pytorchを活用したResNetの転移学習の上でのエラー

python3
1a_transform = transforms.Compose([
2    transforms.RandomResizedCrop(size=224),
3    transforms.RandomHorizontalFlip(),
4    transforms.ToTensor(),
5    transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])])
6
7e_transform = transforms.Compose([
8    transforms.Resize(size=224),
9    transforms.CenterCrop(size=224),
10    transforms.ToTensor(),
11    transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])])
12        
13        
14train_dataset = ImageFolder(f"{DATASETDIR}/Train", transform=augment_transform)
15val_dataset = ImageFolder(f"{DATASETDIR}/Validation", transform=evaluate_transform)
16test_dataset = ImageFolder(f"{DATASETDIR}/Test", transform=evaluate_transform)
17
18model = resnet34(pretrained=True)
19model.fc = nn.Sequential(
20    nn.Dropout(0.5),
21    nn.Linear(in_features=512, out_features=6, bias=True)
22)
23model = model.to("cuda")
24
25loss_func = torch.nn.CrossEntropyLoss()
26optimizer = torch.optim.SGD(model.parameters(), lr=1e-3, momentum=0.9)
27scheduler = torch.optim.lr_scheduler.StepLR(optimizer, step_size=3, gamma=0.9)
28
29def training(model, dataset, optimizer, loss_func):
30    model.train()
31  
32    losses = []
33    for img, target in DataLoader(dataset, shuffle=True, batch_size=32, num_workers=4):
34        img = img.to("cuda")
35        target = target.to("cuda")
36        pred = model(img)
37        loss = loss_func(pred, target)
38        losses.append(loss.item())
39
40        optimizer.zero_grad()
41        loss.backward()
42        optimizer.step()
43    
44    return np.mean(losses)
45
46def evaluation(model, dataset, loss_func):
47    model.eval()
48  
49    losses = []
50    correct = 0
51    with torch.no_grad():
52        for img, target in DataLoader(dataset, shuffle=False, batch_size=64, num_workers=4):
53            img = img.to("cuda")
54            target = target.to("cuda")
55            pred = model(img)
56            loss = loss_func(pred, target)
57            losses.append(loss.item())
58
59            predict = pred.argmax(dim=1)
60            correct += (predict == target).sum().item()
61    return np.mean(losses), correct / len(dataset)
62best_loss = 1e+10
63best_state = None
64earlystop_counter = 0
65for epoch in range(100):
66    train_loss = training(model, train_dataset, optimizer, loss_func)
67    val_loss, val_acc = evaluation(model, val_dataset, loss_func)
68  
69    if val_loss < best_loss:
70        earlystop_counter = 0
71        best_loss = val_loss
72        best_state = model.cpu().state_dict()
73        model.cuda()
74    else:
75        earlystop_counter += 1
76    
77    if earlystop_counter > 5:
78        break
79  
80    scheduler.step()
81    print(f"epoch {str(epoch).zfill(2)}\ttrain_loss: {train_loss}\tval_loss: {val_loss}\tval_acc: {val_acc}")
82
83##以下出現したエラー
84---------------------------------------------------------------------------
85RuntimeError                              Traceback (most recent call last)
86<ipython-input-8-2396f00cf9e1> in <module>()
87      3 earlystop_counter = 0
88      4 for epoch in range(100):
89----> 5     train_loss = training(model, train_dataset, optimizer, loss_func)
90      6     val_loss, val_acc = evaluation(model, val_dataset, loss_func)
91      7 
92
936 frames
94<ipython-input-6-33676f519731> in training(model, dataset, optimizer, loss_func)
95      6         img = img.to("cuda")
96      7         target = target.to("cuda")
97----> 8         pred = model(img)
98      9         loss = loss_func(pred, target)
99     10         losses.append(loss.item())
100
101/usr/local/lib/python3.6/dist-packages/torch/nn/modules/module.py in _call_impl(self, *input, **kwargs)
102    720             result = self._slow_forward(*input, **kwargs)
103    721         else:
104--> 722             result = self.forward(*input, **kwargs)
105    723         for hook in itertools.chain(
106    724                 _global_forward_hooks.values(),
107
108/usr/local/lib/python3.6/dist-packages/torchvision/models/resnet.py in forward(self, x)
109    218 
110    219     def forward(self, x):
111--> 220         return self._forward_impl(x)
112    221 
113    222 
114
115/usr/local/lib/python3.6/dist-packages/torchvision/models/resnet.py in _forward_impl(self, x)
116    201     def _forward_impl(self, x):
117    202         # See note [TorchScript super()]
118--> 203         x = self.conv1(x)
119    204         x = self.bn1(x)
120    205         x = self.relu(x)
121
122/usr/local/lib/python3.6/dist-packages/torch/nn/modules/module.py in _call_impl(self, *input, **kwargs)
123    720             result = self._slow_forward(*input, **kwargs)
124    721         else:
125--> 722             result = self.forward(*input, **kwargs)
126    723         for hook in itertools.chain(
127    724                 _global_forward_hooks.values(),
128
129/usr/local/lib/python3.6/dist-packages/torch/nn/modules/conv.py in forward(self, input)
130    417 
131    418     def forward(self, input: Tensor) -> Tensor:
132--> 419         return self._conv_forward(input, self.weight)
133    420 
134    421 class Conv3d(_ConvNd):
135
136/usr/local/lib/python3.6/dist-packages/torch/nn/modules/conv.py in _conv_forward(self, input, weight)
137    414                             _pair(0), self.dilation, self.groups)
138    415         return F.conv2d(input, weight, self.bias, self.stride,
139--> 416                         self.padding, self.dilation, self.groups)
140    417 
141    418     def forward(self, input: Tensor) -> Tensor:
142
143RuntimeError: Input type (torch.cuda.FloatTensor) and weight type (torch.FloatTensor) should be the same
144

