log_file

21st March, 2017

Using Keras with Tensorflow as backend.
2 LSTM layers with 256 nodes in each layer. 
Dropout used : 0.7

Combinations of optimizers and loss functions used : 
1. Adam, Categorical cross entropy
2. RMSprop, Mean Squared Error
3. SGD, Mean Squared Error

Training info : 
1. Initial training was done using Adam optimizer and Categorical cross entropy loss function. 
   Number of epochs = 400, batch size = 128
   Loss value after 400 epochs = 0.956543
   Weight file : weights-improvement=393-0.956543.hdf5 (File added to the repo)
   Accuracy achieved on testing : 80.9698%
   
2. More training was done using pre-trained weight file. RMSprop and SGD optimizers were used along with MSE loss function. 
   Error further reduced from 0.950654 to 0.023081.
   Number of epochs = 10, batch size = 128
   Loss value after 10 epochs = 0.023081
   Weight file : mt-weights-improvement=10-0.023081.hdf5 (File added to the repo)
   Accuracy achieved on testing : 81.58%