Problem with get_minirocket_features while using CUDA in training

[Scienceseb]

Hi!

So this code is working:

mrf = MiniRocketFeatures(np.asarray(X_train).shape[1], np.asarray(X_train).shape[2]).to(default_device())
X_train = torch.from_numpy((np.asarray(X_train))).to(default_device())
mrf.fit(X_train) 

but I cannot do:
X_feat = get_minirocket_features(np.asarray(X_train), mrf, chunksize=1024)

I get the following error:
TypeError: can't convert cuda:0 device type tensor to numpy. Use Tensor.cpu() to copy the tensor to host memory first.

While when I put everything on the CPU :

mrf = MiniRocketFeatures(np.asarray(X_train).shape[1], np.asarray(X_train).shape[2])
X_train = torch.from_numpy((np.asarray(X_train)))
mrf.fit(X_train)

X_feat = get_minirocket_features(np.asarray(X_train), mrf, chunksize=1024)

It works perfectly.

So the issue is that when the model is train using the GPU (CUDA) it work, but I cannot use the get_minirocket_features function.
While when the model is train using the CPU, I can use the get_minirocket_features but the GPU acceleration is lost.

Thanks for your help.

Kind regards,
Sébastien de Blois

[oguiza]

Hi @Scienceseb,

I've reviewed this and it was indeed a but. I've already fixed it now.

I've also modified the code so that the fit method in MiniRocketFeatures can either take a np.ndarray or torch.Tensor. I think it's easier to remember than X or X_train is always a np.ndarray. In this way you can now do the following:

dsid = 'LSST'
X_train, y_train, X_valid, y_valid = get_UCR_data(dsid, split_data=True)
mrf = MiniRocketFeatures(X_train.shape[1], X_train.shape[2]).to(default_device())
mrf.fit(X_train) # here X_train is a np.ndarray
X_feat = get_minirocket_features(X_train, mrf, chunksize=1024) # X_feat is also a np.ndarray

Even if you pass an np.ndarray, the entire process is run on a gpu if available.

It'd be good if you would test if it works well.

Thanks!

[Scienceseb]

Thanks everything works well!