Loss explosion issue in ConvnLSTMnArchitecture

[frMitola]

Hello, I was following the tutorials and using "clf_cl_n" (ConvnLSTMnArchitecture) of the model zoo.
I noticed that while using this model, sometimes the loss just explode suddenly to very high numbers.
I analyzed the code of ConvnLSTMnArchitecture in tsgm/models/architectures/zoo.py and I noticed that LSTM layer has as activation function "relu", while usually it requests a saturating non-linearity, as it is a recurrent layer (otherwise with linear activation of the ReLU it can explode!).
I modified locally the implementation, simply deleting the "relu" in LSTM and seems to work fine after some trials!