import tensorflow as tf
import numpy as np
import tensorflow.contrib.layers as layers
from tensorflow.python import debug as tf_debug
from tensorflow.examples.tutorials.mnist import input_data
mnist=input_data.read_data_sets("mnist",one_hot=True)
Weights={
"w1":tf.Variable(tf.random_normal([5,5,1,32])),
"w2":tf.Variable(tf.random_normal([3,3,32,49])),
"wf1":tf.Variable(tf.random_normal([7*7*49,1024])),
"out":tf.Variable(tf.random_normal([1024,10]))
}
bias={
"b1": tf.Variable(tf.random_normal([32])),
"b2":tf.Variable(tf.random_normal([49])),
"out":tf.Variable(tf.random_normal([10])),
}
#strides [1, stride,stride, 1]
#ksize [1, height, width, 1]
def ProcessLayers(input):
#batch, height, width, channels
# data N 28*28 1
input = tf.reshape(input, shape=[-1, 28, 28, 1])
#data N 28 28 32
fc1=tf.nn.conv2d(input,Weights["w1"],strides=[1, 1, 1, 1],padding='SAME')
fc1=tf.nn.bias_add(fc1,bias["b1"])
# data N 14*14 32
fc1=tf.nn.max_pool(fc1,ksize=[1,2,2, 1],strides=[1, 2,2, 1],padding='SAME')
# data N 14*14 49
fc2=tf.nn.conv2d(fc1,Weights["w2"],strides=[1, 1, 1, 1],padding='SAME')
fc2=tf.nn.bias_add(fc2,bias["b2"])
# data N 7*7 49
fc2 = tf.nn.max_pool(fc2, ksize=[1,2,2, 1], strides=[1, 2,2, 1],padding='SAME')
fc2=tf.nn.relu(fc2)
#data N 1024
fc2=tf.reshape(fc2, [-1, Weights["wf1"].get_shape().as_list()[0]])
fc2=tf.matmul(fc2,Weights["wf1"])
# dropout 如果放在全连接之前，会导致参数丢失过多
fc2 = tf.nn.dropout(fc2, 0.5)
#data N 10
out=tf.matmul(fc2, Weights["out"])
out=tf.nn.bias_add(out,bias["out"])
return out
x=tf.placeholder(tf.float32,shape=[None,784],name="x")
y=tf.placeholder(tf.float32,shape=[None,10],name="y")
y1=ProcessLayers(x)
entroy=tf.nn.softmax_cross_entropy_with_logits(logits=y1,labels=y)
loss=tf.reduce_mean(entroy)
correct_pred = tf.equal(tf.argmax(y1, 1), tf.argmax(y, 1))
accuracy = tf.reduce_mean(tf.cast(correct_pred, tf.float32))
optimizer=tf.train.AdamOptimizer(learning_rate=0.01)
train_step=optimizer.minimize(loss)
init=tf.global_variables_initializer()
batch_size=200
epochs=int(mnist.train.num_examples/batch_size)
with tf.Session() as sess:
sess.run(init)
"""
x_data = {'images': mnist.train.images}['images']
y_data = mnist.train.labels
"""
x_data_test = {'images': mnist.test.images}['images']
y_data_test = mnist.test.labels
for nStep in range(100):
for n in range(epochs):
batch_x,batch_y=mnist.train.next_batch(batch_size)
sess.run(train_step,feed_dict={x:batch_x,y:batch_y})
lost=sess.run(loss,feed_dict={x:batch_x,y:batch_y})
acc=sess.run(accuracy,feed_dict={x:batch_x,y:batch_y})
test = sess.run(accuracy, feed_dict={x: x_data_test[:256], y: y_data_test[:256]})
print("epoch:%2d,lost:%.2f,acc:%.2f,test:%.2f"%(nStep,lost,acc,test))
----------------------------------
epoch: 0,lost:130.59,acc:0.94,test:0.95
epoch: 1,lost:21.38,acc:0.98,test:0.95
epoch: 2,lost:88.58,acc:0.94,test:0.95
epoch: 3,lost:28.32,acc:0.95,test:0.96
epoch: 4,lost:26.61,acc:0.96,test:0.96
