在练习用VGG16做分类，我用这篇的程式码下去改的https://reurl.cc/RvvG9x VGG16的部分和他这篇不一样不是用预设而是用写的 问题在於，进行训练的时候Testing acc一直都是0.500，而acc都在0.5附近上下跳动 也就是说好像根本没有学习到 进行训练的程式码如下: (基本与那篇相同，这里只是将原文中的C改成model，并且把label的type改成tensor) if __name__ == '__main__': for epoch in range(epochs): start_time = time.time() iter = 0 correct_train, total_train = 0, 0 correct_test, total_test = 0, 0 train_loss_C = 0.0 model.train() # 设定 train 或 eval print('epoch: ' + str(epoch + 1) + ' / ' + str(epochs)) # --------------------------- # Training Stage # --------------------------- for i, (x, label) in enumerate(train_dataloader) : label = label.type(torch.LongTensor) x, label = x.to(device), label.to(device) optimizer_C.zero_grad() # 清空梯度 train_output = model(x) train_loss = criterion(train_output, label) # 计算 loss train_loss.backward() # 将 loss 反向传播 optimizer_C.step() # 更新权重 # 计算训练资料的准确度 (correct_train / total_train) _, predicted = torch.max(train_output.data, 1) total_train += label.size(0) correct_train += (predicted == label).sum() train_loss_C += train_loss.item() iter += 1 print('Training epoch: %d / loss_C: %.3f | acc: %.3f' % \ (epoch + 1, train_loss_C / iter, correct_train / total_train)) # -------------------------- # Testing Stage # -------------------------- model.eval() # 设定 train 或 eval for i, (x, label) in enumerate(test_dataloader) : with torch.no_grad(): label = label.type(torch.LongTensor) x, label = x.to(device), label.to(device) test_output = model(x) test_loss = criterion(test_output, label) # 计算 loss # 计算测试资料的准确度 (correct_test / total_test) _, predicted = torch.max(test_output.data, 1) total_test += label.size(0) correct_test += (predicted == label).sum() print('Testing acc: %.3f' % (correct_test / total_test)) train_acc.append(100 * (correct_train / total_train).cpu()) # training accuracy test_acc.append(100 * (correct_test / total_test).cpu()) # testing accuracy loss_epoch_C.append((train_loss_C / iter)) # loss end_time = time.time() print('Cost %.3f(secs)' % (end_time - start_time)) 指定epochs只跑3次，得到以下结果: epoch: 1 / 3 Training epoch: 1 / loss_C: 0.693 | acc: 0.514 Testing acc: 0.500 Cost 21.822(secs) epoch: 2 / 3 Training epoch: 2 / loss_C: 0.693 | acc: 0.475 Testing acc: 0.500 Cost 23.002(secs) epoch: 3 / 3 Training epoch: 3 / loss_C: 0.693 | acc: 0.476 Testing acc: 0.500 Cost 22.669(secs) --

※ 发信站: 批踢踢实业坊(ptt.cc), 来自: 42.77.144.160 (台湾) ※ 文章网址: https://webptt.com/cn.aspx?n=bbs/DataScience/M.1678336929.A.EBC.html