Filters: create 2 branches of them. Use functional style in Keras

Author: ldacosta

keras_cnn.py

@@ -1,10 +1,10 @@
 import numpy as np
 import keras
 from keras.models import Sequential
-from keras.layers import Dense, Dropout, Flatten
+from keras.layers import Dense, Dropout, Flatten, Merge
 from keras.layers import Conv2D, MaxPooling2D
 from keras.optimizers import SGD
-import tensorflow as tf
+from keras.layers import merge, add
 
 
 class SentimentAnalysisAnalyzer(object):
@@ -20,32 +20,59 @@ def fit(self, x_train, y_train, batch_size, epochs):
     def evaluate(self, x, y, batch_size):
         raise NotImplementedError("'evaluate' has to be implemented")
 
+
+from keras.layers import Input, Dense, Conv2D, MaxPooling2D, Dropout
+from keras.models import Model
+
 class WithKeras(SentimentAnalysisAnalyzer):
 
     def __init__(self, features_in_words, words_in_review):
 
         SentimentAnalysisAnalyzer.__init__(self, features_in_words, words_in_review)
-
-        # # n_features_in_word = mw.model.vector_size
-        # n_features_in_word = 300 # TODO: replace this by previous line
-        # n_words_in_review = 10 # number of words to take from each review
-
-        self.model = Sequential()
-        # input: 100x100 images with 3 channels -> (100, 100, 3) tensors.
-        # this applies 1 convolution filter(s) of size 3x3 each.
-        filter_depth = 5
-        self.model.add(Conv2D(filter_depth, (3,  self.n_features_in_word), activation='relu', input_shape=(self.n_words_in_review, self.n_features_in_word, self.input_channels), strides=(1, 1), padding='valid'))
-        self.model.add(MaxPooling2D(pool_size=(2, 1), strides = (1,1)))
-        self.model.add(Dropout(0.25))
-
-        self.model.add(Flatten())
-        self.model.add(Dense(256, activation='relu'))
-        self.model.add(Dropout(0.5))
-        self.model.add(Dense(2, activation='softmax'))
-
+        # inputs
+        inputs = Input(shape=(self.n_words_in_review, self.n_features_in_word, self.input_channels))
+        # intermediate results
+        # 2D 3x3 convolution followed by a maxpool
+        branch1 =Conv2D(
+                filters=5,
+                kernel_size=(3, self.n_features_in_word),
+                activation='relu',
+                input_shape=(self.n_words_in_review, self.n_features_in_word, self.input_channels),
+                strides=(1, 1),
+                padding='valid')(inputs)
+        branch1 = MaxPooling2D(pool_size=(2, 1), strides = (1,1))(branch1)
+        branch1 = Dropout(0.25)(branch1)
+
+        # 2D 4x4 convolution followed by a maxpool
+        branch2 =Conv2D(
+                filters=5,
+                kernel_size=(4, self.n_features_in_word),
+                activation='relu',
+                input_shape=(self.n_words_in_review, self.n_features_in_word, self.input_channels),
+                strides=(1, 1),
+                padding='valid')(inputs)
+        branch2 = MaxPooling2D(pool_size=(2, 1), strides = (1,1))(branch2)
+        branch2 = Dropout(0.25)(branch2)
+
+        # now let's go fully-connected to a 2-way classification:
+        merged_branches = keras.layers.concatenate([branch1, branch2], axis = 1) # , axis=-1)
+
+        # fully-connected
+        fully_connected = Flatten()(merged_branches)
+        fully_connected = Dense(256, activation='relu')(fully_connected)
+        fully_connected = Dropout(0.5)(fully_connected)
+
+        # ok. done
+        categorization = Dense(2, activation='softmax')(fully_connected)
+
+        # all good. Let's build the model, then
+        self.model = Model(inputs=inputs, outputs=categorization)
         sgd = SGD(lr=0.01, decay=1e-6, momentum=0.9, nesterov=True)
         self.model.compile(loss='categorical_crossentropy', metrics = ['accuracy'], optimizer=sgd)
 
+    def summary(self):
+        return self.model.summary()
+
     def fit(self, x_train, y_train, batch_size, epochs):
         self.model.fit(x_train, y_train, batch_size=batch_size, epochs=epochs)