import torch import torch.nn as nn class DepthwiseSeparableConv(nn.Module): “””Depthwise Separable Convolution””” def __init__(self, in_channels, out_channels, stride=1): super(DepthwiseSeparableConv, self).__init__() # Step 1: Depthwise Convolution # groups=in_channels で各チャンネルを独立に処理 self.depthwise = nn.Conv2d( in_channels, in_channels, kernel_size=3, stride=stride, padding=1, groups=in_channels, # これがポイント！ bias=False ) self.bn1 = nn.BatchNorm2d(in_channels) self.relu1 = nn.ReLU(inplace=True) # Step 2: Pointwise Convolution（1×1畳み込み） self.pointwise = nn.Conv2d( in_channels, out_channels, kernel_size=1, stride=1, padding=0, bias=False ) self.bn2 = nn.BatchNorm2d(out_channels) self.relu2 = nn.ReLU(inplace=True) def forward(self, x): # Depthwise out = self.depthwise(x) out = self.bn1(out) out = self.relu1(out) # Pointwise out = self.pointwise(out) out = self.bn2(out) out = self.relu2(out) return out # テスト conv = DepthwiseSeparableConv(128, 256) x = torch.randn(1, 128, 56, 56) output = conv(x) print(f”入力: {x.shape}”) print(f”出力: {output.shape}”) # パラメータ数の比較 normal_conv = nn.Conv2d(128, 256, kernel_size=3, padding=1) print(f”\n通常の畳み込みのパラメータ数: {sum(p.numel() for p in normal_conv.parameters()):,}”) print(f”Depthwise Separableのパラメータ数: {sum(p.numel() for p in conv.parameters()):,}”)

import torch import torch.nn as nn class InvertedResidualBlock(nn.Module): “””MobileNet v2のInverted Residual Block””” def __init__(self, in_channels, out_channels, stride=1, expansion=6): super(InvertedResidualBlock, self).__init__() self.stride = stride self.use_residual = (stride == 1) and (in_channels == out_channels) hidden_dim = in_channels * expansion layers = [] # 拡張（expansion）: 1×1 Conv if expansion != 1: layers.extend([ nn.Conv2d(in_channels, hidden_dim, kernel_size=1, bias=False), nn.BatchNorm2d(hidden_dim), nn.ReLU6(inplace=True) # ReLU6を使用 ]) # Depthwise Conv: 3×3 layers.extend([ nn.Conv2d(hidden_dim, hidden_dim, kernel_size=3, stride=stride, padding=1, groups=hidden_dim, bias=False), nn.BatchNorm2d(hidden_dim), nn.ReLU6(inplace=True) ]) # 縮小（projection）: 1×1 Conv（Linear = ReLUなし） layers.extend([ nn.Conv2d(hidden_dim, out_channels, kernel_size=1, bias=False), nn.BatchNorm2d(out_channels) # ここにReLUはない（Linear Bottleneck） ]) self.conv = nn.Sequential(*layers) def forward(self, x): if self.use_residual: return x + self.conv(x) # Skip Connection else: return self.conv(x) # テスト block = InvertedResidualBlock(24, 24, stride=1, expansion=6) x = torch.randn(1, 24, 56, 56) output = block(x) print(f”入力: {x.shape}”) print(f”出力: {output.shape}”) print(f”Skip Connection使用: {block.use_residual}”)

# PyTorchで事前学習済みMobileNet v2を読み込む import torch import torch.nn as nn import torchvision.models as models from torchvision.models import MobileNet_V2_Weights # MobileNet v2を読み込み model = models.mobilenet_v2(weights=MobileNet_V2_Weights.IMAGENET1K_V2) # モデルの構造を確認 print(f”最後の分類器: {model.classifier}”) # パラメータ数を確認 total_params = sum(p.numel() for p in model.parameters()) print(f”総パラメータ数: {total_params:,}”) # 転移学習用に分類器を置き換え num_classes = 2 model.classifier[1] = nn.Linear(model.classifier[1].in_features, num_classes) print(f”新しい分類器: {model.classifier}”) # テスト model.eval() x = torch.randn(1, 3, 224, 224) with torch.no_grad(): output = model(x) print(f”出力: {output.shape}”)

