LSTMs learning alphabets in different ways
-
one_char to one_char -- Naive LSTM learning one char to one char with no dependencies and achieved an accuracy of 100%.
-
one_char to one_char state -- An LSTM learning alphabets sequence with state maintained within a batch and achieved an accuracy of 100%.
-
feature window to one_char -- An LSTM learning alphabets with a feature size and 1 time step and achieved an accuracy of 86.96%.
-
time steps window to one_char -- An LSTM learning alphabets with some time steps and 1 feature step and achieved an accuracy of 100%.
-
Var length input to one char -- An LSTM learning alphabet sequences with variable length and outputs one character and achieved an accuracy of 97.50%. It takes a long time to train on CPU. The model weights ("model500epochs.h5") and the model structure("model.json") are also uploaded.