Specify retain_graph=True when calling backward the first time

autograd
1

Hi all, My code doesn’t work unless I specify retain_graph=True eventhough I am calling backward once every iteration. Thanks in advance

    depth_net_encoder.eval()
    depth_net_decoder.eval()
    noise=Variable(torch.tensor(img), requires_grad=True)
    while True  :
        noise=Variable(torch.tensor(noise.data), requires_grad=True)
        dummy=torch.zeros(tgt_img_var.shape).to(device)   
        true_mask=torch.zeros(tgt_img_var.shape).to(device)        
        for u in range(dummy.shape[0]):
            ex1=ex2=0 
            dummy[u,:,ry_list[u]-sz2+ex1:ry_list[u]+sz2+ex1, rx_list[u]-sz2+ex2:rx_list[u]+sz2+ex2]=F.sigmoid(noise)
            true_mask[u,:,ry_list[u]-sz2+ex1:ry_list[u]+sz2+ex1, rx_list[u]-sz2+ex2:rx_list[u]+sz2+ex2]=torch.tensor(1).to(device)
        adv_tgt_img_var = torch.mul((1-true_mask), tgt_img_var) + torch.mul(true_mask, dummy)
    enc_var=depth_net_encoder(adv_tgt_img_var)
        depth_var = depth_net_decoder(enc_var)   
        loss_data=0.0
        for i in range(1):
            o_g_x,o_g_y=torch.abs(X_grad(target_depth[("disp", i)],device)),torch.abs(Y_grad(target_depth[("disp", i)],device))
            o_g=o_g_x+o_g_y
            g_x,g_y=torch.abs(X_grad(depth_var[("disp", i)],device)),torch.abs(Y_grad(depth_var[("disp", i)],device))
            g=g_x+g_y
            if i==0:
                p_O,p_f=o_g,g
      
            gradient=torch.log(1+g_x*o_g_x).sum()+torch.log(1+g_y*o_g_y).sum()
            loss_data=gradient

        loss=loss_data*100#+loss_reg
        loss.backward()
        noise_grad = noise.grad.clone()
        noiset=noise-0.001*noise_grad
        noise=torch.tensor(noiset.cpu().detach().numpy())
        loss_scalar = loss.item()
        print(loss_scalar)```
2

Hi,

What is the error you’re seeing?

Also you should not use .data and Variables anymore.
You can replace noise=Variable(torch.tensor(img), requires_grad=True) by noise=torch.tensor(img, requires_grad=True) for example.

3

My issue is eventhough I am calling backward once every iteration,it asks me to set retain_graph=True . What am I doing wrong?

4

It is usually that the graph is reused.
So either that you do some differentiable computation outside the loop, or that you re-use results from one iteration to the next.

I would usually try to get a repro as small as possible. That helps identify the potential faulty Tensors as mentioned above.
If that doesn’t help, you can use something like torchviz to print the graph and see where the one from the second iteration is attached to the one in the first.