In [1]:
    
import matplotlib.pyplot as plt
import numpy as np
import tensorflow as tf
import os
import sys
from six.moves import cPickle as pickle
%matplotlib inline
    
    
In [2]:
    
#pickle_file = 'train.pickle'
'''
with open(pickle_file, 'rb') as f:
    save = pickle.load(f)
    train_X_1 = save['data']
    train_outcome_1 = save['outcome']
    del save  # hint to help gc free up memory
'''
    
pickle_file = 'train2.pickle'
with open(pickle_file, 'rb') as f:
    save = pickle.load(f)
    train_X_0 = save['data']
    train_outcome_0 = save['outcome']
    del save  # hint to help gc free up memory
    
'''
pickle_file = 'test.pickle'
with open(pickle_file, 'rb') as f:
    save = pickle.load(f)
    test_X_1 = save['data']
    test_outcome_1 = save['outcome']
    del save  # hint to help gc free up memory
    
'''
pickle_file = 'test2.pickle'
with open(pickle_file, 'rb') as f:
    save = pickle.load(f)
    test_X_0 = save['data']
    test_outcome_0 = save['outcome']
    del save  # hint to help gc free up memory
    
In [3]:
    
#reformat the label
#for each digit, add a 'end_digit' as '10'
#for each label, add a digit size
#each of them is a one-hot coding
def label_reformat(label, max_size = 5):
    digit_size = np.asarray([len(x) for x in label])
    digit_size[digit_size > max_size]= max_size
    digit_size = ((np.arange(max_size)+1) == digit_size[:,None]).astype(np.float32)
    
    digits = {}
    end_digit = 10.0
    for i in range(max_size):
        digit_coding = np.asarray( [x[i] if len(x)>i else end_digit for x in label])
        digit_coding = (np.arange(end_digit+1) == digit_coding[:,None]).astype(np.float32)
        digits['digit_'+ str(i)] = digit_coding
        
    return digit_size, digits
    
In [4]:
    
#train_X_0 = np.vstack((train_X_1 ,train_X_2 ))
    
In [5]:
    
train_X_0.shape
    
    Out[5]:
In [6]:
    
train_label = train_outcome_0['label']
for i in range(8):
    pickle_file = 'extra_crop_' + str(i) + '.pickle'
    
    with open(pickle_file, 'rb') as f:
        save = pickle.load(f)
        extra_X_tmp = save['data']
        extra_outcome_tmp = save['outcome']
        del save  # hint to help gc free up memory
        
        train_X_0 = np.vstack((train_X_0 ,extra_X_tmp ))
        train_label = train_label + extra_outcome_tmp['label']
        print '{}th extra incorperated into training data'.format(i)
    
    
In [7]:
    
#train_X_0 = np.vstack((train_X_1 ,train_X_2 ))
image_size = train_X_0.shape[1]
num_channels = train_X_0.shape[3]
batch_size = 200
val_size = 50
test_size = 50
reg = 0.001
#train_label = train_outcome_1['label'] +  train_outcome_2['label']
#train_label = train_outcome_0['label']
train_digit_size, train_digits = label_reformat(train_label)
train_X = train_X_0
val_label = test_outcome_0['label']
val_digit_size, val_digits = label_reformat(val_label)
val_X = test_X_0
val_size = val_X.shape[0]
    
In [8]:
    
print train_digit_size.shape
print train_digits['digit_0'].shape
print train_X.shape
    
    
In [9]:
    
plt.imshow(train_X[0,:,:,:])
plt.show()
print train_digits['digit_0'][0]
print train_digits['digit_1'][0]
    
    
    
In [10]:
    
plt.imshow(val_X[1,:,:,:])
plt.show()
print val_digits['digit_0'][1]
print val_digits['digit_1'][1]
    
    
    
