How to let TMVA & pytorch to use GPU?

Hi, my current working environment is ROOT 6.26.14, Python 3.8.19, Windows 10, Pytorch 2.2

I have decided to start my work based on the example code(“tutorials\tmva\pytorch\”) given in the tutorial. It does work well, but I found that this code does not use GPU and only works with the CPU. However, PyTorch is clearly capable of working with the GPU, so I tried to insert model.cuda() in the middle, and it resulted in an error when running it again. The error message is listed as below:

                     : Option SaveBestOnly: Only model weights with smallest validation loss will be stored

: Failed to run python code: trained_model = fit(model, train_loader, val_loader, num_epochs=numEpochs, batch_size=batchSize,optimizer=optimizer, criterion=criterion, save_best=save_best, scheduler=(schedule, schedulerSteps))
: Python error message:
Traceback (most recent call last):
File “”, line 1, in
File “”, line 87, in train
output = model(X)
File “C:\Users\phoenixAspies.conda\envs\Python38\lib\site-packages\torch\nn\modules\”, line 1511, in _wrapped_call_impl
return self._call_impl(*args, **kwargs)
File “C:\Users\phoenixAspies.conda\envs\Python38\lib\site-packages\torch\nn\modules\”, line 1520, in _call_impl
return forward_call(*args, **kwargs)

I believe my PyTorch and CUDA settings are correct, as when I directly use “import torch; x = torch.randn(4,4); x.cuda()”, the variable is successfully placed on the GPU.

Does anyone know how to let GPU work when using TMVA & PyTorch?

Dear @phoenixAspies ,

Thanks for reaching out to the forum!

I believe @moneta can comment regarding your issue.



I think you need to transfer to the device (the GPU) both the input data and the model. In the case of the RegressionPyTorch function, you need to modify the train and predict functions for this.


Thanks! But I don’t know how to let transfer the data which TMVA proceeds to GPU, for example, in pytorch, I can simply write model.cuda() / data.cuda(). Is there any function in TMVA to do that?

TMVA has native functions to transfer to CUDA, but these work for training using a model build using TMVA directly (with its DL package).
In case of PyTorch you would need to call data.cuda() in the train and predict functions, in this way the data should be transfer to GPU.


Thanks! I found that adding .cuda() to part of the pytorch code is enough to run on GPU now.