search

pyg-team / pytorch_geometric

Graph Neural Network Library for PyTorch
https://pyg.org
MIT License
21.08k stars 3.63k forks source link

[Roadmap] New PyTorch Geometric Releases #1465

Open rusty1s opened 4 years ago

rusty1s commented 4 years ago

[Roadmap]

1.6.1

1.6.2

1.6.3.

1.7.0+

sbonner0 commented 4 years ago

I am not sure if this is helpful, or indeed even the correct place for this - please let me know if not and I can remove it. However I thought it might be useful to note that I have been using pytorch geometric with pytorch 1.6.0 over the past few days and haven't noticed any issues. My exact versions of the relevant packages are:

torch                      1.6.0rc7

torch-cluster              1.5.6

torch-geometric            1.6.0

torch-scatter              2.0.5

torch-sparse               0.6.6

torch-spline-conv          1.2.0

I can run the test suite if it would be helpful?

rusty1s commented 4 years ago

Thanks for letting me know. I also encountered no major issues while upgrading. However, I still need to provide updated wheels for the latest PyTorch release.

yangky11 commented 4 years ago

Hi @rusty1s ,

I tried to install PyTorch Geometric with PyTorch 1.6.0 by following the instructions here. However, I cannot find the updated wheels for torch-sparse and torch-cluster. May I know if they will become available in the near future? Thanks!

rusty1s commented 4 years ago

Sure, they should be available next week :)

aluo-x commented 4 years ago

Just built sparse/scatter/cluster/spline-conv/geometric with PyTorch master, CUDA 11, cudnn8, GCC 9.3, MAGMA 2.5.2. Encountered no issues either compiling or running.

lgray commented 4 years ago

Just to add more to this, I've been using torch 1.6 throughout its development period, including the RCs. Never any problems.

GTaherzadeh commented 4 years ago

Hi there,

I sm trying to read my data using TORCH_GEOMETRIC.IO.TU code. My data type is exactly similar to the enzymes dataset. However, the program stops with the Illegal instruction (core dumped) error at this line "node_slice = torch.cumsum(torch.from_numpy(np.bincount(batch)), 0)" in the split function. It seems that the problem is with the dimension but I can't figure out what the solution is. Could you please help with this.

Please note, I installed the package exactly as it is mentioned in the document.

Thanks

rusty1s commented 4 years ago

Not sure if that's the best place to discuss this issue :)

You can create a separate issue for that. If you also provide your raw data, I can help you identify the issue.

GTaherzadeh commented 4 years ago

Hi Matthias,

Thanks for your quick reply. I have created a new issue here: https://github.com/rusty1s/pytorch_geometric/issues/1601

jannessm commented 3 years ago

Hi Matthias, maybe it would be worth to replace network-x by cugraph, when gpu is supported.

I just have heard about it and thought, that it would make sense to give you a hint, if you havent heard about it before.

Regrads.

[EDIT:] I have just seen, that nx is just used in datasets and examples. Therefore, it seems not to be that util. So just in case you need algorithms from network-x, it could be worth to have a look at cugraph. Or for numpy, have a look at cupy ;).

levulinh commented 3 years ago

Please make sure you have the SuperpixelMnist dataset fixed in the next updates, thank you.

rusty1s commented 3 years ago

cugraph conversions is a good idea, I added it to the roadmap. The SuperpixelDataset is already fixed in master.

rusty1s commented 3 years ago

Just released PyTorch Geometric 1.6.2!

Thanks to many wonderful contributors :) Please let me know if you encounter any issues.

eduardohenriquearnold commented 3 years ago

Can we add the FAConv to the next release? It's already in the source. Thanks!

rusty1s commented 3 years ago

I have plans to release a new version today :)

karthi0804 commented 3 years ago

@rusty1s Very eager to get half support. When can we expect this? Thanks in advance.

rusty1s commented 3 years ago

I'm pretty sure that PyTorch 1.9.0 contains all the necessary tools to allow for half-precision in torch-scatter and torch-sparse. I will look into it.

always-ready2learn commented 2 years ago

Hi @rusty1s I have changes TGN.py model for link prediction into edge classification. The model runs for me but unfortunately loss is very high and remains constant through out which means that model is not learning at all. I am not sure what I am missing here. So far I am using similar TGN link prediction example and have made following changes in the train and test functions as:
y_pred = pos_out.argmax(dim=1) y_true = batch.y loss = criterion(y_pred.float(), y_true.float()) I am not sure what I am missing here. the loss value turns out to be 8.02 and remains constant, same goes for accuracy which turns out to be a constant value of .65.Could you think of what I might be missing please?

rusty1s commented 2 years ago

If you want to perform classification, then I think y_true shouldn't be a floating-point tensor. What's the loss function you are using? Note that this is best discussed in a separate issue/discussion - not necessarily in this roadmap issue here :)

always-ready2learn commented 2 years ago

Hi thanks for your reply, I changed y_true to long and y_pred to float. The loss function I tried with is BCELoss and CrossEntropyLoss. Same issue is occurring with both. Thanks for pointing that, I found it to be similar topic but you are right issue is quite diff. I’ll move this discussion to a separate issue sooner. :)

rajveer43 commented 1 year ago

@rusty1s would like to give it a shot! spspmm backward implementation

rusty1s commented 1 year ago

I think this is no longer an issue since we are using PyTorch's spspmm functionality by now.