tensorflow 下CNN卷积神经网络实现
简介：卷积神经网络原理其实就是基于感受野，感受野讲的是只需识别某个图片的一个小区域就知道有某个东西。比如，在一个会议室里，要识别里面是否有投影仪，只需要看放投影仪的那部分小区域就知道这个会议室有投影仪。在卷积神经网络中也是通过卷积和池化来减少计算的复杂度，正式基于感受野的原理。
什么是卷积神经网络？我们以奠基之作Lenet-5为例子
上图中有输入层，卷积C1、C3、C5层，池化S2、S4层，全连接层、输出层
对于一张输入 32*32的黑白图片，经过5x5的filter(神经元)提取出 28x28 feature map。对于C1层，5x5的filter就有25个可训练参数，加上一个共享偏置，每个卷积核就有26个可训练参数，总共就有26*6=156个参数。有(5x5+1)x28x28x6=122304个连接，5*5神经元和1个bias都要和28x28相连
对于池化S2: filter是2x2的，采用的是加和乘一个参数加偏置，训练的参数有 (1+1)x6=12个参数，连接数(2×2+1)x6×14×14=5880个
由于C3的输入并非全部的S2输出，之后的参数和连接就不计算了。值得注意的是在S4(5x5)-->filter(5x5)-->C5类似于一个全连接，图中并没有标错
CNN源码可以下载
代码解释如下，其中将几乎每一步的解释都备注了在了代码。期间有不懂的可以使用print打印语句。欢迎评论。
# -*- coding: utf-8 -*-
"""Simple, end-to-end, LeNet-5-like convolutional MNIST model example.
This should achieve a test error of 0.7%. Please keep this model as simple and
linear as possible, it is meant as a tutorial for simple convolutional models.
Run with --self_test on the command line to execute a short self-test.
"""
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import gzip
import os
import sys
import time
import numpy
from six.moves import urllib
from six.moves import xrange # pylint: disable=redefined-builtin
import tensorflow as tf
SOURCE_URL = 'http://yann.lecun.com/exdb/mnist/'
WORK_DIRECTORY = '/home/xzy/tf-input-data'
IMAGE_SIZE = 28
NUM_CHANNELS = 1
PIXEL_DEPTH = 255
NUM_LABELS = 10
VALIDATION_SIZE = 500 # 自定义validate 数据集.
SEED = 66478 # Set to None for random seed.
BATCH_SIZE = 64
NUM_EPOCHS = 10 # 自定义训练集上轮循10次
EVAL_BATCH_SIZE = 64 # 定义每批次数据
EVAL_FREQUENCY = 100 # 执行步长
TRAIN_MAX = 5000 # 自定义训练数据大小
tf.app.flags.DEFINE_boolean("self_test", False, "True if running a self test.")
tf.app.flags.DEFINE_boolean('use_fp16', False, "Use half floats instead of full floats if True.")
FLAGS = tf.app.flags.FLAGS
def data_type():
"""返回激活函数、权重、placeholder变量的类型"""
if FLAGS.use_fp16:
return tf.float16
else:
return tf.float32
def maybe_download(filename):
"""如果数据不存在，则下载数据"""
if not tf.gfile.Exists(WORK_DIRECTORY):
tf.gfile.MakeDirs(WORK_DIRECTORY)
filepath = os.path.join(WORK_DIRECTORY, filename)
if not tf.gfile.Exists(filepath):
filepath, _ = urllib.request.urlretrieve(SOURCE_URL + filename, filepath)
with tf.gfile.GFile(filepath) as f:
size = f.size()
print('Successfully downloaded', filename, size, 'bytes.')
return filepath
def extract_data(filename, num_images):
"""提取图像为4D tensor [image index, y, x, channels].
其值从[0, 255]重新调整为[-0.5, 0.5].
