cifar preprocess

Author: anantzoid

main.py

@@ -75,7 +75,10 @@ def train(conf, data):
         from keras.datasets import cifar10
         data = cifar10.load_data()
         labels = data[0][1]
-        data = data[0][0] / 255.0
+        data = data[0][0].astype(np.float32)
+        data[:,0,:,:] -= np.mean(data[:,0,:,:])
+        data[:,1,:,:] -= np.mean(data[:,1,:,:])
+        data[:,2,:,:] -= np.mean(data[:,2,:,:])
         data = np.transpose(data, (0, 2, 3, 1))
         conf.img_height = 32
         conf.img_width = 32

models.py

@@ -64,8 +64,8 @@ def __init__(self, X, conf, h=None):
                 the 2 methods which may be used, with the first one (self.pred)
                 being more likely.
             '''
-            self.pred = tf.reshape(tf.multinomial(tf.nn.softmax(self.fc2), num_samples=1, seed=100), tf.shape(self.X))
-            self.pred_argmax = tf.reshape(tf.argmax(tf.nn.softmax(self.fc2), dimension=tf.rank(self.fc2) - 1), tf.shape(self.X))
+            self.pred_sampling = tf.reshape(tf.multinomial(tf.nn.softmax(self.fc2), num_samples=1, seed=100), tf.shape(self.X))
+            self.pred = tf.reshape(tf.argmax(tf.nn.softmax(self.fc2), dimension=tf.rank(self.fc2) - 1), tf.shape(self.X))
 
 
 class ConvolutionalEncoder(object):