# Problem when moving model to cuda **URL:** \ **Category:** Uncategorized\ **Created:** [July 26, 2020, 1:37am UTC](https://discuss.pytorch.org/t/problem-when-moving-model-to-cuda/90549 "2020-07-26T01:37:45Z")\ **Posts on this page:** 6\ **Page:** 1
**Author:** ![CaoHoangTung](https://discuss.pytorch.org/user_avatar/discuss.pytorch.org/caohoangtung/32/27129_2.png) [@CaoHoangTung](https://discuss.pytorch.org/u/CaoHoangTung)\ **Post date:** [July 26, 2020, 1:37am UTC](https://discuss.pytorch.org/t/problem-when-moving-model-to-cuda/90549/1 "2020-07-26T01:37:45Z")
I’m implementing ResNet50 with PyTorch. It ran perfectly okay on cpu, but when I moved the model to GPU, I encountered this error ```auto RuntimeError Traceback (most recent call last) in () 1 input = torch.rand(32,3,224,224).to(device) ----> 2 output = model(input) 12 frames /usr/local/lib/python3.6/dist-packages/torch/nn/modules/conv.py in _conv_forward(self, input, weight) 348 _pair(0), self.dilation, self.groups) 349 return F.conv2d(input, weight, self.bias, self.stride, --> 350 self.padding, self.dilation, self.groups) 351 352 def forward(self, input): RuntimeError: Input type (torch.cuda.FloatTensor) and weight type (torch.FloatTensor) should be the same ``` Here is my related code ```auto class BottleNeck(nn.Module): def __init__ (self,base_kernel,input_kernel_depth,stride=1): super(). __init__ () self.base_kernel = base_kernel self.input_kernel_depth = input_kernel_depth self.stride = stride self.conv1 = nn.Conv2d(in_channels=input_kernel_depth,out_channels=base_kernel,kernel_size=1,stride=stride) self.bn1 = nn.BatchNorm2d(base_kernel) self.conv2 = nn.Conv2d(in_channels=base_kernel,out_channels=base_kernel,kernel_size=3,padding=1) self.bn2 = nn.BatchNorm2d(base_kernel) self.conv3 = nn.Conv2d(in_channels=base_kernel,out_channels=base_kernel*4,kernel_size=1) self.bn3 = nn.BatchNorm2d(base_kernel*4) def forward(self,input): identity = input output = self.conv1(input) output = self.conv2(output) output = self.conv3(output) upsampler = self.upsampling(self.input_kernel_depth,self.base_kernel*4,self.stride) identity = upsampler(identity) return F.relu(output + identity) def upsampling(self,in_channels,out_channels,stride=1): upsampler = nn.Conv2d(in_channels,out_channels,kernel_size=1,stride=stride) return upsampler class Block(nn.Module): def __init__ (self,input_kernel_depth,stack_count,base_kernel,stride,padding): super(). __init__ () layers = [] layers.append( BottleNeck(base_kernel,input_kernel_depth,stride=stride) ) for i in range(1,stack_count): layers.append( BottleNeck(base_kernel,base_kernel*4,stride=1) ) self.block = nn.Sequential(*layers) def forward(self,input): output = self.block(input) return output class ResNet50(nn.Module): def __init__ (self,n_classes): super(). __init__ () self.n_classes = n_classes self.stage1 = nn.Conv2d(3,64,7,2,3) self.stage2 = nn.Sequential( nn.MaxPool2d(3,2,1).to(device), Block(input_kernel_depth=64,stack_count=3,base_kernel=64,stride=1,padding=0) ) self.stage3 = nn.Sequential( Block(input_kernel_depth=256,stack_count=4,base_kernel=128,stride=2,padding=1) ) self.stage4 = nn.Sequential( Block(input_kernel_depth=512,stack_count=6,base_kernel=256,stride=2,padding=1) ) self.stage5 = nn.Sequential( Block(input_kernel_depth=1024,stack_count=3,base_kernel=512,stride=2,padding=1) ) self.avg_pool = nn.AvgPool2d(kernel_size=7) self.fc1 = nn.Linear(2048,1000) self.fc2 = nn.Linear(1000,n_classes) def forward(self,input): batch_size = input.size(0) # print("Global input",input.shape) stage1_output = self.stage1(input) # print("Stage 1",stage1_output.shape) stage2_output = self.stage2(stage1_output) # print("Stage 2",stage2_output.shape) stage3_output = self.stage3(stage2_output) # print("Stage 3",stage3_output.shape) stage4_output = self.stage4(stage3_output) # print("Stage 4",stage4_output.shape) stage5_output = self.stage5(stage4_output) # print("Stage 5",stage5_output.shape) output = self.avg_pool(stage5_output).view(batch_size,-1) output = F.relu(self.fc1(output)) output = F.log_softmax(self.fc2(output),dim=1) return output model = ResNet50(4) model.to(device) optimizer = optim.SGD(model.parameters(),lr=0.001,momentum=0.9) criterion = nn.NLLLoss() ``` The error above occured when I tried to run some test input ```auto input = torch.rand(32,3,224,224).to(device) output = model(input) ``` ---
