Update learning rate for DS2.

In order to get better accuracy for the 3 layer GRU with 2560 hidden
units, the learning rate has been lowered from 0.00075 to 0.0001. This
allows the network to acheive a WER of around 34.

Further hyperparameter tuning needs to occur to get lower WER.

speech_recognition/pytorch/params.py

@@ -33,7 +33,7 @@
 # Training parameters
 epochs          = 10 # Number of training epochs
 learning_anneal = 1.1 # Annealing applied to learning rate every epoch
-lr              = 0.00075 # initial learning rate
+lr              = 0.0001 # initial learning rate
 momentum        = 0.9 # momentum
 max_norm        = 400 # Norm cutoff to prevent explosion of gradients
 l2              = 0 # L2 regularization