google colabを使ってpytorchを活用したresnet34の転移学習を上記のコードで行うとしたのですが、エラーが出てしました。どなたか解決策をご教授していただけますか？
更新10/28

tiitoi

2020/10/27 09:50

stacktrace を省略せず、画像ではなくテキストで貼ってください

TaichiTsujikawa

2020/10/27 16:52

追加させていただきました。ご確認していただけると幸いです。

tiitoi

2020/10/27 19:45

エラーを見る限り、モデルが GPU 上にないってエラーです。 pred = model(img) を model = model.to("cuda") pred = model(img) に置き換えるとどうなりますか?

TaichiTsujikawa

2020/10/30 10:57

できました！！ありがとうございます！！ ----------------------------------------------------------------------------------------------------------------------------- epoch 00 train_loss: 1.9293387055397033 val_loss: 1.360185702641805 val_acc: 0.6573426573426573 epoch 01 train_loss: 1.1356693744659423 val_loss: 0.6244564255078634 val_acc: 0.8741258741258742 epoch 02 train_loss: 0.5705957621335983 val_loss: 0.3632519543170929 val_acc: 0.8881118881118881 epoch 03 train_loss: 0.35443189442157746 val_loss: 0.29364600280920666 val_acc: 0.8951048951048951 epoch 04 train_loss: 0.37619893848896024 val_loss: 0.28837616990009945 val_acc: 0.8811188811188811 epoch 05 train_loss: 0.288086473941803 val_loss: 0.22834072758754095 val_acc: 0.9300699300699301 epoch 06 train_loss: 0.24443146735429763 val_loss: 0.2042258307337761 val_acc: 0.9300699300699301 epoch 07 train_loss: 0.2630683213472366 val_loss: 0.20026926199595133 val_acc: 0.9230769230769231 epoch 08 train_loss: 0.28614960685372354 val_loss: 0.20936919500430426 val_acc: 0.916083916083916 epoch 09 train_loss: 0.20852348133921622 val_loss: 0.18482509503761926 val_acc: 0.9230769230769231 epoch 10 train_loss: 0.16087361611425877 val_loss: 0.19408025095860162 val_acc: 0.9230769230769231 epoch 11 train_loss: 0.14762699361890555 val_loss: 0.1882300873597463 val_acc: 0.9300699300699301 epoch 12 train_loss: 0.12694987654685974 val_loss: 0.18787698447704315 val_acc: 0.9300699300699301 epoch 13 train_loss: 0.13299292735755444 val_loss: 0.18518761545419693 val_acc: 0.9370629370629371 epoch 14 train_loss: 0.1471386682242155 val_loss: 0.1817475681503614 val_acc: 0.9300699300699301 epoch 15 train_loss: 0.14307339126244187 val_loss: 0.1682852034767469 val_acc: 0.9300699300699301 epoch 16 train_loss: 0.11291725002229214 val_loss: 0.16316993037859598 val_acc: 0.916083916083916 epoch 17 train_loss: 0.09629514738917351 val_loss: 0.17270603279272714 val_acc: 0.916083916083916 epoch 18 train_loss: 0.11534484550356865 val_loss: 0.1768180876970291 val_acc: 0.916083916083916 epoch 19 train_loss: 0.08179103210568428 val_loss: 0.16468751430511475 val_acc: 0.916083916083916 epoch 20 train_loss: 0.13136351406574248 val_loss: 0.1816187178095182 val_acc: 0.916083916083916 epoch 21 train_loss: 0.09004241377115249 val_loss: 0.18401368210713068 val_acc: 0.9230769230769231