cs295k hw3 conv net for mnist

tf.py

import tensorflow as tf
import sys

from tensorflow.examples.tutorials.mnist import input_data

# Initialize weights with a small amount of noise for symmetry breaking, and to prevent 0 gradients.
# For ReLU neurons, good practice to initialize with a slightly positive initial bias to avoid "dead neurons"
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)

# define our convolution with a stride of 1, zero-padded:
def conv2d(x, W):
  return tf.nn.conv2d(x, W, strides=[1, 1, 1, 1], padding='SAME')

# max pooling over 2x2 blocks
def max_pool_2x2(x):
  return tf.nn.max_pool(x, ksize=[1, 2, 2, 1],
                        strides=[1, 2, 2, 1], padding='SAME')

mnist = input_data.read_data_sets('MNIST_data', one_hot=True)

# stuff from before:
img = tf.placeholder(tf.float32, [None,784]) # image
lab = tf.placeholder(tf.float32,[None,10])

# first layer: convolution followed by pooling
# 32 features for each 5x5 filter
W_conv1 = weight_variable([5, 5, 1, 32])
b_conv1 = bias_variable([32])

# reshape our image to a 4d tensor:
img_tensor = tf.reshape(img, [-1,28,28,1])

# convolve the image with the weights, add bias and compute relu, then max pool:
h_conv1 = tf.nn.relu(conv2d(img_tensor, W_conv1) + b_conv1)
h_pool1 = max_pool_2x2(h_conv1)

# second layer: 64 features for each 5x5 filter
W_conv2 = weight_variable([5, 5, 32, 64])
b_conv2 = bias_variable([64])
h_conv2 = tf.nn.relu(conv2d(h_pool1, W_conv2) + b_conv2)
h_pool2 = max_pool_2x2(h_conv2)

# densely-connected layer: image reduced to 7x7, use this layer of 1024 neurons
#  to process the entire image
W_fc1 = weight_variable([7 * 7 * 64, 1024])
b_fc1 = bias_variable([1024])

# reshape pooling layer into 2d tensor (a batch of vectors),
#  apply weights and biases, then relu
h_pool2_flat = tf.reshape(h_pool2, [-1, 7*7*64])
h_fc1 = tf.nn.relu(tf.matmul(h_pool2_flat, W_fc1) + b_fc1)

# apply dropout. according to tf, this requires no rescaling aka magic
keep_prob = tf.placeholder(tf.float32) # probability output kept during dropout
h_fc1_drop = tf.nn.dropout(h_fc1, keep_prob)

# feed-forward layer for softmax over digits:
W_fc2 = weight_variable([1024, 10])
b_fc2 = bias_variable([10])

y_conv=tf.nn.softmax(tf.matmul(h_fc1_drop, W_fc2) + b_fc2)

sess = tf.Session()

# train the model
cross_entropy = tf.reduce_mean(-tf.reduce_sum(lab * tf.log(y_conv), reduction_indices=[1]))
train_step = tf.train.AdamOptimizer(1e-4).minimize(cross_entropy)
correct_prediction = tf.equal(tf.argmax(y_conv,1), tf.argmax(lab,1))
accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
sess.run(tf.initialize_all_variables())
for i in range(2000):
  batch = mnist.train.next_batch(50)
  if i%100 == 0:
    train_accuracy = accuracy.eval(feed_dict={img:batch[0], lab: batch[1], keep_prob: 1.0}, session=sess)
    print("step %d, training accuracy %g"%(i, train_accuracy))
  train_step.run(feed_dict={img: batch[0], lab: batch[1], keep_prob: 0.5}, session=sess)

# evaluate:
print("test accuracy %g"%accuracy.eval(feed_dict={img: mnist.test.images, lab: mnist.test.labels, keep_prob: 1.0}, session=sess))