In [11]:
    
def next_batch(X, y_dsize, y_ds, batch_size=50, replace = True):
    idx = np.random.choice(X.shape[0],batch_size, replace = replace)
    batch_x = X[idx,:,:,:]
    batch_y_dsize = y_dsize[idx,:]
    batch_y_d1 = y_ds['digit_0'][idx,:]
    batch_y_d2 = y_ds['digit_1'][idx,:]
    batch_y_d3 = y_ds['digit_2'][idx,:]
    batch_y_d4 = y_ds['digit_3'][idx,:]
    batch_y_d5 = y_ds['digit_4'][idx,:]
    
    return batch_x, batch_y_dsize, batch_y_d1, batch_y_d2, batch_y_d3, batch_y_d4, batch_y_d5
    
In [12]:
    
def weight_variable(shape):
    initial = tf.truncated_normal(shape, stddev=0.1)
    return tf.Variable(initial)
def bias_variable(shape):
    initial = tf.constant(0.1, shape=shape)
    return tf.Variable(initial)
def conv2d(x, W):
    return tf.nn.conv2d(x, W, strides=[1, 1, 1, 1], padding='SAME')
def max_pool_2x2(x):
    return tf.nn.max_pool(x, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')
def max_pool_2x2_same(x):
    return tf.nn.max_pool(x, ksize=[1, 2, 2, 1], strides=[1, 1, 1, 1], padding='SAME')
    
In [13]:
    
# this is a simpler version of Tensorflow's 'official' version. See:
# https://github.com/tensorflow/tensorflow/blob/master/tensorflow/contrib/layers/python/layers/layers.py#L102
# http://r2rt.com/implementing-batch-normalization-in-tensorflow.html
def batch_norm_wrapper(inputs, is_training, decay = 0.999, epsilon = 1e-3):
    scale = tf.Variable( tf.ones(inputs.get_shape()[-1]) )
    beta = tf.Variable(tf.zeros(inputs.get_shape()[-1]))
    pop_mean = tf.Variable(tf.zeros(inputs.get_shape()[-1]), trainable=False)
    pop_var = tf.Variable(tf.ones(inputs.get_shape()[-1]), trainable=False)
    if is_training:
        #for conv layer, use global normalization
        batch_mean, batch_var = tf.nn.moments(inputs,[0,1,2])
        #print pop_mean.get_shape()
        train_mean = tf.assign(pop_mean, pop_mean * decay + batch_mean * (1 - decay))
        train_var = tf.assign(pop_var, pop_var * decay + batch_var * (1 - decay))
        with tf.control_dependencies([train_mean, train_var]):
            return tf.nn.batch_normalization(inputs, batch_mean, batch_var, beta, scale, epsilon)
    else:
        return tf.nn.batch_normalization(inputs, pop_mean, pop_var, beta, scale, epsilon)
    
    
def batch_norm_wrapper_simple(inputs, is_training, decay = 0.999, epsilon = 1e-3):
    scale = tf.Variable( tf.ones(inputs.get_shape()[-1]) )
    beta = tf.Variable(tf.zeros(inputs.get_shape()[-1]))
    pop_mean = tf.Variable(tf.zeros(inputs.get_shape()[-1]), trainable=False)
    pop_var = tf.Variable(tf.ones(inputs.get_shape()[-1]), trainable=False)
    if is_training:
        #for conv layer, use global normalization
        batch_mean, batch_var = tf.nn.moments(inputs,[0])
        #print pop_mean.get_shape()
        train_mean = tf.assign(pop_mean, pop_mean * decay + batch_mean * (1 - decay))
        train_var = tf.assign(pop_var, pop_var * decay + batch_var * (1 - decay))
        with tf.control_dependencies([train_mean, train_var]):
            return tf.nn.batch_normalization(inputs, batch_mean, batch_var, beta, scale, epsilon)
    else:
        return tf.nn.batch_normalization(inputs, pop_mean, pop_var, beta, scale, epsilon)
    
In [14]:
    
def build_graph(is_training):
    
    # shape = [None, ...], the None element of the shape corresponds to a variable-sized dimension.
    x_image = tf.placeholder(tf.float32, shape=[None, image_size, image_size, num_channels])
    
    y_d1 = tf.placeholder(tf.float32, shape=[None, 11])
    y_d2 = tf.placeholder(tf.float32, shape=[None, 11])
    y_d3 = tf.placeholder(tf.float32, shape=[None, 11])
    y_d4 = tf.placeholder(tf.float32, shape=[None, 11])
    y_d5 = tf.placeholder(tf.float32, shape=[None, 11])
    
    y_dsize = tf.placeholder(tf.float32, shape=[None, 5])
    
