Making character model more legible

and a few other minor changes
mmistele · Jun 8, 2018 · 14eb039 · 14eb039
1 parent 990cb47
commit 14eb039
Show file tree

Hide file tree

Showing 3 changed files with 19 additions and 11 deletions.
diff --git a/keras/models.py b/keras/models.py
@@ -1,7 +1,7 @@
 import numpy as np
 
 from keras.models import Model
-from keras.layers import Dense, Input, Dropout, LSTM, Activation, Masking, Bidirectional
+from keras.layers import Dense, Input, Dropout, LSTM, Activation, Masking, Bidirectional, Lambda
 
 from custom_layers import MeanPool
 
@@ -23,11 +23,17 @@ def Character_Model_1(input_shape):
 
     X = Masking(mask_value = 0., input_shape=input_shape)(sentences)
 
-    X = Bidirectional(LSTM(128, return_sequences = False, dropout=0.2, recurrent_dropout=0.2), merge_mode='ave')(X)
+    # X = Bidirectional(LSTM(128, return_sequences = False, dropout=0.2, recurrent_dropout=0.2), merge_mode='ave')(X)
+    X = LSTM(128, return_sequences = True, dropout=0.2, recurrent_dropout=0.2, name='LSTM1')(X)
+    def get_last(X):
+        return X[:,-1,:]
+    X = Lambda(get_last, name = 'LSTM-last')(X)
+
     # X = Bidirectional(LSTM(128, return_sequences = True, dropout=0.2, recurrent_dropout=0.2))(X)
     # X = LSTM(128, dropout=0.2, recurrent_dropout=0.2)(X)
-    X = Dense(1)(X)
-    X = Activation('sigmoid')(X)
+
+    X = Dense(1, name='Dense1')(X)
+    X = Activation('sigmoid', name='output_layer')(X)
 
     model = Model(inputs = sentences, outputs = X)
     return model

diff --git a/keras/train_character_model.py b/keras/train_character_model.py
@@ -19,6 +19,7 @@
 X_dev, Y_dev = read_csv('data/dev.csv') 
 
 maxLen = max(len(max(X_train, key=len)), len(max(X_dev, key=len)))
+print "Max length: %s" % maxLen
 
 counts = get_char_counts_from_csv('data/train.csv') + get_char_counts_from_csv('data/dev.csv') 
 most_common = counts.most_common()
@@ -34,9 +35,9 @@
 model = Character_Model_1((None, X_train_indices.shape[2]))
 
 optimizer = Adam(lr=0.001, beta_1=0.9, beta_2=0.999, decay=0.0, epsilon=None)
-model.compile(loss='binary_crossentropy', optimizer=optimizer, metrics=['accuracy', 'loss'])
+model.compile(loss='binary_crossentropy', optimizer=optimizer, metrics=['accuracy'])
 tensorboard = TensorBoard(log_dir="logs/character-model/{}".format(time()))
-model.fit(X_train_indices, Y_train, epochs = 1, batch_size = 6, shuffle=True, callbacks = [tensorboard], validation_split = 0.2)
+model.fit(X_train_indices, Y_train, epochs = 20, batch_size = 20, shuffle=True, callbacks = [tensorboard], validation_split = 0.2)
 
 X_dev_indices = strings_to_character_vecs(X_dev, char_to_index, maxLen, alphabet_size)
 loss, acc = model.evaluate(X_dev_indices, Y_dev)

diff --git a/keras/train_word_model.py b/keras/train_word_model.py
@@ -13,9 +13,10 @@
 from utils import read_csv, read_glove_vecs
 
 X_train, Y_train = read_csv('data/train.csv') 
-X_test, Y_test = read_csv('data/dev.csv') 
+X_dev, Y_dev = read_csv('data/dev.csv') 
 
-maxLen = len(max(X_train, key=len).split())
+maxLen = max(len(max(X_train, key=len).split()), len(max(X_dev, key=len).split()))
+print "Max length: %s" % maxLen
 
 word_to_index, index_to_word, word_to_vec_map = read_glove_vecs('data/glove.6B.50d.txt')
 embedding_layer = pretrained_embedding_layer(word_to_vec_map, word_to_index)
@@ -28,8 +29,8 @@
 tensorboard = TensorBoard(log_dir="logs/word-model/{}".format(time()))
 model.fit(X_train_indices, Y_train, epochs = 20, batch_size = 6, shuffle=True, callbacks = [tensorboard])
 
-X_test_indices = strings_to_word_indices(X_test, word_to_index, max_len = maxLen)
-loss, acc = model.evaluate(X_test_indices, Y_test)
+X_dev_indices = strings_to_word_indices(X_dev, word_to_index, max_len = maxLen)
+loss, acc = model.evaluate(X_dev_indices, Y_dev)
 model.save('word-model.h5')
 print()
-print("Test accuracy = ", acc)
+print("Dev accuracy = ", acc)