take away dropout

Author: Lazy Programmer

Diff for: nlp_class3/attention.py

@@ -202,7 +202,11 @@ def softmax_over_time(x):
 # Set up the encoder - simple!
 encoder_inputs_placeholder = Input(shape=(max_len_input,))
 x = embedding_layer(encoder_inputs_placeholder)
-encoder = Bidirectional(LSTM(LATENT_DIM, return_sequences=True, dropout=0.5))
+encoder = Bidirectional(LSTM(
+  LATENT_DIM,
+  return_sequences=True,
+  # dropout=0.5 # dropout not available on gpu
+))
 encoder_outputs = encoder(x)
 
 

Diff for: nlp_class3/wseq2seq.py

@@ -174,7 +174,11 @@
 ##### build the model #####
 encoder_inputs_placeholder = Input(shape=(max_len_input,))
 x = embedding_layer(encoder_inputs_placeholder)
-encoder = LSTM(LATENT_DIM, return_state=True, dropout=0.5)
+encoder = LSTM(
+  LATENT_DIM,
+  return_state=True,
+  # dropout=0.5 # dropout not available on gpu
+)
 encoder_outputs, h, c = encoder(x)
 # encoder_outputs, h = encoder(x) #gru
 
@@ -192,7 +196,12 @@
 
 # since the decoder is a "to-many" model we want to have
 # return_sequences=True
-decoder_lstm = LSTM(LATENT_DIM, return_sequences=True, return_state=True, dropout=0.5)
+decoder_lstm = LSTM(
+  LATENT_DIM,
+  return_sequences=True,
+  return_state=True,
+  # dropout=0.5 # dropout not available on gpu
+)
 decoder_outputs, _, _ = decoder_lstm(
   decoder_inputs_x,
   initial_state=encoder_states