1. add comments

2. add the plots of training result

Author: SunnerLi

img/cifar10_record.png

@@ -5,22 +5,52 @@
 import tensorflow as tf
 import numpy as np
 
-epoches = 200       # Epoches
-M = 5               # Monte-Carlo liked scalar
-loss = {            # Loss
-    'cnn': [],
-    'resnet': [],
-    'rir': []
-}           
+epoches = 200               # Epoches
+M = 5                       # Monte-Carlo liked scalar
+record = {
+    'loss': {               # Loss
+        'cnn': [],
+        'resnet': [],
+        'rir': []
+    },
+    'acc': {                # Accuracy
+        'cnn': [],
+        'resnet': [],
+        'rir': []
+    }
+}
+
+def recordTrainResult(cnn_loss, res_net_loss, rir_loss, cnn_acc, res_net_acc, rir_acc):
+    """
+        Append the training result into the corresponding list
+
+        Arg:    cnn_loss        - The loss value of usual CNN
+                res_net_loss    - The loss value of ResNet
+                rir_loss        - The loss value of ResNet in ResNet Network
+                cnn_acc         - The accuracy value of usual CNN
+                res_net_acc     - The accuracy value of ResNet
+                rir_acc         - The accuracy value of ResNet in ResNet Network
+                
+    """
+    global record
+    record['loss']['cnn'].append(cnn_loss)
+    record['loss']['resnet'].append(res_net_loss)
+    record['loss']['rir'].append(rir_loss) 
+    record['acc']['cnn'].append(cnn_acc)
+    record['acc']['resnet'].append(res_net_acc)
+    record['acc']['rir'].append(rir_acc)
 
 if __name__ == '__main__':
     # Load data
-    train_x, train_y, eval_x, eval_y, test_x, test_y = tl.files.load_mnist_dataset(shape=(-1, 28, 28, 1))
-    train_x -= 0.5
+    #train_x, train_y, eval_x, eval_y, test_x, test_y = tl.files.load_mnist_dataset(shape=(-1, 28, 28, 1))
+    train_x, train_y, test_x, test_y = tl.files.load_cifar10_dataset()
+    #train_x -= 0.5
+    train_x = (train_x - 127.5) / 127.5
     train_y = to_categorical(train_y)
+    print('max: ', np.max(train_x))
 
     # Construct the network
-    imgs_ph = tf.placeholder(tf.float32, [None, 28, 28, 1])
+    imgs_ph = tf.placeholder(tf.float32, [None, 32, 32, 3])
     tags_ph = tf.placeholder(tf.float32, [None, 10])
     usual_cnn = CNN(imgs_ph, tags_ph)
     res_net = ResNet(imgs_ph, tags_ph)
@@ -43,10 +73,8 @@
                 if i % 10 == 0:
                     print('iter: ', i, '\tCNN loss: ', _cnn_loss, '\tacc: ', _cnn_acc, '\tResNet loss: ', _res_net_loss, \
                         '\tacc: ', _res_net_acc, '\tRiR loss: ', _rir_loss, '\tacc: ', _rir_acc)
-                    loss['cnn'].append(_cnn_loss)
-                    loss['resnet'].append(_res_net_loss)
-                    loss['rir'].append(_rir_loss)
+                    recordTrainResult(_cnn_loss, _res_net_loss, _rir_loss, _cnn_acc, _res_net_acc, _rir_acc)
 
     # Visualize
-    loss = meanError(loss, M)
-    draw(loss)
+    record = meanError(record, M)
+    draw(record)

img/mnist_record.png

@@ -2,21 +2,46 @@
 import numpy as np
 
 def draw(loss_dict):
-    curve_line_1, = plt.plot(loss_dict['cnn'], 'k', color='g', label='CNN')
-    curve_point_1, = plt.plot(loss_dict['cnn'], 'bo', color='g')
-    curve_line_2, = plt.plot(loss_dict['resnet'], 'k', color='b', label='ResNet')
-    curve_point_2, = plt.plot(loss_dict['resnet'], 'bo', color='b')
-    curve_line_3, = plt.plot(loss_dict['rir'], 'k', color='orange', label='RiR')
-    curve_point_3, = plt.plot(loss_dict['rir'], 'bo', color='orange')
+    """
+        Draw the line plot with points
+        (You should call meanError to get the mean list before calling this function)
+
+        Arg:    loss_dict   - The dict object which contain training mean result
+    """
+    plt.figure(1)
+
+    # Plot loss
+    plt.subplot(211)
+    curve_line_1, = plt.plot(loss_dict['loss']['cnn'], 'k', color='g', label='CNN')
+    curve_point_1, = plt.plot(loss_dict['loss']['cnn'], 'bo', color='g')
+    curve_line_2, = plt.plot(loss_dict['loss']['resnet'], 'k', color='b', label='ResNet')
+    curve_point_2, = plt.plot(loss_dict['loss']['resnet'], 'bo', color='b')
+    curve_line_3, = plt.plot(loss_dict['loss']['rir'], 'k', color='orange', label='RiR')
+    curve_point_3, = plt.plot(loss_dict['loss']['rir'], 'bo', color='orange')
     plt.legend(handles=[curve_line_1, curve_line_2, curve_line_3])
+    plt.title('Loss')
+
+    # Plot accuracy
+    plt.subplot(212)
+    curve_line_1, = plt.plot(loss_dict['acc']['cnn'], 'k', color='g', label='CNN')
+    curve_point_1, = plt.plot(loss_dict['acc']['cnn'], 'bo', color='g')
+    curve_line_2, = plt.plot(loss_dict['acc']['resnet'], 'k', color='b', label='ResNet')
+    curve_point_2, = plt.plot(loss_dict['acc']['resnet'], 'bo', color='b')
+    curve_line_3, = plt.plot(loss_dict['acc']['rir'], 'k', color='orange', label='RiR')
+    curve_point_3, = plt.plot(loss_dict['acc']['rir'], 'bo', color='orange')
+    plt.legend(handles=[curve_line_1, curve_line_2, curve_line_3])
+    plt.title('Accuracy')
+
+    # Show
+    plt.savefig('record.png')                                                 # save before show to avoid refreshing
     plt.show()
-    plt.savefig('loss.png')
+    
 
 def meanError(loss_dict, scalar):
-    for key in loss_dict.keys():
-        origin_list = np.reshape(loss_dict[key], [scalar, -1])
-        loss_dict[key] = np.asarray(np.mean(origin_list, axis=0))
-        # print(loss_dict[key])
+    for record_type in loss_dict.keys():
+        for net_type in loss_dict[record_type].keys():
+            origin_list = np.reshape(loss_dict[record_type][net_type], [scalar, -1])
+            loss_dict[record_type][net_type] = np.asarray(np.mean(origin_list, axis=0))
     return loss_dict
 
 if __name__ == '__main__':