RuntimeError: does not have a grad_fn

autograd
1

I’m trying to use the model output as indexes to create a 3d point cloud. Because in order to calculate the loss I want to convolve it.
code:

for epoch in range(num_epochs):
    optimizer.zero_grad()  # Zero the gradients
    output = model(masked_coords)
    points_3d = gen_3d(200,200,35, mask, output)
    psf_output = conv_3D(points_3d,IPSF)
    loss = loss_fn(psf_output, target)
    loss. Backward()  # Backpropagation
    optimizer.step()  # Update weights

This is my gen_3d function:

def gen_3d(x,y,z, mask, depths):
    pc = torch.zeros(x,y,z)
    _pc = pc.view(-1,z)
    m = mask.flatten()
    d = (depths*z).int().flatten()
    _pc[m, d] += 1
    return torch.nn.functional.pad(pc, (0, 0, 10, 10, 10, 10))

Error:

image
image780×619 31.6 KB

seems like the process is not differentiable. Is there a way to solve this?

2

Converting depth to an integer detaches the tensor from the computation graph so the error is expected. I don’t know how the point cloud will be used and if there is a valid solution so let’s wait for other inputs.

3

Thanks for the reply. Is there a way to bypass this and connect the gradients? (ex. coping the gradients into respective indices)