    #first layer
    W_conv1 = weight_variable([5, 5, num_channels, 32])
    b_conv1 = bias_variable([32])
    
    z_conv1 = conv2d(x_image, W_conv1) + b_conv1
    z_conv1_BN = batch_norm_wrapper(z_conv1, is_training)
    h_conv1 = tf.nn.relu(z_conv1_BN)
    
    h_pool1 = max_pool_2x2(h_conv1)
    
    #second layer
    W_conv2 = weight_variable([5, 5, 32, 64])
    b_conv2 = bias_variable([64])
    
    z_conv2 = conv2d(h_pool1, W_conv2) + b_conv2
    z_conv2_BN = batch_norm_wrapper(z_conv2, is_training)
    h_conv2 = tf.nn.relu( z_conv2_BN )
    
    h_pool2 = max_pool_2x2_same(h_conv2)
    
    #third layer
    W_conv3 = weight_variable([5, 5, 64, 128])
    b_conv3 = bias_variable([128])
    z_conv3 = conv2d(h_pool2, W_conv3) + b_conv3
    z_conv3_BN = batch_norm_wrapper(z_conv3, is_training)
    h_conv3 = tf.nn.relu( z_conv3_BN )
    h_pool3 = max_pool_2x2(h_conv3)
    
    #fourth layer    
    W_conv4 = weight_variable([5, 5, 128, 160])
    b_conv4 = bias_variable([160])
    z_conv4 = conv2d(h_pool3, W_conv4) + b_conv4
    z_conv4_BN = batch_norm_wrapper(z_conv4, is_training)
    h_conv4 = tf.nn.relu( z_conv4_BN )
    
    h_pool4 = max_pool_2x2_same(h_conv4)
    
    
    W_conv5 = weight_variable([5, 5, 160, 180])
    b_conv5 = bias_variable([180])
    z_conv5 = conv2d(h_pool4, W_conv5) + b_conv5
    z_conv5_BN = batch_norm_wrapper(z_conv5, is_training)
    h_conv5 = tf.nn.relu( z_conv5_BN )
    h_pool5 = max_pool_2x2_same(h_conv5)
    
    
    W_conv6 = weight_variable([5, 5, 180, 180])
    b_conv6 = bias_variable([180])
    z_conv6 = conv2d(h_pool5, W_conv6) + b_conv5
    z_conv6_BN = batch_norm_wrapper(z_conv6, is_training)
    h_conv6 = tf.nn.relu( z_conv6_BN )
    h_pool6 = max_pool_2x2_same(h_conv6)
    
    
    W_fc1 = weight_variable([16 * 16 * 180, 1024])
    b_fc1 = bias_variable([1024])
    h_pool6_flat = tf.reshape(h_pool6, [-1, 16*16*180])
    z_fc1 = tf.matmul(h_pool6_flat, W_fc1) + b_fc1
    z_fc1_BN = batch_norm_wrapper_simple(z_fc1, is_training)
    h_fc1 = tf.nn.relu(z_fc1_BN)
    
    keep_prob = tf.placeholder(tf.float32)
    h_fc1_drop = tf.nn.dropout(h_fc1, keep_prob)
    
    #first digit
    W_fc2_d1 = weight_variable([1024, 11])
    b_fc2_d1 = bias_variable([11])
    
    y_conv_d1 = tf.matmul(h_fc1_drop, W_fc2_d1) + b_fc2_d1
    
    #second digit
    W_fc2_d2 = weight_variable([1024, 11])
    b_fc2_d2 = bias_variable([11])
    
    y_conv_d2 = tf.matmul(h_fc1_drop, W_fc2_d2) + b_fc2_d2
    
    #third digit
    W_fc2_d3 = weight_variable([1024, 11])
    b_fc2_d3 = bias_variable([11])
    
