import torch import torch.nn as nn class DenseLayer(nn.Module): “””DenseNet内の1層（Bottleneck構造付き）””” def __init__(self, in_channels, growth_rate): super(DenseLayer, self).__init__() # Bottleneck: BN → ReLU → Conv1×1 self.bn1 = nn.BatchNorm2d(in_channels) self.relu1 = nn.ReLU(inplace=True) self.conv1 = nn.Conv2d(in_channels, 4 * growth_rate, kernel_size=1, bias=False) # 3×3畳み込み: BN → ReLU → Conv3×3 self.bn2 = nn.BatchNorm2d(4 * growth_rate) self.relu2 = nn.ReLU(inplace=True) self.conv2 = nn.Conv2d(4 * growth_rate, growth_rate, kernel_size=3, padding=1, bias=False) def forward(self, x): out = self.conv1(self.relu1(self.bn1(x))) out = self.conv2(self.relu2(self.bn2(out))) return torch.cat([x, out], dim=1) # 連結！ # テスト layer = DenseLayer(in_channels=64, growth_rate=32) x = torch.randn(1, 64, 56, 56) output = layer(x) print(f”入力: {x.shape}”) print(f”出力: {output.shape}”)

class DenseBlock(nn.Module): “””複数のDense Layerからなるブロック””” def __init__(self, num_layers, in_channels, growth_rate): “”” Args: num_layers: ブロック内の層数 in_channels: 入力チャンネル数 growth_rate: 各層が出力するチャンネル数 “”” super(DenseBlock, self).__init__() layers = [] for i in range(num_layers): # 各層の入力チャンネル数 = 初期 + これまでの層の出力の合計 layer_in_channels = in_channels + i * growth_rate layers.append(DenseLayer(layer_in_channels, growth_rate)) self.layers = nn.ModuleList(layers) def forward(self, x): for layer in self.layers: x = layer(x) # 各層で連結が行われる return x # テスト：6層のDense Block block = DenseBlock(num_layers=6, in_channels=64, growth_rate=32) x = torch.randn(1, 64, 56, 56) output = block(x) print(f”入力: {x.shape}”) print(f”出力: {output.shape}”)

class TransitionLayer(nn.Module): “””Dense Block間の圧縮層””” def __init__(self, in_channels, out_channels): “”” Args: in_channels: 入力チャンネル数 out_channels: 出力チャンネル数（通常は半分） “”” super(TransitionLayer, self).__init__() self.bn = nn.BatchNorm2d(in_channels) self.relu = nn.ReLU(inplace=True) # 1×1畳み込みでチャンネル数を削減 self.conv = nn.Conv2d(in_channels, out_channels, kernel_size=1, bias=False) # Average Poolingで空間サイズを半分に self.pool = nn.AvgPool2d(kernel_size=2, stride=2) def forward(self, x): out = self.conv(self.relu(self.bn(x))) out = self.pool(out) return out # テスト transition = TransitionLayer(in_channels=256, out_channels=128) x = torch.randn(1, 256, 56, 56) output = transition(x) print(f”入力: {x.shape}”) print(f”出力: {output.shape}”)

# PyTorchで事前学習済みDenseNetを読み込む import torch import torchvision.models as models from torchvision.models import DenseNet121_Weights # DenseNet-121を読み込み（ImageNetで事前学習済み） model = models.densenet121(weights=DenseNet121_Weights.IMAGENET1K_V1) # モデル構造の確認 print(model) # パラメータ数を確認 total_params = sum(p.numel() for p in model.parameters()) print(f”\n総パラメータ数: {total_params:,}”) # 転移学習用にclassifierを置き換え num_classes = 2 # 例：猫 vs 犬 num_features = model.classifier.in_features # 1024 model.classifier = nn.Linear(num_features, num_classes) print(f”新しい分類器: {model.classifier}”)