**Author:** ![ptrblck](https://discuss.pytorch.org/user_avatar/discuss.pytorch.org/ptrblck/32/1823_2.png) [@ptrblck](https://discuss.pytorch.org/u/ptrblck)\ **Post date:** [July 26, 2020, 9:57am UTC](https://discuss.pytorch.org/t/problem-when-moving-model-to-cuda/90549/2 "2020-07-26T09:57:14Z")
Based on the error message it seems that some parameters of the model weren’t pushed to the GPU. In particular creating a new module in the `forward` will not push it directly to the GPU, if you don’t specify it: ```python upsampler = self.upsampling(self.input_kernel_depth,self.base_kernel*4,self.stride) ``` You could use `upsampler = self.upsampling(self.input_kernel_depth,self.base_kernel*4,self.stride).to(identity.device)` to solve this issue. Note however, that you are re-initializing this module with random shape. I’m not familiar with your use case, but this module will not be trained. ---
**Author:** ![CaoHoangTung](https://discuss.pytorch.org/user_avatar/discuss.pytorch.org/caohoangtung/32/27129_2.png) [@CaoHoangTung](https://discuss.pytorch.org/u/CaoHoangTung)\ **Post date:** [July 27, 2020, 9:24am UTC](https://discuss.pytorch.org/t/problem-when-moving-model-to-cuda/90549/3 "2020-07-27T09:24:14Z")
Thank you for the reply. It is indeed the solution to my case. However, I do want this module to be trained. What is the proper way to do that? ---
**Author:** ![ptrblck](https://discuss.pytorch.org/user_avatar/discuss.pytorch.org/ptrblck/32/1823_2.png) [@ptrblck](https://discuss.pytorch.org/u/ptrblck)\ **Post date:** [July 27, 2020, 10:32pm UTC](https://discuss.pytorch.org/t/problem-when-moving-model-to-cuda/90549/4 "2020-07-27T22:32:33Z")
You would have to initialize this module in the ` __init__ ` method of your module (as the other layers) and just call it in `forward`. ---
**Author:** ![DBraun](https://discuss.pytorch.org/user_avatar/discuss.pytorch.org/dbraun/32/24595_2.png) [@DBraun](https://discuss.pytorch.org/u/DBraun)\ **Post date:** [July 28, 2020, 1:24am UTC](https://discuss.pytorch.org/t/problem-when-moving-model-to-cuda/90549/5 "2020-07-28T01:24:22Z")
@ptrblck Could you inspect some similar code here? [https://github.com/NathanUA/U-2-Net/blob/b77cd6da3204efcb03e18e15dd3b9eb24d47f969/model/u2net.py#L24](https://github.com/NathanUA/U-2-Net/blob/b77cd6da3204efcb03e18e15dd3b9eb24d47f969/model/u2net.py#L24) `_upsample_like` is called only in the `forward` pass just like OP’s `self.upsampling` I wonder if this is the cause of my libtorch `module.forward()` not working even though my python jit trace saves and imports in libtorch without errors. If you think it is the cause, what is a concise fix considering `_upsample_like` is written a lot in all the `forward` implementations? My other thread is here [Debugging runtime error module-\>forward(inputs) libtorch 1.4](https://discuss.pytorch.org/t/debugging-runtime-error-module-forward-inputs-libtorch-1-4/82415) ---
**Author:** ![ptrblck](https://discuss.pytorch.org/user_avatar/discuss.pytorch.org/ptrblck/32/1823_2.png) [@ptrblck](https://discuss.pytorch.org/u/ptrblck)\ **Post date:** [July 28, 2020, 8:59am UTC](https://discuss.pytorch.org/t/problem-when-moving-model-to-cuda/90549/6 "2020-07-28T08:59:47Z")
Your code looks good, since `_upsample_like` onle uses the functional API via `F.upsample` without any parameters. You could of course create an `nn.Upsample` module, but this wouldn’t make a difference and I think your current code is clear and shouldn’t be changed. I’ll take a look at the linked issue.