Error of assert xt.grad is not None

[jiajinghu19]

Hi, I am trying to evaluate on tsp100 data, and have this error saying

File "/T2TCO/diffusion/pl_tsp_model.py", line 247, in test_step
    g_xt = self.guided_categorical_denoise_step(points, g_xt, t1, device, edge_index, target_t=t2)
  File "/T2TCO/diffusion/pl_tsp_model.py", line 91, in guided_categorical_denoise_step
    assert xt.grad is not None

The function has this assert statement to assert xt has gradient, but the input g_xt is converted to long type in line 238, and long type tensor cannot have gradient. Can someone help me with this? Thanks

(g_xt > 0).long()

[yangco-le]

Hi, @jiajinghu19

Please kindly note that this tensor has been converted to float format in line 61 xt = xt.float(). The gradient here is estimated based on x_t. This issue is likely caused by the incorrect version of PyTorch Lightning. PyTorch Lightning itself does not support gradient computation during the testing phase. Therefore, we added two lines of code, torch.set_grad_enabled(True) and with torch.inference_mode(False):, to enable gradient computation. If you still encounter gradient computation issues, please check the version of the libraries or refer to relevant solutions for the version you are using, such as those available at Lightning-AI/pytorch-lightning#13948.

Should you have any further inquiries, please feel free to reach out. By the way, kindly mention me in the comments (@ me) so that I can promptly receive notifications and address your question.