有好看圖片的軟件網(wǎng)站模板,微信小程序營(yíng)銷推廣,如何運(yùn)營(yíng)一個(gè)品牌的推廣,百度seo排名優(yōu)化費(fèi)用浙大疏錦行 一、nn.Module 核心自帶方法 nn.Module 封裝了模型的核心邏輯#xff0c;以下是高頻使用的自帶方法#xff0c;按功能分類#xff1a; 1. 模型狀態(tài)控制#xff08;訓(xùn)練 / 評(píng)估模式#xff09; 方法作用model.train()切換為訓(xùn)練模式#xff1a;啟用 Dropou…浙大疏錦行一、nn.Module核心自帶方法nn.Module封裝了模型的核心邏輯以下是高頻使用的自帶方法按功能分類1. 模型狀態(tài)控制訓(xùn)練 / 評(píng)估模式方法作用model.train()切換為訓(xùn)練模式啟用 Dropout、BatchNorm 等層的訓(xùn)練行為默認(rèn)模式model.eval()切換為評(píng)估模式關(guān)閉 Dropout、固定 BatchNorm 均值 / 方差用于推理 / 驗(yàn)證model.training屬性返回布爾值True 訓(xùn)練模式False 評(píng)估模式示例import torch import torch.nn as nn class SimpleCNN(nn.Module): def __init__(self): super().__init__() self.conv nn.Conv2d(3, 16, 3) self.dropout nn.Dropout(0.5) # 訓(xùn)練時(shí)隨機(jī)失活評(píng)估時(shí)關(guān)閉 def forward(self, x): x self.conv(x) x self.dropout(x) return x model SimpleCNN() print(model.training) # True默認(rèn)訓(xùn)練模式 model.eval() print(model.training) # False評(píng)估模式dropout失效 model.train() print(model.training) # True切回訓(xùn)練模式2. 設(shè)備遷移CPU/GPU方法作用model.to(device)將模型所有參數(shù) / 緩沖區(qū)移到指定設(shè)備cuda/cpu/mps返回模型實(shí)例model.cuda()快捷方式移到默認(rèn) GPU等價(jià)于model.to(cuda)model.cpu()快捷方式移到 CPU等價(jià)于model.to(cpu)示例device torch.device(cuda if torch.cuda.is_available() else cpu) model model.to(device) # 模型移到GPU/CPU # 驗(yàn)證設(shè)備 print(next(model.parameters()).device) # 輸出cuda:0 或 cpu3. 參數(shù)管理查看 / 遍歷參數(shù)方法作用model.parameters()返回生成器包含所有可訓(xùn)練參數(shù)nn.Parameter類型model.named_parameters()返回生成器(參數(shù)名參數(shù)張量)便于定位參數(shù)model.named_parameters()返回生成器(參數(shù)名參數(shù)張量)便于定位參數(shù)model.state_dict()返回字典{參數(shù)名參數(shù)值}用于保存模型參數(shù)model.load_state_dict()加載參數(shù)字典用于恢復(fù)模型示例# 查看所有參數(shù)名稱和形狀 for name, param in model.named_parameters(): print(f參數(shù)名{name}形狀{param.shape}設(shè)備{param.device}) # 統(tǒng)計(jì)總參數(shù)量手動(dòng)實(shí)現(xiàn)無(wú)第三方庫(kù)時(shí)用 total_params sum(p.numel() for p in model.parameters()) trainable_params sum(p.numel() for p in model.parameters() if p.requires_grad) print(f總參數(shù){total_params}可訓(xùn)練參數(shù){trainable_params})4. 結(jié)構(gòu)遍歷查看模型層方法作用model.children()返回生成器僅包含直接子層如 Sequential 內(nèi)的第一層不遞歸model.named_children()返回生成器(層名子層)僅直接子層model.modules()返回生成器遞歸包含所有層包括嵌套層model.named_modules()返回生成器(層名層)遞歸所有層示例# 定義嵌套模型 class NestedModel(nn.Module): def __init__(self): super().