👍

insuranceqa.py

@@ -184,7 +184,7 @@ test_model.load_weights(os.path.join(models_path, 'iqa_model_for_prediction.h5')
 
 # found through experimentation that ~24 epochs generalized the best
 print('Fitting model')
-for i in range(30):
+for i in range(10000):
     print('----- %d -----' % i)
     np.random.shuffle(ab_data)
     train_model.fit([q_data, ag_data, ab_data], targets, nb_epoch=1, batch_size=128, validation_data=[[qv_data, avg_data, avb_data], v_targets], shuffle=True)

keras_attention_model.py

@@ -82,7 +82,7 @@ def make_model(maxlen_question, maxlen_answer, n_words, n_lstm_dims=141, n_embed
 
     # merge together
     # note: `cos` refers to "cosine similarity", i.e. similar vectors should go to 1
-    # for training's sake, "abs" limits range to be tween 0 and 1 (binary classification)
+    # for training's sake, "abs" limits range to between 0 and 1 (binary classification)
     good_out = merge([q_out, ag_out], name='good', mode='cos', dot_axes=1)
     bad_out = merge([q_out, ab_out], name='bad', mode='cos', dot_axes=1)