self.fc(x) def count_params(model): return sum(p.numel() for p in model.parameters())