__init__() self.block1 nn.Sequential( nn.Conv2d(3, 16, 3), nn.ReLU() ) self.block2 nn.Linear(16*30*30, 10) model NestedModel() # children()僅直接子層block1、block2 print( children() ) for name, layer in model.named_children(): print(name, layer) # modules()遞歸所有層包括Sequential內(nèi)的Conv2d、ReLU print( modules() ) for name, layer in model.named_modules(): print(name, layer)5. 前向傳播與梯度方法作用model.forward(x)手動(dòng)調(diào)用前向傳播不推薦建議直接model(x)調(diào)用__call__model(x)等價(jià)于model.__call__(x)自動(dòng)執(zhí)行 forward 鉤子hook邏輯model.zero_grad()清空所有參數(shù)的梯度訓(xùn)練時(shí)反向傳播前必須調(diào)用示例x torch.randn(1, 3, 32, 32).to(device) output model(x) # 推薦調(diào)用__call__等價(jià)于model.forward(x) 鉤子 model.zero_grad() # 清空梯度 output.sum().backward() # 反向傳播計(jì)算梯度二、torchsummary庫(kù)的summary方法torchsummary是早期輕量庫(kù)核心功能是快速打印模型層結(jié)構(gòu)、輸出形狀、總參數(shù)量?jī)H支持單輸入模型對(duì)嵌套模型 / 多輸入支持差維護(hù)較少。1. 安裝與基本用法pip install torchsummaryfrom torchsummary import summary # 定義模型輸入3通道32×32圖像 class SimpleCNN(nn.Module): def __init__(self): super().__init__() self.conv1 nn.Conv2d(3, 16, 3, padding1) self.pool nn.MaxPool2d(2, 2) self.conv2 nn.Conv2d(16, 32, 3, padding1) self.fc1 nn.Linear(32 * 8 * 8, 128) self.fc2 nn.Linear(128, 10) def forward(self, x): x self.pool(nn.functional.relu(self.conv1(x))) x self.pool(nn.functional.relu(self.conv2(x))) x x.view(-1, 32 * 8 * 8) x nn.functional.relu(self.fc1(x)) x self.fc2(x) return x # 設(shè)備配置 模型初始化 device torch.device(cuda if torch.cuda.is_available() else cpu) model SimpleCNN().to(device) # 調(diào)用summary參數(shù)模型輸入形狀通道,高,寬batch_size可選 summary(model, input_size(3, 32, 32), batch_size1)2. 輸出解讀---------------------------------------------------------------- Layer (type) Output Shape Param # Conv2d-1 [1, 16, 32, 32] 448 MaxPool2d-2 [1, 16, 16, 16] 0 Conv2d-3 [1, 32, 16, 16] 4,640 MaxPool2d-4 [1, 32, 8, 8] 0 Linear-5 [1, 128] 262,272 Linear-6 [1, 10] 1,290 Total params: 268,650 Trainable params: 268,650 Non-trainable params: 0 ---------------------------------------------------------------- Input size (MB): 0.01 Forward/backward pass size (MB): 0.29 Params size (MB): 1.02 Estimated Total Size (MB): 1.32 ----------------------------------------------------------------3. 優(yōu)缺點(diǎn)優(yōu)點(diǎn)缺點(diǎn)極簡(jiǎn)、無(wú)多余依賴僅支持單輸入模型輸出簡(jiǎn)潔、易理解對(duì)嵌套模型 / 多分支模型支持差快速查看參數(shù)量 / 形狀無(wú)批次維度、無(wú)內(nèi)存占用細(xì)分支持 GPU/CPU維護(hù)停滯僅兼容 PyTorch 舊版本三、torchinfo庫(kù)的summary方法推薦torchinfo是torchsummary的升級(jí)版原torchsummaryX解決了多輸入、嵌套模型、維度展示不清晰的問(wèn)題功能更全面是當(dāng)前 PyTorch 模型可視化的首選。1. 