刚开始学习深度学习，使用 mnist 进行练习，刚开始自己定义了有3个卷积层（卷积+池化层），一个 fc 层，跑出的结果居然是50%，太低了，一脸懵逼，后来想到估计是自己 batch 的次数太少了（只有10 * 10次），想想也是，训练次数少（刚开始每个batch 是5000），每个 batch 会更新一次网络参数，100次肯定低到没朋友（没有达到拟合数据的数量级别），现在修改成了500。另外还有学习率的调参（现在使用的是0.01，尝试过0.02，0.05，其他参数不变的情况下，效果都不如0.01好），具体调参的话，后续打算用网格搜索法测试下。代码如下：

import tensorflow as tf
import argparse
from tensorflow.examples.tutorials.mnist import input_data
from sklearn import metrics
import numpy as np

class MnistCnnMethod:
    def __init__(self):
        self.x = tf.placeholder(tf.float32, [None, 28, 28, 1], name='x')
        self.y = tf.placeholder(tf.float32, [None, 10], name='y')
        self.keep_prob = tf.placeholder(tf.float32, name='keep_prob')
        self.cnn()

    def cnn(self):
        with tf.name_scope('init_weights'):
            w1 = tf.Variable(tf.random_normal([3, 3, 1, 32]), trainable=True, name='first_layer_w')
            w2 = tf.Variable(tf.random_normal([3, 3, 32, 32]), trainable=True, name='second_layer_w')
            w3 = tf.Variable(tf.random_normal([3, 3, 32, 64]), trainable=True, name='third_layer_w')
            w4 = tf.Variable(tf.random_normal([64*4*4, 1024]), trainable=True, name='full_connect_w')
            w_o = tf.Variable(tf.random_normal([1024, 10]), trainable=True, name='softmax_out_w')

        with tf.name_scope('first_layer'):
            lc1 = tf.nn.relu(tf.nn.conv2d(self.x, w1, strides=[1, 1, 1, 1], padding='SAME'))
            lp1 = tf.nn.max_pool(lc1, ksize=[1, 2, 2, 1],  strides=[1, 2, 2, 1], padding='SAME')
            lp1 = tf.nn.dropout(lp1, self.keep_prob)

        with tf.name_scope('second_layer'):
            lc2 = tf.nn.relu(tf.nn.conv2d(lp1, w2, strides=[1, 1, 1, 1], padding='SAME'))
            lp2 = tf.nn.max_pool(lc2, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')
            lp2 = tf.nn.dropout(lp2, self.keep_prob)

        with tf.name_scope('third_layer'):
            lc3 = tf.nn.relu(tf.nn.conv2d(lp2, w3, strides=[1, 1, 1, 1], padding='SAME'))
            lp3 = tf.nn.max_pool(lc3, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')
            lp3 = tf.reshape(lp3, [-1, w4.get_shape().as_list()[0]])
            lp3 = tf.nn.dropout(lp3, self.keep_prob)

        with tf.name_scope('full_connect'):
            lc4 = tf.nn.relu(tf.matmul(lp3, w4))
            ld4 = tf.nn.dropout(lc4, self.keep_prob)
            pyx = tf.matmul(ld4, w_o)

        with tf.name_scope('predict'):
            self.predict = tf.argmax(pyx, 1)

        with tf.name_scope('loss'):
            entropy = tf.nn.softmax_cross_entropy_with_logits_v2(logits=pyx, labels=self.y)
            cost = tf.reduce_mean(entropy, name='cost')
            self.train_step = tf.train.AdamOptimizer(learning_rate=0.01).minimize(cost, var_list=tf.trainable_variables())
            tf.summary.scalar('cost', cost)

        with tf.name_scope('summary'):
            self.merged = tf.summary.merge_all()




def train(mnistD, model):
    batch_size = 256
    keep_prob = 0.9
    with tf.Session() as sess:
        saver = tf.train.Saver()
        writer = tf.summary.FileWriter('./log', sess.graph)
        sess.run(tf.global_variables_initializer())
        train_loops = 500
        j = 0
        for i in range(train_loops):
            print(j)
            j += 1
            batch_x, batch_y = mnistD.train.next_batch(batch_size)
            batch_x = np.reshape(batch_x, [-1, 28, 28, 1])
            if i == train_loops - 1:
                sess.run(model.train_step, feed_dict={model.x: batch_x, model.y: batch_y, model.keep_prob: keep_prob})
                saver.save(sess, 'model/model.ckpt')
            else:
                summary, _ = sess.run([model.merged, model.train_step],
                                      feed_dict={model.x: batch_x, model.y: batch_y, model.keep_prob: keep_prob})
                writer.add_summary(summary, i)


def test(mnistD, model):
    with tf.Session() as sess:
        batch_x = mnistD.test.images
        batch_y = mnistD.test.labels
        batch_y = [np.where(r==1)[0][0] for r in batch_y]
        tf.train.Saver().restore(sess=sess, save_path='model/model.ckpt')
        pre = sess.run(model.predict, feed_dict={model.x: np.reshape(batch_x, [-1, 28, 28, 1]), model.keep_prob: 1.})
        print(metrics.classification_report(batch_y, pre))


if __name__ == '__main__':
    parser = argparse.ArgumentParser()
    parser.add_argument(
        '--data_dir',
        type=str,
        default='./mnist',
        help='Directory for storing input data')
    parser.add_argument(
        '--type',
        type=str,
        default='train',
        help='input train or test'
    )
    FLAGS, _ = parser.parse_known_args()
    cnnModel = MnistCnnMethod()
    mnistData = input_data.read_data_sets(FLAGS.data_dir, one_hot=True)
    if FLAGS.type == 'train':
        train(mnistData, cnnModel)
    else:
        test(mnistData, cnnModel)

cnn 结构：

这里面 numpy 居然也有 reshape 功能，汗，不知道。