Adding a CNN-based classifier for the task of entries/other classification

classification/.gitignore

@@ -0,0 +1 @@
+*.pyc

classification/README.md

@@ -0,0 +1,12 @@
+## Classification of images into entry/other
+
+### Proposed Technique
+* Convert all images into Black&White.
+* Downsize all images into (150, 250)
+* Define a simple CNN-classifier and train it on the given data
+* Batch-normalization is used to handle the variance in given data, while automatic class-weights are used to balance the error function (as the class distribution is biased)
+* To account for the low amount of data given, a small learning rate is used (to avoid overfitting)
+
+### Running it
+* Run `python trainClassifier.py <images_folder> <label_file>` from the current directory to train an end-to-end model.
+* For example, run `python trainClassifier.py images/freecen/ data/gold/combined_classifications_20180227.csv`

classification/model.py

@@ -0,0 +1,29 @@
+import keras
+from keras.models import Sequential
+from keras.layers import Dense, Dropout, Flatten, Activation
+from keras.layers import Conv2D, MaxPooling2D, BatchNormalization
+
+# Define a simple CNN model
+def getSimpleCNN(input_shape, num_classes):
+	model = Sequential()
+	model.add(Conv2D(16, kernel_size=(3, 3), input_shape=input_shape))
+	model.add(BatchNormalization())
+	model.add(Activation('relu'))	
+	model.add(Conv2D(32, (3, 3)))
+	model.add(BatchNormalization())
+        model.add(Activation('relu'))
+	model.add(MaxPooling2D(pool_size=(2, 2)))
+	model.add(Dropout(0.25))
+	model.add(Flatten())
+	model.add(Dense(64))
+	model.add(BatchNormalization())
+        model.add(Activation('relu'))
+	model.add(Dropout(0.5))
+	model.add(Dense(num_classes, activation='softmax'))
+
+	model.compile(loss=keras.losses.categorical_crossentropy,
+        	      optimizer=keras.optimizers.Adadelta(lr=0.1),
+	              metrics=['accuracy'])
+
+	return model
+

classification/readData.py

@@ -0,0 +1,35 @@
+import numpy as np 
+from PIL import Image
+import os
+from tqdm import tqdm
+from scipy.misc import imresize
+import csv
+
+# Read label classification file, construct data
+def getData(imageDirPrefix, filePath):
+	X = []
+	Y = []
+	with open(filePath, 'r') as f:
+		reader = csv.reader(f)
+		for line in tqdm(reader):
+			filePath = line[0]
+			imgClass = line[1]
+			# Read image as a black&white image
+			image = np.asarray(Image.open(os.path.join(imageDirPrefix, filePath)).convert('L'))
+			# Resize into a smaller image
+			image = imresize(image, (150, 250))
+			X.append(image)
+			Y.append(imgClass)
+	X = np.array(X)
+	X = X.reshape(X.shape + (1,))
+	# Also store the mapping between class-names and indices
+	mappingDict = dict([(y,x) for x,y in enumerate(sorted(set(Y)))])
+	Y = np.array([ mappingDict[x] for x in Y])
+	return X, Y, mappingDict
+
+
+if __name__ == "__main__":
+	import sys
+	X, Y, mapping = getData(sys.argv[1], sys.argv[2])
+	print X.shape, Y.shape
+

classification/trainClassifier.py

@@ -0,0 +1,19 @@
+import readData
+import model
+import keras
+
+
+if __name__ == "__main__":
+	import sys
+	# Load data
+	X, Y, mapping = readData.getData(sys.argv[1], sys.argv[2])
+	num_classes = len(mapping.keys())
+	input_shape = X.shape[1:]
+	# Loada simple CNN for tha classification task
+	model = model.getSimpleCNN(input_shape, num_classes)
+	Y = keras.utils.to_categorical(Y, num_classes)
+	batch_size = 8
+	epochs = 20
+	# Train our model on the available data
+	model.fit(X, Y, batch_size=batch_size, epochs=epochs, validation_split=0.2, class_weight='auto')
+