import torch import torch.nn as nn class InceptionModule(nn.Module): “””Inception Module with Dimension Reduction””” def __init__(self, in_channels, ch1x1, ch3x3_reduce, ch3x3, ch5x5_reduce, ch5x5, pool_proj): “”” Args: in_channels: 入力チャンネル数 ch1x1: 1×1ブランチの出力チャンネル数 ch3x3_reduce: 3×3ブランチの1×1削減後のチャンネル数 ch3x3: 3×3ブランチの出力チャンネル数 ch5x5_reduce: 5×5ブランチの1×1削減後のチャンネル数 ch5x5: 5×5ブランチの出力チャンネル数 pool_proj: プーリングブランチの1×1後のチャンネル数 “”” super(InceptionModule, self).__init__() # ブランチ1: 1×1畳み込み self.branch1 = nn.Sequential( nn.Conv2d(in_channels, ch1x1, kernel_size=1), nn.BatchNorm2d(ch1x1), nn.ReLU(inplace=True) ) # ブランチ2: 1×1 → 3×3 self.branch2 = nn.Sequential( nn.Conv2d(in_channels, ch3x3_reduce, kernel_size=1), nn.BatchNorm2d(ch3x3_reduce), nn.ReLU(inplace=True), nn.Conv2d(ch3x3_reduce, ch3x3, kernel_size=3, padding=1), nn.BatchNorm2d(ch3x3), nn.ReLU(inplace=True) ) # ブランチ3: 1×1 → 5×5 self.branch3 = nn.Sequential( nn.Conv2d(in_channels, ch5x5_reduce, kernel_size=1), nn.BatchNorm2d(ch5x5_reduce), nn.ReLU(inplace=True), nn.Conv2d(ch5x5_reduce, ch5x5, kernel_size=5, padding=2), nn.BatchNorm2d(ch5x5), nn.ReLU(inplace=True) ) # ブランチ4: MaxPool → 1×1 self.branch4 = nn.Sequential( nn.MaxPool2d(kernel_size=3, stride=1, padding=1), nn.Conv2d(in_channels, pool_proj, kernel_size=1), nn.BatchNorm2d(pool_proj), nn.ReLU(inplace=True) ) def forward(self, x): # 各ブランチを並列に計算 branch1_out = self.branch1(x) branch2_out = self.branch2(x) branch3_out = self.branch3(x) branch4_out = self.branch4(x) # チャンネル方向に連結 return torch.cat([branch1_out, branch2_out, branch3_out, branch4_out], dim=1) # テスト inception = InceptionModule( in_channels=192, ch1x1=64, ch3x3_reduce=96, ch3x3=128, ch5x5_reduce=16, ch5x5=32, pool_proj=32 ) x = torch.randn(1, 192, 28, 28) output = inception(x) print(f”入力: {x.shape}”) print(f”出力: {output.shape}”) print(f”出力チャンネル数: 64+128+32+32 = {64+128+32+32}”)

# PyTorchで事前学習済みInception v3を読み込む import torch import torchvision.models as models from torchvision.models import Inception_V3_Weights # Inception v3を読み込み # 注意：Inception v3は入力サイズが299×299 model = models.inception_v3(weights=Inception_V3_Weights.IMAGENET1K_V1) # 訓練時には補助分類器も使われる（aux_logits=True） # 推論時にはメイン分類器のみ使用 print(f”補助分類器の有無: {model.aux_logits}”) # 転移学習用に分類器を置き換え num_classes = 2 model.fc = nn.Linear(model.fc.in_features, num_classes) # 補助分類器も置き換え（訓練時に使用） if model.aux_logits: model.AuxLogits.fc = nn.Linear(model.AuxLogits.fc.in_features, num_classes) print(f”メイン分類器: {model.fc}”) # 推論時の使用例 model.eval() # 評価モードに（補助分類器は使われない） x = torch.randn(1, 3, 299, 299) # 299×299が必要！ with torch.no_grad(): output = model(x) print(f”出力: {output.shape}”)

# PyTorchで事前学習済みResNeXtを読み込む import torch import torchvision.models as models from torchvision.models import ResNeXt50_32X4D_Weights # ResNeXt-50 (32×4d) を読み込み # 32 = Cardinality（並列経路の数） # 4d = 各経路のチャンネル幅 model = models.resnext50_32x4d(weights=ResNeXt50_32X4D_Weights.IMAGENET1K_V2) # パラメータ数を確認 total_params = sum(p.numel() for p in model.parameters()) print(f”ResNeXt-50 (32×4d) パラメータ数: {total_params:,}”) # 転移学習用に分類器を置き換え num_classes = 2 model.fc = nn.Linear(model.fc.in_features, num_classes) # テスト x = torch.randn(1, 3, 224, 224) model.eval() with torch.no_grad(): output = model(x) print(f”出力: {output.shape}”)