import torch import torch.nn as nn import torchvision.models as models from torchvision.models import ResNet50_Weights # =================================================== # Feature Extraction の実装 # =================================================== # 1. 事前学習済みモデルを読み込む model = models.resnet50(weights=ResNet50_Weights.IMAGENET1K_V2) # 2. 全てのパラメータを固定（学習させない） # requires_grad = False にすると、その層は訓練時に更新されない for param in model.parameters(): param.requires_grad = False # 3. 最後の全結合層を新しいタスク用に置き換える # 元のfc層は1000クラス（ImageNet）用 # 新しいタスクのクラス数に合わせて置き換え num_classes = 2 # 例：犬 vs 猫 # model.fc.in_features で元の入力次元を取得 num_features = model.fc.in_features print(f”FC層の入力次元: {num_features}”) # 新しいFC層を作成 # 新しく作成した層は自動的に requires_grad=True model.fc = nn.Sequential( nn.Dropout(0.5), # 過学習防止のためのDropout nn.Linear(num_features, num_classes) ) # 4. 学習可能なパラメータ数を確認 trainable_params = sum(p.numel() for p in model.parameters() if p.requires_grad) total_params = sum(p.numel() for p in model.parameters()) print(f”学習可能なパラメータ: {trainable_params:,}”) print(f”全パラメータ: {total_params:,}”) print(f”学習する割合: {trainable_params / total_params * 100:.2f}%”)

import torch import torch.nn as nn import torch.optim as optim import torchvision.models as models from torchvision.models import ResNet50_Weights # =================================================== # Full Fine-tuning の実装 # =================================================== # 1. 事前学習済みモデルを読み込む model = models.resnet50(weights=ResNet50_Weights.IMAGENET1K_V2) # 2. 最後の層を置き換える（全層は訓練可能のまま） num_classes = 2 model.fc = nn.Sequential( nn.Dropout(0.5), nn.Linear(model.fc.in_features, num_classes) ) # 3. 層ごとに異なる学習率を設定 # これがFine-tuningの重要なポイント optimizer = optim.Adam([ # 初期層: 非常に小さい学習率（汎用特徴を保持） {‘params’: model.conv1.parameters(), ‘lr’: 1e-6}, {‘params’: model.bn1.parameters(), ‘lr’: 1e-6}, {‘params’: model.layer1.parameters(), ‘lr’: 1e-6}, # 中間層: 小さい学習率 {‘params’: model.layer2.parameters(), ‘lr’: 1e-5}, {‘params’: model.layer3.parameters(), ‘lr’: 1e-5}, # 後半層: やや大きい学習率 {‘params’: model.layer4.parameters(), ‘lr’: 1e-4}, # 新しい層: 大きい学習率 {‘params’: model.fc.parameters(), ‘lr’: 1e-3} ]) # 学習率の確認 print(“層ごとの学習率:”) for i, param_group in enumerate(optimizer.param_groups): print(f” Group {i}: lr = {param_group[‘lr’]}”)

import torch import torch.nn as nn import torch.optim as optim import torchvision.models as models from torchvision.models import ResNet50_Weights # =================================================== # 段階的Fine-tuning（Progressive Fine-tuning）の実装 # =================================================== def progressive_finetuning(model, train_loader, val_loader, device): “”” 3段階でFine-tuningを行う関数 Phase 1: FC層のみ訓練 Phase 2: layer4 + FC を訓練 Phase 3: 全層を訓練 “”” criterion = nn.CrossEntropyLoss() # =============================== # Phase 1: FC層のみ訓練 # =============================== print(“=” * 50) print(“Phase 1: FC層のみ訓練”) print(“=” * 50) # 全層を固定 for param in model.parameters(): param.requires_grad = False # FC層のみ学習可能に for param in model.fc.parameters(): param.requires_grad = True # オプティマイザ（FC層のみ） optimizer = optim.Adam(model.fc.parameters(), lr=1e-3) # 5エポック訓練 for epoch in range(5): train_loss = train_one_epoch(model, train_loader, criterion, optimizer, device) val_acc = evaluate(model, val_loader, device) print(f”Epoch {epoch+1}/5 – Loss: {train_loss:.4f}, Val Acc: {val_acc:.2f}%”) # =============================== # Phase 2: layer4 + FC を訓練 # =============================== print(“\n” + “=” * 50) print(“Phase 2: layer4 + FC を訓練”) print(“=” * 50) # layer4を学習可能に for param in model.layer4.parameters(): param.requires_grad = True # 異なる学習率 optimizer = optim.Adam([ {‘params’: model.layer4.parameters(), ‘lr’: 1e-5}, {‘params’: model.fc.parameters(), ‘lr’: 1e-4} ]) # 10エポック訓練 for epoch in range(10): train_loss = train_one_epoch(model, train_loader, criterion, optimizer, device) val_acc = evaluate(model, val_loader, device) print(f”Epoch {epoch+1}/10 – Loss: {train_loss:.4f}, Val Acc: {val_acc:.2f}%”) # =============================== # Phase 3: 全層を訓練 # =============================== print(“\n” + “=” * 50) print(“Phase 3: 全層を訓練”) print(“=” * 50) # 全層を学習可能に for param in model.parameters(): param.requires_grad = True # 層ごとに異なる学習率 optimizer = optim.Adam([ {‘params’: model.conv1.parameters(), ‘lr’: 1e-6}, {‘params’: model.bn1.parameters(), ‘lr’: 1e-6}, {‘params’: model.layer1.parameters(), ‘lr’: 1e-6}, {‘params’: model.layer2.parameters(), ‘lr’: 1e-6}, {‘params’: model.layer3.parameters(), ‘lr’: 1e-5}, {‘params’: model.layer4.parameters(), ‘lr’: 1e-5}, {‘params’: model.fc.parameters(), ‘lr’: 1e-4} ]) # 15エポック訓練 for epoch in range(15): train_loss = train_one_epoch(model, train_loader, criterion, optimizer, device) val_acc = evaluate(model, val_loader, device) print(f”Epoch {epoch+1}/15 – Loss: {train_loss:.4f}, Val Acc: {val_acc:.2f}%”) return model def train_one_epoch(model, loader, criterion, optimizer, device): “””1エポックの訓練””” model.train() total_loss = 0 for images, labels in loader: images, labels = images.to(device), labels.to(device) optimizer.zero_grad() outputs = model(images) loss = criterion(outputs, labels) loss.backward() optimizer.step() total_loss += loss.item() return total_loss / len(loader) def evaluate(model, loader, device): “””評価””” model.eval() correct = 0 total = 0 with torch.no_grad(): for images, labels in loader: images, labels = images.to(device), labels.to(device) outputs = model(images) _, predicted = outputs.max(1) total += labels.size(0) correct += predicted.eq(labels).sum().item() return 100. * correct / total # 使用例 # model = models.resnet50(weights=ResNet50_Weights.IMAGENET1K_V2) # model.fc = nn.Linear(model.fc.in_features, num_classes) # model = model.to(device) # model = progressive_finetuning(model, train_loader, val_loader, device)

