alphabet = "ABCDEFGHIJKLMNOPQRSTUVWXYZ"
Credit: Deep Learning with Python by Jason Brownlee
1. alphabet_LSTM_1to1 for Learning One-Char to One-Char Mapping, 1 to 1
500 epochs, took a while to train
Model Accuracy: 84.00%
2. alphabet_LSTM_3to1 for 3 to One-Char Mapping, 3 to 1
Fast
Model Accuracy: 82.61%
3. alphabet_LSTM_3stepsto1 for a time steps to One-Char Mapping, 3 to 1
Same as the above model, but the input data was shaped differently, so that the previous information is modeled as time steps, instead of window features
Model Accuracy: 95.65%
4. alphabet_LSTM_batchto1 with a larger batch
Same as 1to1, but a larger batch that is the same size as the input.
Fast!
100.00%
5. alphabet_LSTM_statebatchto1 with state! (How LSTM is meant to be used)
Very fast!
Model Accuracy: 100.00%
6. alphabet_LSTM_varto1 stateless variable length to One-Char Mapping
Very Very slow.
Model Accuracy: 98.40%. Maybe it learned something!
Conclusion: LSTM with state has the best performance and speed, but it may not have learned the generic information. Stateless LSTM performs the best without knowing the whole sequence.
