veason-silverbullet / nnr Goto Github PK

I test DFM on MIND-small. Find the performance of DFM outperforms all the baselines except your proposed model.
This is puzzling, since DFM is proposed in 2018 and the framework is naive.
Do you have similar situation? Could you share your thinking on this?

When I run this code on the torch-1.10.1, facing the issue as follows:

RuntimeError: one of the variables needed for gradient computation has been modified by an inplace operation: [torch.cuda.FloatTensor [400, 1]], which is output 0 of TBackward, is at version 2; expected version 1 instead. Hint: the backtrace further above shows the operation that failed to compute its gradient. The variable in question was changed in there or anywhere later. Good luck!

This points to the code in trainer.py:
loss.backward()
But the code is worked on the torch-1.9.0.

I try to run libfm in general recommendation methods, but the os.system in libfm_main.py doesn't work as I get an error as follows:

FileNotFoundError: [Errno 2] No such file or directory: './test/res/libfm/1/libfm'

I run this code on Ubunutu 20.4 and windows10 respectively, but the same problem.
How should I do?

When running the baseline FIM on A100(40GB), the training time is too long (over 15hrs can not finish an epoch).
I try to maximize the bachsize, but the situation is not solved. Moreover, the volatile GPU util is 20%.
What's the problem?
How much time is needed, and how much computing resource is needed?