TensorFlow手写数字识别mnist example源码分析

TensorFlow手写数字识别mnist example源码分析 
TensorFlow 默认安装在 /usr/lib/Python/site-packages/tensorflow/

实例文件位于tensorflow/models/image/mnist/convolutional.py，为TensorFlow自带的example文件。

# Copyright 2015 The TensorFlow Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# ==============================================================================
"""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 argparse
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/'  # 数据源
WORK_DIRECTORY = 'data'  # 工作目录，存放下载的数据
# MNIST 数据集特征：
#     图像尺寸 28x28
IMAGE_SIZE = 28
NUM_CHANNELS = 1  # 黑白图像
PIXEL_DEPTH = 255  # 像素值0~255
NUM_LABELS = 10  # 标签分10个类别
VALIDATION_SIZE = 5000  # 验证集大小
SEED = 66478  # 随机数种子，可设为 None 表示真的随机
BATCH_SIZE = 64  # 批处理大小为64
NUM_EPOCHS = 10  # 数据全集一共过10遍网络
EVAL_BATCH_SIZE = 64  # 验证集批处理大小也是64
# 验证时间间隔，每训练100个批处理，做一次评估
EVAL_FREQUENCY = 100  # Number of steps between evaluations.

FLAGS = None


def data_type():
    """Return the type of the activations, weights, and placeholder variables."""
    if FLAGS.use_fp16:
        return tf.float16
    else:
        return tf.float32


def maybe_download(filename):
    """如果下载过了数据，就不再重复下载"""
    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):
    # 抽取数据，变为 4维张量[图像索引，y, x, c]
    # 去均值、做归一化，范围变到[-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 * NUM_CHANNELS)
        data = numpy.frombuffer(buf, dtype=numpy.uint8).astype(numpy.float32)
        data = (data - (PIXEL_DEPTH / 2.0)) / PIXEL_DEPTH
        data = data.reshape(num_images, IMAGE_SIZE, IMAGE_SIZE, NUM_CHANNELS)
        return data


def extract_labels(filename, num_images):
    """Extract the labels into a vector of int64 label IDs."""
    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):
    """Generate a fake dataset that matches the dimensions of MNIST."""
    """生成一个匹配MNIST数据集大小的伪造数据集"""
    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):
    """Return the error rate based on dense predictions and sparse labels."""
    return 100.0 - (
        100.0 *
        numpy.sum(numpy.argmax(predictions, 1) == labels) /
        predictions.shape[0])


# 主函数
def main(_):
    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:
        # 下载数据
        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')

        # 载入数据到numpy
        train_data = extract_data(train_data_filename, 60000)
        train_labels = extract_labels(train_labels_filename, 60000)
        test_data = extract_data(test_data_filename, 10000)
        test_labels = extract_labels(test_labels_filename, 10000)

        # 产生评测集
        validation_data = train_data[:VALIDATION_SIZE, ...]
        validation_labels = train_labels[:VALIDATION_SIZE]
        train_data = train_data[VALIDATION_SIZE:, ...]
        train_labels = train_labels[VALIDATION_SIZE:]
        num_epochs = NUM_EPOCHS # 数据全集一共过10遍网络
    train_size = train_labels.shape[0]

    # 训练样本和标签将从这里送入网络。
    # 每训练迭代步，占位符节点将被送入一个批处理数据
    # 训练数据节点
    train_data_node = tf.placeholder(
        data_type(),
        shape=(BATCH_SIZE, IMAGE_SIZE, IMAGE_SIZE, NUM_CHANNELS))
    # 训练标签节点
    train_labels_node = tf.placeholder(tf.int64, shape=(BATCH_SIZE,))
    # 评测数据节点
    eval_data = tf.placeholder(
        data_type(),
        shape=(EVAL_BATCH_SIZE, IMAGE_SIZE, IMAGE_SIZE, NUM_CHANNELS))
    # 下面这些变量是网络的可训练权值
    # conv1 权值维度为 32 x channels x 5 x 5, 32 为特征图数目
    conv1_weights = tf.Variable(
        tf.truncated_normal([5, 5, NUM_CHANNELS, 32],  # 5x5 filter, depth 32.
                            stddev=0.1,
                            seed=SEED, dtype=data_type()))
    # conv1 偏置
    conv1_biases = tf.Variable(tf.zeros([32], dtype=data_type()))
    # conv2 权值维度为 64 x 32 x 5 x 5
    conv2_weights = tf.Variable(tf.truncated_normal(
        [5, 5, 32, 64], stddev=0.1,
        seed=SEED, dtype=data_type()))
    # conv2 偏置
    conv2_biases = tf.Variable(tf.constant(0.1, shape=[64], dtype=data_type()))
    # 全连接层 fc1 权值，神经元数目为512
    fc1_weights = tf.Variable(  # fully connected, depth 512.
        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 权值，维度与标签类数目一致
    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()))