    y_conv_d3 = tf.matmul(h_fc1_drop, W_fc2_d3) + b_fc2_d3
    
    #fourth digit
    W_fc2_d4 = weight_variable([1024, 11])
    b_fc2_d4 = bias_variable([11])
    
    y_conv_d4 = tf.matmul(h_fc1_drop, W_fc2_d4) + b_fc2_d4
    
    #fifth digit
    W_fc2_d5 = weight_variable([1024, 11])
    b_fc2_d5 = bias_variable([11])
    
    y_conv_d5 = tf.matmul(h_fc1_drop, W_fc2_d5) + b_fc2_d5
    
    #digit size
    W_fc2_dsize = weight_variable([1024, 5])
    b_fc2_dsize = bias_variable([5])
    
    y_conv_dsize = tf.matmul(h_fc1_drop, W_fc2_dsize) + b_fc2_dsize
    
    
    cross_entropy = ( tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(y_conv_d1, y_d1)) 
                     + tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(y_conv_d2, y_d2))
                     + tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(y_conv_d3, y_d3))
                     + tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(y_conv_d4, y_d4))
                     + tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(y_conv_d5, y_d5))
                     + tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(y_conv_dsize, y_dsize))
                     ) + reg *(tf.nn.l2_loss(W_conv1) + tf.nn.l2_loss(W_conv2) 
                               + tf.nn.l2_loss(W_conv3) + tf.nn.l2_loss(W_conv4)
                               + tf.nn.l2_loss(W_conv5) + tf.nn.l2_loss(W_conv6)
                               + tf.nn.l2_loss(W_fc1)
                               + tf.nn.l2_loss(W_fc2_d1) + tf.nn.l2_loss(W_fc2_d2) 
                               + tf.nn.l2_loss(W_fc2_d3) + tf.nn.l2_loss(W_fc2_d4) 
                               + tf.nn.l2_loss(W_fc2_d5) + tf.nn.l2_loss(W_fc2_dsize) 
                     )
    
    train_step = tf.train.AdamOptimizer(1e-4).minimize(cross_entropy)
    
    #let's just check the first digit
    #correct_prediction = tf.equal(tf.argmax(y_conv_d1,1), tf.argmax(y_d1,1))
    correct_prediction = ( tf.cast( tf.equal(tf.argmax(y_conv_d1,1), tf.argmax(y_d1,1)) , tf.float32) 
                          * tf.cast( tf.equal(tf.argmax(y_conv_d1,1), tf.argmax(y_d1,1)) , tf.float32)
                          * tf.cast( tf.equal(tf.argmax(y_conv_d1,1), tf.argmax(y_d1,1)) , tf.float32)
                          * tf.cast( tf.equal(tf.argmax(y_conv_d1,1), tf.argmax(y_d1,1)) , tf.float32)
                          * tf.cast( tf.equal(tf.argmax(y_conv_d1,1), tf.argmax(y_d1,1)) , tf.float32)
                          )
    accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
        
    return ( (x_image, y_dsize, y_d1, y_d2, y_d3, y_d4, y_d5, keep_prob) , train_step, accuracy, z_fc1_BN, tf.train.Saver())
    
In [15]:
    
#Build training graph, train and save the trained model
#sess.close()
tf.reset_default_graph()
( (x_image, y_dsize, y_d1, y_d2, y_d3, y_d4, y_d5, keep_prob), 
 train_step, accuracy, z_fc1_BN, saver) = build_graph(is_training=True)
num_steps = 10000
summary_frequency = 40
BNs_train, BNs_test, acc_train, acc_test = [], [], [], []
with tf.Session() as sess:
    sess.run(tf.initialize_all_variables())
    for i in range(num_steps):
        (batch_x, batch_y_dsize, 
         batch_y_d1, batch_y_d2, 
         batch_y_d3, batch_y_d4, batch_y_d5) = next_batch(train_X, 
                                                          train_digit_size,train_digits, batch_size)
    
        feed_dict={x_image: batch_x, y_dsize: batch_y_dsize,
                   y_d1: batch_y_d1, y_d2: batch_y_d2, y_d3: batch_y_d3,
                   y_d4: batch_y_d4, y_d5: batch_y_d5, keep_prob: 0.5}
    
        train_step.run(feed_dict=feed_dict)
    
        if i%summary_frequency == 0:
            #train_accuracy = accuracy.eval(feed_dict=feed_dict)
            res_train = sess.run([accuracy,z_fc1_BN],feed_dict=feed_dict)
            print("step %d, training accuracy %g"%(i, res_train[0]))
            
            acc_train.append(res_train[0])
            BNs_train.append(np.mean(res_train[1],axis=0).flatten()[:10])
            