"""
print('Extracting', filename)
with gzip.open(filename) as bytestream:
bytestream.read(16)
buf = bytestream.read(IMAGE_SIZE * IMAGE_SIZE * num_images)
data = numpy.frombuffer(buf, dtype=numpy.uint8).astype(numpy.float32)
"""
print('--------------------------------')
print('data1=', data)
print('--------------------------------')
"""
data = (data - (PIXEL_DEPTH / 2.0)) / PIXEL_DEPTH
"""
print('################################')
print('data2=', data)
print('################################')
"""
data = data.reshape(num_images, IMAGE_SIZE, IMAGE_SIZE, 1)
"""
print('%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%')
print('data3=', data)
print('%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%')
"""
return data
def extract_labels(filename, num_images):
"""提取labels成一个int64 IDs的vector"""
print('Extracting', filename)
with gzip.open(filename) as bytestream:
bytestream.read(8)
buf = bytestream.read(1 * num_images)
labels = numpy.frombuffer(buf, dtype=numpy.uint8).astype(numpy.int64)
return labels
def fake_data(num_images):
"""生成与MNIST的维度相匹配的假数据集"""
data = numpy.ndarray(
shape=(num_images, IMAGE_SIZE, IMAGE_SIZE, NUM_CHANNELS),
dtype=numpy.float32)
labels = numpy.zeros(shape=(num_images,), dtype=numpy.int64)
for image in xrange(num_images):
label = image % 2
data[image, :, :, 0] = label - 0.5
labels[image] = label
return data, labels
def error_rate(predictions, labels):
"""基于密集预测和稀疏标签返回错误率"""
return 100.0 - (100.0 * numpy.sum(numpy.argmax(predictions, 1) == labels) /
predictions.shape[0])
def main(argv=None): # pylint: disable=unused-argument
if FLAGS.self_test:
print('Running self-test.')
train_data, train_labels = fake_data(256)
validation_data, validation_labels = fake_data(EVAL_BATCH_SIZE)
test_data, test_labels = fake_data(EVAL_BATCH_SIZE)
num_epochs = 1
else:
# 获取数据
train_data_filename = maybe_download('train-images-idx3-ubyte.gz')
train_labels_filename = maybe_download('train-labels-idx1-ubyte.gz')
test_data_filename = maybe_download('t10k-images-idx3-ubyte.gz')
test_labels_filename = maybe_download('t10k-labels-idx1-ubyte.gz')
# 提取数据并转成numpy的array.
train_data = extract_data(train_data_filename, 60000)
train_labels = extract_labels(train_labels_filename, 60000)
test_data = extract_data(test_data_filename, 1000)
test_labels = extract_labels(test_labels_filename, 1000)
# 产生一个验证集，并分片处理.VALIDATION_SIZE = 5000.
validation_data = train_data[:VALIDATION_SIZE, ...] # 0-5000分片
validation_labels = train_labels[:VALIDATION_SIZE]
train_data = train_data[VALIDATION_SIZE:TRAIN_MAX+VALIDATION_SIZE, ...]
train_labels = train_labels[VALIDATION_SIZE:TRAIN_MAX+VALIDATION_SIZE]
num_epochs = NUM_EPOCHS # NUM_EPOCHS = 10
train_size = train_labels.shape[0]
""" 训练样本和标签占位声明，每次训练都会喂一个批次的数据，
通过向Run()方法传递{feed_dict}参数"""
train_data_node = tf.placeholder(data_type(),
shape=(BATCH_SIZE, IMAGE_SIZE, IMAGE_SIZE, NUM_CHANNELS))
train_labels_node = tf.placeholder(tf.int64, shape=(BATCH_SIZE,))
eval_data = tf.placeholder(data_type(),
shape=(EVAL_BATCH_SIZE, IMAGE_SIZE, IMAGE_SIZE, NUM_CHANNELS))
"""持久化训练权重，通过{tf.initialize_all_variables().run()}执行操作"""
# 5x5 filter, depth 32.
conv1_weights = tf.Variable(tf.truncated_normal([5, 5, NUM_CHANNELS, 32],
stddev=0.1, seed=SEED, dtype=data_type()))
conv1_biases = tf.Variable(tf.zeros([32], dtype=data_type()))
conv2_weights = tf.Variable(tf.truncated_normal(
[5, 5, 32, 64], stddev=0.1, seed=SEED, dtype=data_type()))
conv2_biases = tf.Variable(tf.constant(0.1, shape=[64], dtype=data_type()))
