I tried to train sequential (LSTM and GRU) models to predict the chemical structure of drugs for drug discovery process. I got very large cross entropy loss (>100). The input to the model is SMILES string of chemical structures encoded as a one-hot encoded array.

I tried to change parameters like number of epochs, learning rate, played with train/test data split, etc., and also tried to add more layers to the model but the cross-entropy loss still doesn’t drop below 100. What do you think it is? I’m stuck guys