            (batch_x, batch_y_dsize, batch_y_d1,
             batch_y_d2, batch_y_d3, batch_y_d4, batch_y_d5) = next_batch(val_X, 
                                                                          val_digit_size, 
                                                                          val_digits, 
                                                                          batch_size, replace = False)
            feed_dict={x_image: batch_x, 
                       y_dsize: batch_y_dsize,y_d1: batch_y_d1, 
                       y_d2: batch_y_d2, y_d3: batch_y_d3,y_d4: batch_y_d4, 
                       y_d5: batch_y_d5, keep_prob: 1}
        
            res = sess.run([accuracy,z_fc1_BN],feed_dict=feed_dict)
            acc_test.append(res[0])
            # record the first 10 mean value of BN2 over the entire test set
            BNs_test.append(np.mean(res[1],axis=0).flatten()[:10])
        
    saved_model = saver.save(sess, 'temp-bn-save')
    
BNs_train, BNs_test, acc_train, acc_test = ( np.array(BNs_train), 
                                            np.array(BNs_test), 
                                            np.array(acc_train), 
                                            np.array(acc_test) )
    
    
In [16]:
    
fig, ax = plt.subplots()
ax.plot(range(0,len(acc_train)*summary_frequency,summary_frequency),acc_train, label='Training')
ax.plot(range(0,len(acc_test)*summary_frequency,summary_frequency),acc_test, label='Validation')
ax.set_xlabel('Training steps')
ax.set_ylabel('Accuracy')
ax.set_ylim([0,1])
ax.set_title('Batch Normalization Accuracy')
ax.legend(loc=4)
plt.show()
    
    
In [22]:
    
np.mean(acc_test[200:]), acc_test[249]
    
    Out[22]:
In [23]:
    
fig, axes = plt.subplots(5, 2, figsize=(6,12))
fig.tight_layout()
for i, ax in enumerate(axes):
    ax[0].set_title("training BN")
    ax[1].set_title("validation BN")
    ax[0].plot(BNs_train[:,i])
    ax[1].plot(BNs_test[:,i])
    
    
In [25]:
    
sess.close()
tf.reset_default_graph()
( (x_image, y_dsize, y_d1, y_d2, y_d3, y_d4, y_d5, keep_prob), 
 train_step, accuracy, z_conv2, saver) = build_graph(is_training=True)
predictions = []
correct = 0
with tf.Session() as sess:
    sess.run(tf.initialize_all_variables())
    saver.restore(sess, saved_model)
    for i in range(1000):
    
        (batch_x, batch_y_dsize, batch_y_d1, 
         batch_y_d2, batch_y_d3, batch_y_d4, batch_y_d5) = next_batch(val_X, val_digit_size, val_digits, 
                                                                      10, replace = False)
        feed_dict={x_image: batch_x, 
                   y_dsize: batch_y_dsize,y_d1: batch_y_d1, 
                   y_d2: batch_y_d2, y_d3: batch_y_d3,y_d4: batch_y_d4, 
                   y_d5: batch_y_d5, keep_prob: 1}
    
        corr = sess.run(accuracy,
                         feed_dict=feed_dict)
        correct += corr
    #predictions.append(pred[0])
#print("PREDICTIONS:", predictions)
print("ACCURACY:", correct/1000)
    
    
In [61]:
    
a = [1,2,3,4]
a[-1]
    
    Out[61]:
In [63]:
    
a[1:]
    
    Out[63]:
In [72]:
    
[1]
    
    Out[72]:
In [ ]: