cuDNN error: CUDNN_STATUS_NOT_INITIALIZED

1

I receive the following error:

    filtered_y_y = F.conv2d(y_c, sobel_y_filter, stride=1, padding=1)
RuntimeError: cuDNN error: CUDNN_STATUS_NOT_INITIALIZED

The problem occurs, when I apply a sobel filter to a 2d feature map and run it on cuda. The code looks as follows:

def AbsDetJacobian(batch_meshgrid, device):
    """
        :param batch_meshgrid: takes meshgrid tensor of dim [bn, 2, h, w] (conceptually meshgrid represents a two dimensional function f = [fx, fy] on [bn, h, w] )
        :return: returns Abs det of  Jacobian of f of dim [bn, 1, h, w]
        """
    y_c = batch_meshgrid[:, 0, :, :].unsqueeze(1)
    x_c = batch_meshgrid[:, 1, :, :].unsqueeze(1)
    sobel_x_filter = 1 / 4 * torch.tensor([[-1, 0, 1], [-2, 0, 2], [-1, 0, 1]]).reshape(1, 1, 3, 3).to(device)
    sobel_y_filter = sobel_x_filter.permute(0, 1, 3, 2).to(device)


    filtered_y_y = F.conv2d(y_c, sobel_y_filter, stride=1, padding=1)
    filtered_y_x = F.conv2d(y_c, sobel_x_filter, stride=1, padding=1)
    filtered_x_y = F.conv2d(x_c, sobel_y_filter, stride=1, padding=1)
    filtered_x_x = F.conv2d(x_c, sobel_x_filter, stride=1, padding=1)

    Det = torch.abs(filtered_y_y * filtered_x_x - filtered_y_x * filtered_x_y)

    return Det

Does anyone have an idea what’s going on here?

2

I cannot reproduce this issue using this code snippet:

device = 'cuda'
batch_meshgrid = torch.empty(4, 2, 24, 24).uniform_(-1, 1).to(device)

y_c = batch_meshgrid[:, 0, :, :].unsqueeze(1)
x_c = batch_meshgrid[:, 1, :, :].unsqueeze(1)
sobel_x_filter = 1 / 4 * torch.tensor([[-1, 0, 1], [-2, 0, 2], [-1, 0, 1]]).reshape(1, 1, 3, 3).to(device)
sobel_y_filter = sobel_x_filter.permute(0, 1, 3, 2).to(device)


filtered_y_y = F.conv2d(y_c, sobel_y_filter, stride=1, padding=1)
filtered_y_x = F.conv2d(y_c, sobel_x_filter, stride=1, padding=1)
filtered_x_y = F.conv2d(x_c, sobel_y_filter, stride=1, padding=1)
filtered_x_x = F.conv2d(x_c, sobel_x_filter, stride=1, padding=1)

Det = torch.abs(filtered_y_y * filtered_x_x - filtered_y_x * filtered_x_y)
print(Det.shape)

Could you post the shapes you’ve used to run into this issue?

3

Have you solved the issue? If so, would you mind posting the solution? Thank you