#train.pyimport torchimport torch.nn as nnfrom torchvision import transforms, datasetsimport torchvisionimport jsonimport matplotlib.pyplot as pltimport osimport torch.optim as optimfrom model import GoogLeNetdevice = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")print(device)data_transform = {    "train": transforms.Compose([transforms.RandomResizedCrop(224),                                 transforms.RandomHorizontalFlip(),                                 transforms.ToTensor(),                                 transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))]),    "val": transforms.Compose([transforms.Resize((224, 224)),                               transforms.ToTensor(),                               transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))])}#data_root = os.path.abspath(os.path.join(os.getcwd(), "../.."))  # get data root pathdata_root = os.getcwd()image_path = data_root + "/flower_data/"  # flower data set pathtrain_dataset = datasets.ImageFolder(root=image_path + "train",                                     transform=data_transform["train"])train_num = len(train_dataset)# {'daisy':0, 'dandelion':1, 'roses':2, 'sunflower':3, 'tulips':4}flower_list = train_dataset.class_to_idxcla_dict = dict((val, key) for key, val in flower_list.items())# write dict into json filejson_str = json.dumps(cla_dict, indent=4)with open('class_indices.json', 'w') as json_file:    json_file.write(json_str)batch_size = 32train_loader = torch.utils.data.DataLoader(train_dataset,                                           batch_size=batch_size, shuffle=True,                                           num_workers=0)validate_dataset = datasets.ImageFolder(root=image_path + "val",                                        transform=data_transform["val"])val_num = len(validate_dataset)validate_loader = torch.utils.data.DataLoader(validate_dataset,                                              batch_size=batch_size, shuffle=False,                                              num_workers=0)# test_data_iter = iter(validate_loader)# test_image, test_label = test_data_iter.next()# net = torchvision.models.googlenet(num_classes=5)# model_dict = net.state_dict()# pretrain_model = torch.load("googlenet.pth")# del_list = ["aux1.fc2.weight", "aux1.fc2.bias",#             "aux2.fc2.weight", "aux2.fc2.bias",#             "fc.weight", "fc.bias"]# pretrain_dict = {k: v for k, v in pretrain_model.items() if k not in del_list}# model_dict.update(pretrain_dict)# net.load_state_dict(model_dict)net = GoogLeNet(num_classes=5, aux_logits=True, init_weights=True)net.to(device)loss_function = nn.CrossEntropyLoss()optimizer = optim.Adam(net.parameters(), lr=0.0003)best_acc = 0.0save_path = './googleNet.pth'for epoch in range(30):    # train    net.train()    running_loss = 0.0    for step, data in enumerate(train_loader, start=0):        images, labels = data        optimizer.zero_grad()        logits, aux_logits2, aux_logits1 = net(images.to(device))        loss0 = loss_function(logits, labels.to(device))        loss1 = loss_function(aux_logits1, labels.to(device))        loss2 = loss_function(aux_logits2, labels.to(device))        loss = loss0 + loss1 * 0.3 + loss2 * 0.3        loss.backward()        optimizer.step()        # print statistics        running_loss += loss.item()        # print train process        rate = (step + 1) / len(train_loader)        a = "*" * int(rate * 50)        b = "." * int((1 - rate) * 50)        print("rtrain loss: {:^3.0f}%[{}->{}]{:.3f}".format(int(rate * 100), a, b, loss), end="")    print()    # validate    net.eval()    acc = 0.0  # accumulate accurate number / epoch    with torch.no_grad():        for val_data in validate_loader:            val_images, val_labels = val_data            outputs = net(val_images.to(device))  # eval model only have last output layer            predict_y = torch.max(outputs, dim=1)[1]            acc += (predict_y == val_labels.to(device)).sum().item()        val_accurate = acc / val_num        if val_accurate > best_acc:            best_acc = val_accurate            torch.save(net.state_dict(), save_path)        print('[epoch %d] train_loss: %.3f  test_accuracy: %.3f' %              (epoch + 1, running_loss / step, val_accurate))print('Finished Training')