[BUG] TypeError in explain function for TFT

[studentangerer]

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

in tft_explainer.py:211 -> can be fixed with
attention_heads = self.model.model._attn_out_weights.detach().cpu().numpy().sum(axis=-2)

another TypeError occurs in tft_explainer.py:506

To Reproduce
When executing darts/examples/13-TFT-examples.ipynb, without any modifications.

In particular the following line causes the error.
explainability_result = explainer.explain()

System:

• Python version: 3.9.16
• darts version: 0.25.0

[dennisbader]

Hi @studentangerer, this is indeed a bug when running TFTModel on GPU. We'll fix this soon.

As a hotfix, you could save the model and then load it to cpu:

model.save("my_model.pt")
model_cpu = TFTModel.load("my_model.pt", map_location="cpu")
model_cpu.to_cpu()

explainer = TFTExplainer(model_cpu)

After that the explainer should work as expected.