input must be a CUDA tensor error when using AdaPool3d

[Tloops]

Great work!
However, when i'm using AdaPool3d, and i encoutered the error below:

Traceback (most recent call last):
  File "<masked>/main.py", line 339, in <module>
    main()
  File "<masked>/main.py", line 329, in main
    train_loss, train_acc = train(model, train_loader, epoch, criterion, optimizer)
  File "<masked>/main.py", line 113, in train
    outputs = model(inputs)
  File "<masked>/miniconda3/envs/python39/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1511, in _wrapped_call_impl
    return self._call_impl(*args, **kwargs)
  File "<masked>/miniconda3/envs/python39/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1520, in _call_impl
    return forward_call(*args, **kwargs)
  File "<masked>/miniconda3/envs/python39/lib/python3.9/site-packages/torch/nn/parallel/data_parallel.py", line 183, in forward
    return self.module(*inputs[0], **module_kwargs[0])
  File "<masked>/miniconda3/envs/python39/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1511, in _wrapped_call_impl
    return self._call_impl(*args, **kwargs)
  File "<masked>/miniconda3/envs/python39/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1520, in _call_impl
    return forward_call(*args, **kwargs)
  File "<masked>/models/cnn3d_adapool.py", line 52, in forward
    x = self.pool(x)
  File "<masked>/miniconda3/envs/python39/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1511, in _wrapped_call_impl
    return self._call_impl(*args, **kwargs)
  File "<masked>/miniconda3/envs/python39/lib/python3.9/site-packages/torch/nn/modules/module.py", line 1520, in _call_impl
    return forward_call(*args, **kwargs)
  File "<masked>/miniconda3/envs/python39/lib/python3.9/site-packages/adaPool-0.1-py3.9-linux-x86_64.egg/adaPool/idea.py", line 1495, in forward
    return adapool3d(x, beta=self.beta, kernel_size=self.kernel_size, stride=self.stride, return_mask=self.return_mask, native=self.native)
  File "<masked>/miniconda3/envs/python39/lib/python3.9/site-packages/adaPool-0.1-py3.9-linux-x86_64.egg/adaPool/idea.py", line 771, in adapool3d
    x = beta*CUDA_ADAPOOL3d_EDSCW.apply(x, kernel_size, stride, return_mask) + (1. - beta)*CUDA_ADAPOOL3d_EM.apply(x, kernel_size, stride, return_mask)
  File "<masked>/miniconda3/envs/python39/lib/python3.9/site-packages/torch/autograd/function.py", line 553, in apply
    return super().apply(*args, **kwargs)  # type: ignore[misc]
  File "<masked>/miniconda3/envs/python39/lib/python3.9/site-packages/torch/cuda/amp/autocast_mode.py", line 123, in decorate_fwd
    return fwd(*args, **kwargs)
  File "<masked>/miniconda3/envs/python39/lib/python3.9/site-packages/adaPool-0.1-py3.9-linux-x86_64.egg/adaPool/idea.py", line 505, in forward
    adapool_cuda.forward_3d_edscw(input.contiguous(), kernel, stride, output, return_mask, mask)
RuntimeError: input.is_cuda() INTERNAL ASSERT FAILED at "CUDA/adapool_cuda.cpp":616, please report a bug to PyTorch. input must be a CUDA tensor

I've checked the if the input is on the cuda by printing input.is_cuda just before the adapool_cuda.forward_3d_edscw function call, and it displays True. But, when it comes to the cpp file, the input became NOT a CUDA tensor. I'm really comfused about that. Hope to receive your reply soon. Thanks!

[alexandrosstergiou]

Hi @Tloops,

Can you share a minimal snippet of your code?
A quick solution would be to explicitly do .cuda() in the call.

[Tloops]

Thanks for your reply!

I've tried adapool_cuda.forward_3d_edscw(input.contiguous().cuda(), kernel, stride, output, return_mask, mask) in idea.py, but I still get the same error above.

A minimal snippet of my code (the model part) is as follows:

import adapool_cuda
from adaPool import AdaPool3d

class CNN3D(nn.Module):
    def __init__(self, num_classes):
        super(CNN3D, self).__init__()
        self.cnn = nn.Sequential(
            BasicConv(1, 8, kernel_size=(3, 3, 3), stride=(1, 1, 1), padding=(1, 1, 1)),
            nn.MaxPool3d(kernel_size=(2, 2, 2), stride=(2, 2, 2), padding=(1, 0, 1)),
            BasicConv(8, 16, kernel_size=(3, 3, 3), stride=(1, 1, 1), padding=(1, 1, 1)),
            nn.MaxPool3d(kernel_size=(2, 2, 2), stride=(2, 2, 2), padding=(1, 0, 0)),
            BasicConv(16, 32, kernel_size=(3, 3, 3), stride=(1, 1, 1), padding=(1, 1, 1)),
            nn.MaxPool3d(kernel_size=(2, 2, 2), stride=(2, 2, 2), padding=(1, 0, 1)),
            BasicConv(32, 64, kernel_size=(3, 3, 3), stride=(1, 1, 1), padding=(1, 1, 1)),
            nn.MaxPool3d(kernel_size=(2, 2, 2), stride=(2, 2, 2), padding=(0, 0, 1)),
            BasicConv(64, 128, kernel_size=(3, 3, 3), stride=(1, 1, 1), padding=(1, 1, 1)),
            nn.MaxPool3d(kernel_size=(2, 2, 2), stride=(2, 2, 2), padding=(1, 1, 0)),
        )
        self.pool = AdaPool3d(kernel_size=(6,7,6), beta=torch.Tensor([[[0.5]]]))
        self.fc = nn.Linear(128, num_classes)
        self.softmax = nn.Softmax()
    
    def forward(self, x):
        x = self.cnn(x)
        x = self.pool(x)
        x = x.view(x.size(0), -1)
        x = self.fc(x)
        x = self.softmax(x)
        return x

[alexandrosstergiou]

You need the beta tensor to also be in cuda, i.e:

pool = AdaPool3d(kernel_size=(6,7,6), beta=torch.Tensor([[[0.5]]]).cuda())

[Tloops]

You need the beta tensor to also be in cuda, i.e:

pool = AdaPool3d(kernel_size=(6,7,6), beta=torch.Tensor([[[0.5]]]).cuda())

I've tried it but I still get the same error above. Really strange... 😕

[alexandrosstergiou]

I used the script below on two different machines:

import torch
import adapool_cuda
from adaPool import AdaPool3d

pool = AdaPool3d(kernel_size=(6,7,6), beta=torch.Tensor([[[0.5]]]).cuda())
x = torch.rand(1,1,16,224,224).cuda()
out = pool(x)
print(out.shape)

Wasn't able to reproduce the error