@mr_cell, Hi.
I’ve read your code and looked at your snapshots. Unfortunately I can’t read the langauge in which the snapshots labels were written but here are some pointers that work for me in CNN-LSTM Models.
-Increase your hidden size to something a little larger, like 128.
-Decrease your number of Epochs <<-Definitely!.
-Train your model on a larger dataset if you have one. If not, find data that is almost similar, something with the same input shape, and is a related task to what you want to achieve. Train your model on that data then use transfer learnining to finetune it to your original data. This will allow the model to learn the complex stuff from the other data then just figure out how to work with your new data. Basically it gives you better results from small data.
- Tell me how it goes. Kuddos.