Facing issue while training.

[richas46]

Hi, thankyou for sharing your wonderful work.

I've followed the steps given in the setup. When I proceed further to the training step and run python train_net.py --cfg_file configs/scannet/0050.yaml gpus 0, exp_name scannet_0050, I'm facing a RuntimeError: Function 'MmBackward' returned nan values in its 1th output.

How should I solve this error?

[richas46]

This is what the detailed error looks like -

Traceback (most recent call last): File "train_net.py", line 127, in main() File "train_net.py", line 123, in main train(cfg, network) File "train_net.py", line 57, in train trainer.train(epoch, train_loader, optimizer, recorder) File "C:\Users\Local_Admin\Desktop\manhattan_sdf-main\lib\train\trainers\trainer.py", line 86, in train output, loss, loss_stats, image_stats = self.network(batch) File "C:\Users\Local_Admin\anaconda3\envs\manhattan\lib\site-packages\torch\nn\modules\module.py", line 722, in call_impl result = self.forward(*input, **kwargs) File "lib/train/trainers/manhattan_sdf.py", line 14, in forward output = self.net(batch) File "C:\Users\Local_Admin\anaconda3\envs\manhattan\lib\site-packages\torch\nn\modules\module.py", line 722, in call_impl result = self.forward(*input, **kwargs) File "lib/networks/network.py", line 209, in forward perturb=pertube File "lib/networks/network.py", line 166, in volume_render ret_i = render_rayschunk(rays_o[:, i:i+rayschunk], rays_d[:, i:i+rayschunk]) File "lib/networks/network.py", line 114, in render_rayschunk radiances, semantics, sdf, nablas = batchify_query(model.forward, pts, view_dirs.unsqueeze(-2).expand_as(pts)) File "C:\Users\Local_Admin\Desktop\manhattan_sdf-main\lib\utils\net_utils.py", line 153, in batchify_query raw_ret_i = query_fn(*args_i) File "lib/networks/network.py", line 44, in forward sdf, nablas, geometry_feature = self.forward_surface_with_nablas(x) File "lib/networks/network.py", line 40, in forward_surface_with_nablas sdf, nablas, h = self.sdf_net.forward_with_nablas(x) File "C:\Users\Local_Admin\Desktop\manhattan_sdf-main\lib\networks\base.py", line 192, in forward_with_nablas only_inputs=True File "C:\Users\Local_Admin\anaconda3\envs\manhattan\lib\site-packages\torch\autograd_init.py", line 192, in grad inputs, allow_unused) RuntimeError: Function 'MmBackward' returned nan values in its 1th output.

[ghy0324]

Hi, Thanks for your interest! I'm not sure what the reason is. Do you use the conda envrionment file we provide? And is this bug reproducible?

[richas46]

Yes I'm using the conda environment provided.