Resume my custom dataset of scannet failed for pytorch_geometry

[ruomingzhai]

I try to train the scannet datatset in custom_dataset mode , it goes well in the first 20 epoch without pyg,and collpases at 30 epoch.And when I try to continue the training with --RESUME option , it comes out a bug like the following: Traceback (most recent call last): File "learning/main.py", line 463, in <module> main() File "learning/main.py", line 168, in main model = create_model(args, dbinfo) File "learning/main.py", line 427, in create_model model.ecc = graphnet.GraphNetwork(args.model_config, nfeat, [dbinfo['edge_feats']] + args.fnet_widths, args.fnet_orthoinit, args.fnet_llbias,args.fnet_bnidx, args.edge_mem_limit, use_pyg = args.use_pyg, cuda = args.cuda) File "/mnt/superpoint_graph_zrm/learning/../learning/graphnet.py", line 78, in __init__ gconv = RNNGraphConvModule(cell, fnet, nfeat, vv = vv, nrepeats=nrepeats, cat_all=cat_all, edge_mem_limit=edge_mem_limit, use_pyg = use_pyg, cuda = cuda) File "/mnt/superpoint_graph_zrm/learning/../learning/modules.py", line 145, in __init__ self.nn = NNConv(nfeat, nfeat, vv = vv) NameError: name 'NNConv' is not defined This bug makes no sense because it went well in the first 30 epoch and I did not change any code during this training process. Can anyone give me some clues? Thanks!

[loicland]

Hi,

do you have PyTorch Geometric installed on your computer?

[ruomingzhai]

Hi,

do you have PyTorch Geometric installed on your computer?

I did not install the pytorch geometric because I thought it is optional choice.my programmes went well without pytorch geometric until the 30th epoch.I don't know what is going on.

[loicland]

can you check the value of args.use_pyg just after the checkpoint is loaded?

print(args.use_pyg )

[ruomingzhai]

can you check the value of args.use_pyg just after the checkpoint is loaded?

print(args.use_pyg )

and the program works now, I found that this "if use_pyg:" return true anyway no matter what the value of use_pyg is.Because I add the print(args.use_pyg) you recommand and it shows 0,but the code inside the "if use_pyg:" still be excuted.So I change and add two lines as following: self.use_pyg = False use_pyg=False in modules.py in line 143 & 144

[ruomingzhai]

can you check the value of args.use_pyg just after the checkpoint is loaded? print(args.use_pyg )

and the program works now, I found that this "if use_pyg:" return true anyway no matter what the value of use_pyg is.Because I add the print(args.use_pyg) you recommand and it shows 0,but the code inside the "if use_pyg:" still be excuted.So I change and add two lines as following: self.use_pyg = False use_pyg=False in modules.py in line 143 & 144

can you check the value of args.use_pyg just after the checkpoint is loaded? print(args.use_pyg )

and the program works now, I found that this "if use_pyg:" return true anyway no matter what the value of use_pyg is.Because I add the print(args.use_pyg) you recommand and it shows 0,but the code inside the "if use_pyg:" still be excuted.So I change and add two lines as following: self.use_pyg = False use_pyg=False in modules.py in line 143 & 144

can you check the value of args.use_pyg just after the checkpoint is loaded?

print(args.use_pyg )

I install pyg and the problem is sloved, thanks for your guidance!

[kangx326]

Hi, do you have PyTorch Geometric installed on your computer?

I did not install the pytorch geometric because I thought it is optional choice.my programmes went well without pytorch geometric until the 30th epoch.I don't know what is going on.

Hi, can you share your implementation on ScanNet-v2? I'm working on this right now and need some help.

[loicland]

In the end, we never got around to implement it.

This repo use a superpoint approach on ScanNet, maybe it will help you.