# fully connected, depth 512.
fc1_weights = tf.Variable(tf.truncated_normal([IMAGE_SIZE // 4 * IMAGE_SIZE // 4 * 64, 512],
stddev=0.1, seed=SEED, dtype=data_type()))
fc1_biases = tf.Variable(tf.constant(0.1, shape=[512], dtype=data_type()))
fc2_weights = tf.Variable(tf.truncated_normal(
[512, NUM_LABELS], stddev=0.1, seed=SEED, dtype=data_type()))
fc2_biases = tf.Variable(tf.constant(0.1, shape=[NUM_LABELS], dtype=data_type()))
"""训练子图和评估子图共用一个模型，通过共享训练参数"""
def model(data, train=False):
"""2维卷积, 使用 'SAME'处理边界 (feature map输出和输入有相同的维数
{strides} 是一个 4D array，他的形状为[image index, y, x, depth].
data=[bitchSize=64, 28, 28, channels=1]
conv1_weights=[5, 5, channels=1, filters=32]
print("---------initial data---------")
print('data=', data)
print('conv1_weights=', conv1_weights)
"""
conv = tf.nn.conv2d(data, conv1_weights, strides=[1, 1, 1, 1], padding='SAME')
relu = tf.nn.relu(tf.nn.bias_add(conv, conv1_biases))
"""
最大池化
ksize：池化窗口的大小，取一个四维向量，一般是[1, height, width, 1]，
因为我们不想在batch和channels上做池化
"""
pool = tf.nn.max_pool(relu, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')
"""first conv-pool
conv=[64, 28, 28, filters=32]
pool=[64, 14, 14, channels=32]
conv2_weights=[5, 5, channels=32, filters=64]
print('---------first conv-pool---------')
print('pool=', pool)
print('conv=', conv)
print('conv2_weights=', conv2_weights)
"""
conv = tf.nn.conv2d(pool, conv2_weights, strides=[1, 1, 1, 1], padding='SAME')
relu = tf.nn.relu(tf.nn.bias_add(conv, conv2_biases))
pool = tf.nn.max_pool(relu, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')
"""second conv-pool
conv=[64, 14, 14, filters=64]
pool=[64, 7, 7, channels=64]
print('----------second conv-pool---------')
print('conv=', conv)
print('pool=', pool)
"""
# 重构feature map成2D矩阵，传给全链接网络
pool_shape = pool.get_shape().as_list()
reshape = tf.reshape(pool,
[pool_shape[0], pool_shape[1] * pool_shape[2] * pool_shape[3]])
"""
print('-------reshape---------')
print('pool_shape=', pool_shape)
print('reshape=', reshape)
print('fc1_weights=',fc1_weights)
"""
# 全链接网络层
hidden = tf.nn.relu(tf.matmul(reshape, fc1_weights) + fc1_biases)
# 训练数据时候，采用dropout随机丢弃50%的像素点
if train:
hidden = tf.nn.dropout(hidden, 0.5, seed=SEED)
return tf.matmul(hidden, fc2_weights) + fc2_biases
""" 训练计算公式: logits + cross-entropy loss.
交叉熵：之前使用sigmoid{f(z) = 1/(1+\exp(-z))}函数计算神经元与真实值的欧式距离来判断偏差，
反向去修正权重和偏置，但函数接近1的时候，导数会非常小，学习的速度就会变得非常缓慢。解决这个缺点
就是使用交叉熵。参看http://blog.csdn.net/bixiwen_liu/article/details/52922008
"""
logits = model(train_data_node, True)
loss = tf.reduce_mean(tf.nn.sparse_softmax_cross_entropy_with_logits(
logits=logits, labels=train_labels_node))
# 防止过拟合，对全链接网络的参数进行正则化，常见的是dropout.
regularizers = (tf.nn.l2_loss(fc1_weights) + tf.nn.l2_loss(fc1_biases) +
tf.nn.l2_loss(fc2_weights) + tf.nn.l2_loss(fc2_biases))
# 加入一个正则项.
loss += 5e-4 * regularizers
# 优化器: 每个批次都会增加，并且控制着学习率的衰变
batch = tf.Variable(0, dtype=data_type())
# 每次训练全部样本，学习率都会衰变，使用指数衰变
learning_rate = tf.train.exponential_decay(
0.01, # 开始学习率
batch * BATCH_SIZE, # 数据集大小=批次×每批次数据大小.