# PyTorchで事前学習済みEfficientNetを読み込む import torch import torch.nn as nn import torchvision.models as models from torchvision.models import EfficientNet_B0_Weights, EfficientNet_B3_Weights # EfficientNet-B0を読み込み model_b0 = models.efficientnet_b0(weights=EfficientNet_B0_Weights.IMAGENET1K_V1) # EfficientNet-B3を読み込み model_b3 = models.efficientnet_b3(weights=EfficientNet_B3_Weights.IMAGENET1K_V1) # パラメータ数を確認 params_b0 = sum(p.numel() for p in model_b0.parameters()) params_b3 = sum(p.numel() for p in model_b3.parameters()) print(f”EfficientNet-B0 パラメータ数: {params_b0:,}”) print(f”EfficientNet-B3 パラメータ数: {params_b3:,}”) # 転移学習用に分類器を置き換え（B0の場合） num_classes = 2 model_b0.classifier[1] = nn.Linear(model_b0.classifier[1].in_features, num_classes) # テスト（B0は224×224、B3は300×300） model_b0.eval() x_b0 = torch.randn(1, 3, 224, 224) with torch.no_grad(): output = model_b0(x_b0) print(f”EfficientNet-B0 出力: {output.shape}”)

# PyTorchで動的量子化を実行 import torch import torchvision.models as models import os # MobileNet v2を読み込み model = models.mobilenet_v2(weights=’IMAGENET1K_V2′) model.eval() # 動的量子化（Dynamic Quantization） # Linear層とConv層を量子化 quantized_model = torch.quantization.quantize_dynamic( model, {torch.nn.Linear}, # 量子化する層の種類 dtype=torch.qint8 # 8bit整数に量子化 ) # モデルサイズの比較 def get_model_size(model, path): torch.save(model.state_dict(), path) size = os.path.getsize(path) / (1024 * 1024) # MB os.remove(path) return size original_size = get_model_size(model, ‘temp_original.pt’) quantized_size = get_model_size(quantized_model, ‘temp_quantized.pt’) print(f”元のモデルサイズ: {original_size:.2f} MB”) print(f”量子化後のサイズ: {quantized_size:.2f} MB”) print(f”削減率: {original_size / quantized_size:.2f}倍”)

# PyTorchでプルーニングを実行 import torch import torch.nn.utils.prune as prune import torchvision.models as models # MobileNet v2を読み込み model = models.mobilenet_v2(weights=’IMAGENET1K_V2′) # 最初の畳み込み層を取得 first_conv = model.features[0][0] print(“プルーニング前:”) print(f” ゼロの数: {torch.sum(first_conv.weight == 0).item()}”) print(f” 総パラメータ数: {first_conv.weight.numel()}”) # L1ノルムに基づいて30%の重みをプルーニング prune.l1_unstructured(first_conv, name=’weight’, amount=0.3) print(“\nプルーニング後（30%削除）:”) print(f” ゼロの数: {torch.sum(first_conv.weight == 0).item()}”) print(f” スパース性: {torch.sum(first_conv.weight == 0).item() / first_conv.weight.numel() * 100:.1f}%”)

import torch import torch.nn as nn import torch.nn.functional as F def distillation_loss(student_logits, teacher_logits, labels, temperature=3.0, alpha=0.7): “”” 知識蒸留の損失関数 Args: student_logits: Studentモデルの出力 teacher_logits: Teacherモデルの出力 labels: 正解ラベル temperature: 温度パラメータ（高いほど滑らか） alpha: Hard Lossの重み “”” # Hard Loss: 正解ラベルとの交差エントロピー hard_loss = F.cross_entropy(student_logits, labels) # Soft Loss: Teacher出力とのKLダイバージェンス # 温度で割ってソフトにする soft_student = F.log_softmax(student_logits / temperature, dim=1) soft_teacher = F.softmax(teacher_logits / temperature, dim=1) # KLダイバージェンス（温度の2乗でスケール） soft_loss = F.kl_div( soft_student, soft_teacher, reduction=’batchmean’ ) * (temperature ** 2) # 合計損失 total_loss = alpha * hard_loss + (1 – alpha) * soft_loss return total_loss, hard_loss, soft_loss # 使用例 batch_size = 4 num_classes = 10 # ダミーデータ student_logits = torch.randn(batch_size, num_classes) teacher_logits = torch.randn(batch_size, num_classes) labels = torch.randint(0, num_classes, (batch_size,)) total_loss, hard_loss, soft_loss = distillation_loss( student_logits, teacher_logits, labels, temperature=3.0, alpha=0.7 ) print(f”Total Loss: {total_loss.item():.4f}”) print(f”Hard Loss: {hard_loss.item():.4f}”) print(f”Soft Loss: {soft_loss.item():.4f}”)