Neural net predictor for Googlenet labels

83% F1 score on test set!

net2.py

@@ -0,0 +1,73 @@
+from __future__ import print_function
+import numpy as np
+import sys
+np.random.seed(1337)  # for reproducibility
+
+from keras import models
+from keras.models import Sequential
+from keras.layers.core import Dense, Dropout, Activation, Flatten, Reshape
+from keras.layers.convolutional import Convolution2D, MaxPooling2D
+from keras.preprocessing.image import ImageDataGenerator
+from keras.utils import np_utils
+import json
+import score
+import read_data
+
+if __name__ == '__main__':
+    dropout = 0.5
+    batch_size = 64
+    nb_classes = 9
+    nb_epoch = 200
+
+    model = Sequential()
+
+    model.add(Flatten(input_shape=(3,1000)))
+
+    model.add(Dropout(dropout))
+    model.add(Dense(1024))
+    model.add(Activation('relu'))
+
+    model.add(Dropout(dropout))
+    model.add(Dense(512))
+    model.add(Activation('relu'))
+
+    model.add(Dense(nb_classes))
+
+    model.compile(loss='hinge', optimizer='adadelta')
+
+    # image_labels: shape (num_image, 3, 1000)
+    image_labels = np.load('data/googlenet_predictions.npy')
+    with open('data/googlenet_predictions_order.json', 'r') as jfile:
+        image_label_order = json.load(jfile)
+    biz_csv = read_data.read_biz_csv()
+
+    num_biz = len(biz_csv)
+    x_whole = np.zeros((num_biz, image_labels.shape[1], image_labels.shape[2])) 
+    y_whole = np.zeros((num_biz, 9))
+    # aggregate data by business: take average of predictions
+    for idx, (biz_id, (label, photo_ids)) in enumerate(biz_csv.items()):
+        for photo_id in photo_ids:
+            image_idx = image_label_order[str(photo_id)]
+            x_whole[idx] += image_labels[image_idx]
+        x_whole[idx] /= len(photo_ids)
+        y_whole[idx] = label
+
+    train_frac = 0.9
+    x_train, x_test = np.vsplit(x_whole, [int(num_biz*train_frac)])
+    y_train, y_test = np.vsplit(y_whole, [int(num_biz*train_frac)])
+
+    print('X_train shape:', x_train.shape)
+    print('Y_train shape:', y_train.shape)
+    print('Training on %s biz, testing on %s biz' % \
+            (x_train.shape[0], x_test.shape[0]))
+
+    model.fit(x_train, y_train, batch_size=batch_size,
+            nb_epoch=nb_epoch, verbose=1, validation_split=0.1)
+
+    test_pred = np.sign(model.predict(x_test))
+    test_loss = model.evaluate(x_test, y_test)
+    np.savetxt('pred.csv', test_pred, delimiter=',')
+
+    print('Test loss: ', test_loss)
+    print('Test accuracy: ', score.accuracy(test_pred, y_test))
+    print('F1 score: ', score.f1score(test_pred, y_test))

score.py

@@ -0,0 +1,14 @@
+import numpy as np
+
+def accuracy(y_pred, y_test):
+    total = y_test.size
+    correct = np.sum(y_pred == y_test)
+    return correct/total
+
+def f1score(y_pred, y_test):
+    tp = np.sum(np.logical_and(y_pred == 1, y_test == 1))
+    fp = np.sum(np.logical_and(y_pred == 1, y_test == -1))
+    fn = np.sum(np.logical_and(y_pred == -1, y_test == 1))
+    precision = tp/(tp + fp)
+    recall = tp/(tp + fn)
+    return 2*precision*recall/(precision + recall)