安裝與基本用法pip install torchinfofrom torchinfo import summary # 復(fù)用上面的SimpleCNN模型 device torch.device(cuda if torch.cuda.is_available() else cpu) model SimpleCNN().to(device) # 核心參數(shù)model, input_size, batch_dim, device, col_width等 summary( model, input_size(1, 3, 32, 32), # (batch_size, 通道, 高, 寬) batch_dim0, # 批次維度的位置默認(rèn)0 devicedevice, # 模型設(shè)備 col_width20, # 列寬 col_names[input_size, output_size, num_params, trainable], # 顯示列 row_settings[var_names] # 顯示層變量名 )2. 輸出解讀 Layer (type (var_name)) Input Shape Output Shape Param # Trainable SimpleCNN (SimpleCNN) [1, 3, 32, 32] [1, 10] -- -- ├─Conv2d (conv1) [1, 3, 32, 32] [1, 16, 32, 32] 448 True ├─MaxPool2d (pool) [1, 16, 32, 32] [1, 16, 16, 16] -- -- ├─Conv2d (conv2) [1, 16, 16, 16] [1, 32, 16, 16] 4,640 True ├─MaxPool2d (pool) [1, 32, 16, 16] [1, 32, 8, 8] -- -- ├─Linear (fc1) [1, 2048] [1, 128] 262,272 True ├─Linear (fc2) [1, 128] [1, 10] 1,290 True Total params: 268,650 Trainable params: 268,650 Non-trainable params: 0 Total mult-adds (M): 2.15 Input size (MB): 0.01 Forward/backward pass size (MB): 0.29 Params size (MB): 1.02 Estimated Total Size (MB): 1.32 四、推理的寫法評(píng)估模式def evaluate_classification(model, dataloader, device): 分類模型評(píng)估計(jì)算準(zhǔn)確率、F1-score宏平均、混淆矩陣 # 1. 切換到評(píng)估模式必須 model.eval() # 2. 初始化指標(biāo)容器 all_preds [] all_labels [] # 3. 關(guān)閉梯度計(jì)算加速省顯存 with torch.no_grad(): for batch_idx, (x, y) in enumerate(dataloader): # 數(shù)據(jù)移到設(shè)備 x x.to(device, dtypetorch.float32) y y.to(device, dtypetorch.long) # 4. 推理前向傳播 outputs model(x) # 輸出(batch_size, num_classes) preds torch.argmax(outputs, dim1) # 取概率最大的類別 # 5. 收集預(yù)測(cè)結(jié)果和真實(shí)標(biāo)簽轉(zhuǎn)回CPU便于計(jì)算指標(biāo) all_preds.extend(preds.cpu().numpy()) all_labels.extend(y.cpu().numpy()) # 可選打印進(jìn)度 if (batch_idx 1) % 10 0: print(fBatch [{batch_idx1}/{len(dataloader)}] 完成) # 6. 計(jì)算評(píng)估指標(biāo) accuracy accuracy_score(all_labels, all_preds) f1_macro f1_score(all_labels, all_preds, averagemacro) # 宏平均F1適合類別均衡 f1_weighted f1_score(all_labels, all_preds, averageweighted) # 加權(quán)F1適合類別不均衡 # 7. 打印結(jié)果 print(*50) print(f分類模型評(píng)估結(jié)果) print(f準(zhǔn)確率 (Accuracy): {accuracy:.4f}) print(f宏平均F1-score: {f1_macro:.4f}) print(f加權(quán)F1-score: {f1_weighted:.4f}) print(*50) return { accuracy: accuracy, f1_macro: f1_macro, f1_weighted: f1_weighted, preds: all_preds, labels: all_labels } # 執(zhí)行評(píng)估 eval_results evaluate_classification(model, test_loader, device)