In [1]:
import tensorflow as tf
import numpy as np
from tensorflow.examples.tutorials.mnist import input_data
mnist_data = input_data.read_data_sets('/tmp/data', one_hot=True)
In [2]:
## Visualize a sample subset of data
import matplotlib.pyplot as plt
%matplotlib inline
import numpy as np
fig,ax = plt.subplots(5,10,figsize=(10,5))
for i in range(5):
for j in range(10):
index = (i-1)*5 + j
ax[i][j].imshow(np.reshape(1-mnist_data.test.images[index],(28,28)), cmap='Greys_r')
ax[i][j].set_xticks([])
ax[i][j].set_yticks([])
fig.show()
In [3]:
### Function for initializing weights
def weight_variable(shape):
initial = tf.truncated_normal(shape, stddev=0.1)
return tf.Variable(initial)
### Function for initializing biases
def bias_variable(shape):
initial = tf.constant(0.1, shape=shape)
return tf.Variable(initial)
### Function for 2-D Convolution
def conv2d(x, W):
return tf.nn.conv2d(x, W, strides=[1, 1, 1, 1], padding='SAME')
### Function for MaxPooling(2x )
def max_pool_2x2(x):
return tf.nn.max_pool(x, ksize=[1, 2, 2, 1],
strides=[1, 2, 2, 1], padding='SAME')
In [4]:
## place-holders for input samples/labels
x = tf.placeholder(tf.float32, shape=[None, 784])
y_ = tf.placeholder(tf.float32, shape=[None, 10])
## Reshape input:
#### --> reshape input to a 4-D tensor (-1,28,28,1)
x_image = tf.reshape(x, [-1,28,28,1])
#~~~~~~~~~~~~~~~~~~~~~~~~ 1st layer ~~~~~~~~~~~~~~~~~~~~~~~~~~#
## 1st convolutional layer: convolution filter shape (5,5,1,32)
#### --> input channels: 1
#### --> each filter size: (5x5)
#### --> output: 32 channels
W_conv1 = weight_variable(shape=[5, 5, 1, 32])
b_conv1 = bias_variable([32])
### Apply convolution, non-linear activation, and max-ppoling
h_conv1 = conv2d(x_image, W_conv1) + b_conv1
h_relu1 = tf.nn.relu(h_conv1)
h_pool1 = max_pool_2x2(h_relu1)
#~~~~~~~~~~~~~~~~~~~~~~~~ 2nd layer ~~~~~~~~~~~~~~~~~~~~~~~~~~#
## 2nd convolutional layer: convolution filter shape (5,5,32,64)
#### --> input channels: 32
#### --> each filter size: (5x5)
#### --> output: 64 channels
W_conv2 = weight_variable(shape=[5, 5, 32, 64])
b_conv2 = bias_variable([64])
### Apply convolution, non-linear activation, and max-ppoling
h_conv2 = conv2d(h_pool1, W_conv2) + b_conv2
h_relu2 = tf.nn.relu(h_conv2)
h_pool2 = max_pool_2x2(h_relu2)
#~~~~~~~~~~~~~~~~~~~~~~~~ 3rd layer ~~~~~~~~~~~~~~~~~~~~~~~~~~#
## 3rd Layer: Fully-Connected
W_fc3 = weight_variable([7 * 7 * 64, 1024])
b_fc3 = bias_variable([1024])
### Apply fully connected layer
h_pool2_flat = tf.reshape(h_pool2, [-1, 7*7*64])
h_fc3 = tf.nn.relu(tf.matmul(h_pool2_flat, W_fc3) + b_fc3)
## drop-out: placeholder for
## probability of dropoout in fully connected layer
keep_prob = tf.placeholder(tf.float32)
h_fc3_drop = tf.nn.dropout(h_fc3, keep_prob)
#~~~~~~~~~~~~~~~~~~~~~~ ReadOut layer ~~~~~~~~~~~~~~~~~~~~~~~#
W_fc4 = weight_variable([1024, 10])
b_fc4 = bias_variable([10])
### Apply fully connected layer
y_pred = tf.matmul(h_fc3_drop, W_fc4) + b_fc4
In [5]:
cross_entropy_loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(y_pred, y_))
optimizer = tf.train.AdamOptimizer(1e-4).minimize(cross_entropy_loss)
label_pred = tf.argmax(y_pred,1)
correct_prediction = tf.equal(label_pred, tf.argmax(y_,1))
accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
In [6]:
init = tf.initialize_all_variables()
losses, trainaccs = [], []
sess = tf.Session()
with sess:
sess.run(init)
for i in range(20000):
batch = mnist_data.train.next_batch(50)
#train_accuracy = accuracy.eval(
# feed_dict={x:batch[0], y_: batch[1], keep_prob: 1.0})
_, train_acc, loss = sess.run([optimizer, accuracy, cross_entropy_loss],
feed_dict={x: batch[0], y_: batch[1], keep_prob: 0.5})
losses.append(loss)
trainaccs.append(train_acc)
if i%1000 == 0:
print("step {:05d}, training accuracy {:.2f}".format(i, train_acc))
fig,ax = plt.subplots(2,1,figsize=(10,6))
ax[0].plot(range(len(losses)), losses, alpha=0.5)
ax[0].set_ylim((-0.09,2))
ax[1].plot(range(len(trainaccs)), trainaccs, alpha=0.5)
ax[1].set_ylim((0,1.09))
plt.show()
### reporting test accuracy:
#print("test accuracy {:g}".format(accuracy.eval(
# feed_dict={x: mnist_data.test.images,
# y_: mnist_data.test.labels, keep_prob: 1.0})))
## predicting all test values at once failed on my laptop
## splitting the test dataset:
test_splits = np.split(np.arange(mnist_data.test.images.shape[0]), 20)
test_preds = np.zeros(mnist_data.test.images.shape[0])
for indx in test_splits:
y_test = sess.run(label_pred,
feed_dict={x:mnist_data.test.images[indx], keep_prob:1.0})
test_preds[indx] = y_test
print("test accuracy {:g}".format(accuracy.eval(feed_dict={
label_pred:test_preds, y_:mnist_data.test.labels
})))
In [21]:
indx = np.where(test_preds != np.argmax(mnist_data.test.labels, axis=1))[0]
fig,ax = plt.subplots(5,15,figsize=(20,6))
for i in range(5):
for j in range(15):
k = (i-1)*5 + j
ax[i][j].imshow(np.reshape(1-mnist_data.test.images[indx[k]],(28,28)), cmap='Greys_r')
ax[i][j].set_xticks([])
ax[i][j].set_yticks([])
ax[i][j].text(x=20, y=7, s="{:d}".format(int(test_preds[indx[k]])),
color='red', fontsize=20)
fig.show()
In [ ]: