tensorflow學習筆記（六）：cnn過程可視化

import tensorflow as tf
from tensorflow.examples.tutorials.mnist import input_data
import os
import matplotlib.pyplot as plt

os.environ['TF_CPP_MIN_LOG_LEVEL'] = '3'


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

# print mnist.train.images.shape,mnist.train.labels.shape
# (55000, 784) (55000, 10)
# 784 = 28*28
# print mnist.test.images.shape,mnist.test.labels.shape\
# (10000, 784) (10000, 10)
# print mnist.validation.images.shape,mnist.validation.labels.shape
# (5000, 784) (5000, 10)

def Weight_value(shape):
    init = tf.random_normal(shape, stddev=)
    return tf.Variable(init, name="weight")
def bias_value(shape):
    init = tf.constant(, shape=shape)
    return tf.Variable(init)
def conv2d(x, W):
    return tf.nn.conv2d(x, W, strides=[, , , ], padding="SAME")
def pool_2x2(x):
    return tf.nn.max_pool(x, ksize=[, , , ], strides=[, , , ], padding="SAME")

xs = tf.placeholder(tf.float32, [None, ])
ys = tf.placeholder(tf.float32, [None, ])

x_image = tf.reshape(xs, [-, , , ])

# layer1 conv1  [-1, 28, 28, 32]
W_conv1 = Weight_value([, , , ])
b_conv1 = bias_value([])
h_conv1 = tf.nn.relu(conv2d(x_image, W_conv1)+b_conv1)
# layer2 pool1 [-1, 14, 14, 32]
h_pool1 = pool_2x2(h_conv1)
# layer3 conv2 [-1, 14, 14, 64]
W_conv2 = Weight_value([, , , ])
b_conv2 = bias_value([])
h_conv2 = tf.nn.relu(conv2d(h_pool1, W_conv2)+b_conv2)
# layer4 pool2 [-1,7,7,64]
h_pool2 = pool_2x2(h_conv2)
# layer5 fc1 [-1,1024]
h_pool2_flat = tf.reshape(h_pool2, [-, **])
W_fc1 = Weight_value([**, ])
b_fc1 = bias_value([])
h_fc1 = tf.nn.relu(tf.matmul(h_pool2_flat, W_fc1)+b_fc1)
#layer6 dropout
keep_prob = tf.placeholder(tf.float32)
h_fc1_drop = tf.nn.dropout(h_fc1, keep_prob)
# layer7 fc2 [-1,10]
W_fc2 = Weight_value([, ])
b_fc2 = bias_value([])
y_conv = tf.nn.softmax(tf.matmul(h_fc1_drop, W_fc2)+b_fc2)

# cross_entropy
cross_entropy = tf.reduce_mean(-tf.reduce_sum(ys*tf.log(y_conv), reduction_indices=[]))
# optimizer
train_step = tf.train.AdamOptimizer().minimize(cross_entropy)
# accuracy
correct_prediction = tf.equal(tf.argmax(ys, ), tf.argmax(y_conv, ))
accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
# init
init = tf.global_variables_initializer()
# sess
with tf.Session() as sess:
    sess.run(init)
    img1 = mnist.train.images[]
    img1.shape = [, ]
    result = sess.run(h_conv1, feed_dict={xs : img1})
    for i in range():
        show_img = result[:, :, :, i]
        # print type(show_img)
        show_img.shape = [, ]
        plt.subplot(, , i + )
        plt.imshow(show_img, cmap='gray')
        plt.axis('off')
    plt.show()











    # for i in range(1001):
    #     x_batch, y_batch = mnist.train.next_batch(50)
    #     sess.run(train_step, feed_dict={xs:x_batch, ys:y_batch, keep_prob:0.5})
    #     if i%100 == 0:
    #         x_test, y_test = mnist.test.next_batch(50)
    #         print i, ' step train ', sess.run(accuracy, feed_dict={xs: x_batch, ys: y_batch, keep_prob: 1})
    #         print i, ' step test', sess.run(accuracy, feed_dict={xs:x_test, ys:y_test, keep_prob: 1})
           

可以看出32個卷積核對圖檔特征的提取