Make nea compatible with Keras 2.0.8

README.md

@@ -2,6 +2,14 @@
 
 An automatic essay scoring system based on convolutional and recurrent neural networks, including GRU and LSTM.
 
+## Environment
+
+* Keras 2.0.8
+* Theano 0.9.0
+* Numpy 1.13.3
+* Scipy 1.0.0
+* Python 2.7.9
+
 ### Set Up ###
 
 * Install Keras (with Theano backend)

nea/models.py

@@ -49,7 +49,7 @@ def create_model(args, initial_mean_value, overal_maxlen, vocab):
 		model.add(Dense(num_outputs))
 		if not args.skip_init_bias:
 			bias_value = (np.log(initial_mean_value) - np.log(1 - initial_mean_value)).astype(K.floatx())
-			model.layers[-1].b.set_value(bias_value)
+			model.layers[-1].bias = bias_value
 		model.add(Activation('sigmoid'))
 		model.emb_index = 0
 
@@ -70,10 +70,10 @@ def create_model(args, initial_mean_value, overal_maxlen, vocab):
 		model.add(Dense(num_outputs))
 		if not args.skip_init_bias:
 			bias_value = (np.log(initial_mean_value) - np.log(1 - initial_mean_value)).astype(K.floatx())
-			model.layers[-1].b.set_value(bias_value)
+			model.layers[-1].bias = bias_value
 		model.add(Activation('sigmoid'))
 		model.emb_index = 0
-	
+
 	elif args.model_type == 'breg':
 		logger.info('Building a BIDIRECTIONAL REGRESSION model')
 		from keras.layers import Dense, Dropout, Embedding, LSTM, Input, merge
@@ -130,7 +130,7 @@ def create_model(args, initial_mean_value, overal_maxlen, vocab):
 		from w2vEmbReader import W2VEmbReader as EmbReader
 		logger.info('Initializing lookup table')
 		emb_reader = EmbReader(args.emb_path, emb_dim=args.emb_dim)
-		model.layers[model.emb_index].W.set_value(emb_reader.get_emb_matrix_given_vocab(vocab, model.layers[model.emb_index].W.get_value()))
+		model.layers[model.emb_index].set_weights(emb_reader.get_emb_matrix_given_vocab(vocab, model.layers[model.emb_index].get_weights()))
 		logger.info('  Done')
 
 	return model

nea/my_layers.py

@@ -54,11 +54,12 @@ def __init__(self, mask_zero=True, **kwargs):
 
 	def call(self, x, mask=None):
 		if self.mask_zero:
-			return K.cast(x.sum(axis=1) / mask.sum(axis=1, keepdims=True), K.floatx())
+			mask = K.cast(mask, K.floatx())
+			return K.cast(K.sum(x, axis=1) / K.sum(mask, axis=1, keepdims= True), K.floatx())
 		else:
 			return K.mean(x, axis=1)
 
-	def get_output_shape_for(self, input_shape):
+	def compute_output_shape(self, input_shape):
 		return (input_shape[0], input_shape[2])
 
 	def compute_mask(self, x, mask):
@@ -73,6 +74,6 @@ class Conv1DWithMasking(Convolution1D):
 	def __init__(self, **kwargs):
 		self.supports_masking = True
 		super(Conv1DWithMasking, self).__init__(**kwargs)
-	
+
 	def compute_mask(self, x, mask):
 		return mask

nea/w2vEmbReader.py

@@ -28,9 +28,9 @@ def __init__(self, emb_path, emb_dim=None):
 				counter = 0
 				for line in emb_file:
 					tokens = line.split()
-					assert len(tokens) == self.emb_dim + 1, 'The number of dimensions does not match the header info'
 					word = tokens[0]
-					vec = tokens[1:]
+					vec = tokens[1].split(',')
+					assert len(vec) == self.emb_dim, 'The number of dimensions does not match the header info'
 					self.embeddings[word] = vec
 					counter += 1
 				assert counter == self.vocab_size, 'Vocab size does not match the header info'
@@ -41,13 +41,14 @@ def __init__(self, emb_path, emb_dim=None):
 				self.embeddings = {}
 				for line in emb_file:
 					tokens = line.split()
+					word = tokens[0]
+					vec = tokens[1].split(',')
 					if self.emb_dim == -1:
-						self.emb_dim = len(tokens) - 1
+						self.emb_dim = len(vec)
 						assert self.emb_dim == emb_dim, 'The embeddings dimension does not match with the requested dimension'
 					else:
-						assert len(tokens) == self.emb_dim + 1, 'The number of dimensions does not match the header info'
-					word = tokens[0]
-					vec = tokens[1:]
+						assert len(vec) == self.emb_dim, 'The number of dimensions does not match the header info'
+
 					self.embeddings[word] = vec
 					self.vocab_size += 1
 
@@ -63,7 +64,7 @@ def get_emb_matrix_given_vocab(self, vocab, emb_matrix):
 		counter = 0.
 		for word, index in vocab.iteritems():
 			try:
-				emb_matrix[index] = self.embeddings[word]
+				emb_matrix[0][index] = self.embeddings[word]
 				counter += 1
 			except KeyError:
 				pass

train_nea.py

@@ -1,5 +1,6 @@
 #!/usr/bin/env python
 
+import os
 import argparse
 import logging
 import numpy as np
@@ -9,6 +10,8 @@
 import nea.utils as U
 import pickle as pk
 
+
+os.environ['KERAS_BACKEND']='theano'
 logger = logging.getLogger(__name__)
 
 ###############################################################################################################################
@@ -164,9 +167,9 @@
 ## Plotting model
 #
 
-from keras.utils.visualize_util import plot
+from keras.utils.vis_utils import plot_model
 
-plot(model, to_file = out_dir + '/model.png')
+plot_model(model, to_file = out_dir + '/model.png')
 
 ###############################################################################################################################
 ## Save model architecture
@@ -194,10 +197,11 @@
 total_train_time = 0
 total_eval_time = 0
 
+
 for ii in range(args.epochs):
 	# Training
 	t0 = time()
-	train_history = model.fit(train_x, train_y, batch_size=args.batch_size, nb_epoch=1, verbose=0)
+	train_history = model.fit(train_x, train_y, batch_size=args.batch_size, epochs=1, verbose=0)
 	tr_time = time() - t0
 	total_train_time += tr_time