# timmのインストール # !pip install timm import timm # 利用可能なモデル一覧を取得 all_models = timm.list_models() print(f”全モデル数: {len(all_models)}”) # 事前学習済みモデルのみ pretrained_models = timm.list_models(pretrained=True) print(f”事前学習済みモデル数: {len(pretrained_models)}”) # 特定のキーワードで検索 efficientnet_models = timm.list_models(‘*efficientnet*’) print(f”\nEfficientNet系モデル: {len(efficientnet_models)}”) for m in efficientnet_models[:5]: print(f” – {m}”)

import timm import torch # =================================================== # timmでモデルを作成する方法 # =================================================== # 方法1: 事前学習済み + クラス数指定（最も一般的） model = timm.create_model( ‘efficientnet_b0’, # モデル名 pretrained=True, # 事前学習済み重みを使用 num_classes=10 # 出力クラス数を変更 ) # モデル情報を確認 print(f”モデル: {model.default_cfg[‘architecture’]}”) print(f”入力サイズ: {model.default_cfg[‘input_size’]}”) print(f”パラメータ数: {sum(p.numel() for p in model.parameters()):,}”) # 方法2: 分類層なしで作成（特徴抽出器として使用） feature_extractor = timm.create_model( ‘resnet50’, pretrained=True, num_classes=0 # 分類層なし ) # 特徴次元を確認 dummy_input = torch.randn(1, 3, 224, 224) with torch.no_grad(): features = feature_extractor(dummy_input) print(f”\n特徴次元: {features.shape}”)

import timm import torch import torch.nn as nn import torch.optim as optim # =================================================== # timmを使った転移学習の完全なコード # =================================================== # 1. モデル作成 model = timm.create_model( ‘efficientnet_b0’, pretrained=True, num_classes=2 # 2クラス分類 ) # 2. Feature Extractionの設定 # timmのモデルは classifier 属性で分類層にアクセス # （モデルによって属性名が異なる場合がある） # 全層を固定 for param in model.parameters(): param.requires_grad = False # 分類層のみ学習可能に # EfficientNetの場合は classifier 属性 for param in model.classifier.parameters(): param.requires_grad = True # 3. オプティマイザ optimizer = optim.Adam(model.classifier.parameters(), lr=1e-3) # 4. データ変換（timmの推奨設定を使用） from timm.data import resolve_data_config from timm.data.transforms_factory import create_transform # モデルに最適な変換を自動生成 config = resolve_data_config({}, model=model) transform = create_transform(**config) print(“推奨データ変換:”) print(transform) # 5. 学習可能なパラメータを確認 trainable = sum(p.numel() for p in model.parameters() if p.requires_grad) total = sum(p.numel() for p in model.parameters()) print(f”\n学習可能パラメータ: {trainable:,} / {total:,} ({100*trainable/total:.2f}%)”)

import timm import torch import time # =================================================== # 複数モデルの比較 # =================================================== models_to_compare = [ (‘efficientnet_b0’, 224), (‘resnet50’, 224), (‘mobilenetv2_100’, 224), (‘convnext_tiny’, 224), (‘swin_tiny_patch4_window7_224’, 224) ] print(“モデル比較:”) print(“-” * 70) print(f”{‘モデル’:<35} {'パラメータ':<15} {'推論時間':<15}") print("-" * 70) for model_name, size in models_to_compare: # モデル作成 model = timm.create_model(model_name, pretrained=False, num_classes=10) model.eval() # パラメータ数 params = sum(p.numel() for p in model.parameters()) / 1e6 # 推論速度測定 x = torch.randn(1, 3, size, size) # ウォームアップ with torch.no_grad(): _ = model(x) # 計測 start = time.time() with torch.no_grad(): for _ in range(50): _ = model(x) elapsed = (time.time() - start) / 50 * 1000 # ms print(f"{model_name:<35} {params:>10.1f}M {elapsed:>10.2f}ms”) print(“-” * 70)