How to put tensor on a custom Function to cuda device?

autograd
1

Hi,
I custom a Function, in forward method. I have a numpy array x and a cuda tensor input.How to copy value of x to input tensor ? I have try, in forward method, torch.to(device) don’t work.

Thanks

class MyFunction(torch.autograd.Function):

    @staticmethod
    def forward(ctx, input):
        x = np.zeros(shape)
        # do something for x
        # and how to copy data of x to gpu tensor input
        
        return ...

    @staticmethod
    def backward(ctx, grad_output):
          '''
2

Here is an answer, that might work for you.
You should create Tensor from NumPy array and then transfer it to device

3

Hi, I create a tensor from numpy array. But the tensor to cuda device don’t work.

4

At first you should check if CUDA devices are available.
Then set the device variable with some value (e.g. 'cpu', 'cuda:0') and pass it to your_tensor.to() function.
Note: set a constant string value for the device is not an only option (if you want use tensor.to() for transfering to device), you may pass there a device value of some other tensor