train_size, # 训练多大更新学习率
0.95, # 衰变率--单位时间内衰变的几率.
staircase=True) # true代表训练train_size完更新一次，false代表每个样本都更新
# 使用momentum（动量--它模拟的是物体运动时的惯性，即更新的时候在一定程度上保留之前更新的方向） 优化器.
optimizer = tf.train.MomentumOptimizer(learning_rate, 0.9).minimize(loss, global_step=batch)
# 计算当前训练minibatch的预测概率.
train_prediction = tf.nn.softmax(logits)
# 计算测试和验证minibatch上的预测概率
eval_prediction = tf.nn.softmax(model(eval_data))
# 通过feeding多个批次的数据到{eval_data},并从{eval_predictions}获取结果并保存.
def eval_in_batches(data, sess):
"""获取一个小批次数据集的所有预测概率."""
size = data.shape[0]
if size < EVAL_BATCH_SIZE: # 数据集小于每批次数据EVAL_BATCH_SIZE=64
raise ValueError("batch size for evals larger than dataset: %d" % size)
predictions = numpy.ndarray(shape=(size, NUM_LABELS), dtype=numpy.float32) # NUM_LABELS=10
for begin in xrange(0, size, EVAL_BATCH_SIZE): # 每次增量为EVAL_BATCH_SIZE,一直到大于size
end = begin + EVAL_BATCH_SIZE
if end <= size:
predictions[begin:end, :] = sess.run(eval_prediction,
feed_dict={eval_data: data[begin:end, ...]})
else:
batch_predictions = sess.run(eval_prediction,
feed_dict={eval_data: data[-EVAL_BATCH_SIZE:, ...]})
predictions[begin:, :] = batch_predictions[begin - size:, :]
return predictions
# 创建session开始训练数据
st_time = time.clock()
start_time = time.time()
with tf.Session() as sess:
# 初始化所有的训练参数
tf.initialize_all_variables().run()
print('Initialized!')
# 根据 steps进行循环.
# num_epochs = 10(非self_test)或 1(self_test) BATCH_SIZE = 64 train_size=55000
for step in xrange(int(num_epochs * train_size) // BATCH_SIZE):
# 计算当前 minibatch 在数据集里的偏移.
offset = (step * BATCH_SIZE) % (train_size - BATCH_SIZE)
batch_data = train_data[offset:(offset + BATCH_SIZE), ...]
batch_labels = train_labels[offset:(offset + BATCH_SIZE)]
# 向数据字典里喂数据
feed_dict = {train_data_node: batch_data, train_labels_node: batch_labels}
# 取数据
_, l, lr, predictions = sess.run([optimizer, loss, learning_rate, train_prediction],
feed_dict=feed_dict)
if step % EVAL_FREQUENCY == 0:
elapsed_time = time.time() - start_time
start_time = time.time()
print('Step %d (epoch %.2f), %.1f ms' %
(step, float(step) * BATCH_SIZE / train_size,
1000 * elapsed_time / EVAL_FREQUENCY))
print('Minibatch loss: %.3f, learning rate: %.6f' % (l, lr))
print('Minibatch error: %.1f%%' % error_rate(predictions, batch_labels))
print('Validation error: %.1f%%' % error_rate(
eval_in_batches(validation_data, sess), validation_labels))
sys.stdout.flush()
# 打印最终的结果
test_error = error_rate(eval_in_batches(test_data, sess), test_labels)
print('Test error: %.1f%%' % test_error)
if FLAGS.self_test:
print('test_error', test_error)
assert test_error == 0.0, 'expected 0.0 test_error, got %.2f' % (test_error,)
ed_time = time.clock()
print('******total train time:', ed_time - st_time)
if __name__ == '__main__':
tf.app.run()
参考资料
lenet-5 http://blog.csdn.net/xuanyuansen/article/details/41800721
卷积神经网络 http://blog.csdn.net/u011453773/article/details/51597072