    # 两个网络：训练网络和评测网络
    # 它们共享权值
    # 实现 LeNet-5 模型，该函数输入为数据，输出为fc2的响应
    # 第二个参数区分训练网络还是评测网络
    def model(data, train=False):
        # 二维卷积，使用“不变形”补零（即输出特征图与输入尺寸一致）。
        conv = tf.nn.conv2d(data,
                            conv1_weights,
                            strides=[1, 1, 1, 1],
                            padding='SAME')
        # 加偏置、过激活函数一块完成
        relu = tf.nn.relu(tf.nn.bias_add(conv, conv1_biases))
        # 最大值下采样
        pool = tf.nn.max_pool(relu,
                              ksize=[1, 2, 2, 1],
                              strides=[1, 2, 2, 1],
                              padding='SAME')
        # 第二个卷积层，步长为1
        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')
        # 特征图变形为2维矩阵，便于送入全连接层
        pool_shape = pool.get_shape().as_list()
        reshape = tf.reshape(
            pool,
            [pool_shape[0], pool_shape[1] * pool_shape[2] * pool_shape[3]])
        # 全连接层，注意“+”运算自动广播偏置
        hidden = tf.nn.relu(tf.matmul(reshape, fc1_weights) + fc1_biases)
        # 训练阶段，增加 50% dropout；而评测阶段无需该操作
        if train:
            hidden = tf.nn.dropout(hidden, 0.5, seed=SEED)
        return tf.matmul(hidden, fc2_weights) + fc2_biases


        # 训练阶段计算： 对数+交叉熵 损失函数

    logits = model(train_data_node, True)
    loss = tf.reduce_mean(tf.nn.sparse_softmax_cross_entropy_with_logits(
        logits, train_labels_node))

    # 全连接层参数进行 L2 正则化
    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))
    # 将正则项加入损失函数
    loss += 5e-4 * regularizers

    # 优化器： 设置一个变量，每个批处理递增，控制学习速率衰减
    batch = tf.Variable(0, dtype=data_type())
    # Decay once per epoch, using an exponential schedule starting at 0.01.
    #  指数衰减
    learning_rate = tf.train.exponential_decay(
        0.01,  # 基本学习速率
        batch * BATCH_SIZE,  # 当前批处理在数据全集中的位置
        train_size,  # Decay step.
        0.95,  # 衰减率
        staircase=True)
    # Use simple momentum for the optimization.
    optimizer = tf.train.MomentumOptimizer(learning_rate,
                                           0.9).minimize(loss,
                                                         global_step=batch)

    # 用softmax 计算训练批处理的预测概率
    train_prediction = tf.nn.softmax(logits)

    # 用 softmax 计算评测批处理的预测概率
    eval_prediction = tf.nn.softmax(model(eval_data))

    # Small utility function to evaluate a dataset by feeding batches of data to
    # {eval_data} and pulling the results from {eval_predictions}.
    # Saves memory and enables this to run on smaller GPUs.
    def eval_in_batches(data, sess):
        """通过运行在小批量数据，得到所有预测结果."""
        size = data.shape[0]
        if size < EVAL_BATCH_SIZE:
            raise ValueError("batch size for evals larger than dataset: %d" % size)
        predictions = numpy.ndarray(shape=(size, NUM_LABELS), dtype=numpy.float32)
        for begin in xrange(0, size, EVAL_BATCH_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

    # Create a local session to run the training.
    start_time = time.time()
    with tf.Session() as sess:
        # 初始化操作准备参数
        tf.global_variables_initializer().run()
        print('Initialized!')
        # 循环训练
        for step in xrange(int(num_epochs * train_size) // BATCH_SIZE):
            # 计算当前minibatch的offset
            # Note that we could use better randomization across epochs.
            offset = (step * BATCH_SIZE) % (train_size - BATCH_SIZE)
            batch_data = train_data[offset:(offset + BATCH_SIZE), ...]
            batch_labels = train_labels[offset:(offset + BATCH_SIZE)]
            # This dictionary maps the batch data (as a numpy array) to the
            # node in the graph it should be fed to.
            feed_dict = {train_data_node: batch_data,
                         train_labels_node: batch_labels}
            # 运行优化器更新权重
            sess.run(optimizer, feed_dict=feed_dict)
            # print some extra information once reach the evaluation frequency
            if step % EVAL_FREQUENCY == 0:
                # fetch some extra nodes' data
                l, lr, predictions = sess.run([loss, learning_rate, train_prediction],
                                              feed_dict=feed_dict)
                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()
        # Finally print the result!
        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,)


if __name__ == '__main__':
    parser = argparse.ArgumentParser()
    parser.add_argument(
        '--use_fp16',
        default=False,
        help='Use half floats instead of full floats if True.',
        action='store_true')
    parser.add_argument(
        '--self_test',
        default=False,
        action='store_true',
        help='True if running a self test.')

    FLAGS, unparsed = parser.parse_known_args()
    tf.app.run(main=main, argv=[sys.argv[0]] + unparsed)