4-CNN-demo-02-激活函数和dropout使用

最新推荐文章于 2024-07-26 22:21:09 发布

HJZ11

最新推荐文章于 2024-07-26 22:21:09 发布

阅读量1.1k

点赞数

分类专栏： # 深度学习4-CNN

本文链接：https://blog.csdn.net/HJZ11/article/details/103743183

版权

深度学习4-CNN 专栏收录该内容

21 篇文章 0 订阅

订阅专栏

在这里插入图片描述



import tensorflow as tf


"""
学习relu激活函数 和 dropout的使用
"""

tf.set_random_seed(43)

def activation_func():
    with tf.Graph().as_default():
        hidden_layer_weights = tf.truncated_normal(shape=[4, 3], stddev=1.0)
        output_weights = tf.truncated_normal(shape=[3, 2], stddev=1.0)
        weights = [
            tf.get_variable('w1', initializer=hidden_layer_weights),
            tf.get_variable('w2', initializer=output_weights)
        ]
        biases = [
            tf.get_variable('b1', initializer=tf.zeros(3)),
            tf.get_variable('b2', initializer=tf.zeros(2))
        ]
        input_x = tf.constant(value=[[1, 2, 3, 4],
                                     [-1, -23, -32, -12],
                                     [11, 12, 2, 3]], dtype=tf.float32)
        hidden_input = tf.matmul(input_x, weights[0]) + biases[0]
        activate_out = tf.nn.relu(hidden_input)
        relu6_out = tf.nn.relu6(hidden_input)

        # Leaky relu（自己实现）
        alpha = 0.2
        leaky_relu_out = tf.maximum(hidden_input, hidden_input*alpha)
        # tf.nn.leaky_relu()   渗漏relu激活函数的api。

        final_output = tf.matmul(activate_out, weights[1]) + biases[1]

        with tf.Session() as sess:
            sess.run(tf.global_variables_initializer())
            hidden_input_, activate_out_, relu6_out_, leaky_relu_out_ = sess.run(
                [hidden_input, activate_out, relu6_out, leaky_relu_out]
            )
            print('激活之前的值为:{}'.format(hidden_input_),
                  '\n',
                  'relu 激活值为:{}'.format(activate_out_),
                  '\n',
                  'relu6 激活值为:{}'.format(relu6_out_),
                  '\n',
                  'leaky relu激活值为:{}'.format(leaky_relu_out_))


def dense(inputs, units, name, activation=None, use_bias=True, trainable=True):
    """
            tf.layers.dense(
                inputs,      # 输入的tensor
                units,       # 隐藏层的节点数量（神经元的个数。）
                activation=None,  # 是否使用激活函数，None 代表不用激活函数
                use_bias=True,    # 是否创建偏置项
                kernel_initializer=None,  # 权重的初始化器
                bias_initializer=init_ops.zeros_initializer(),  # 偏置项初始化器
                kernel_regularizer=None,
                bias_regularizer=None,
                activity_regularizer=None,
                kernel_constraint=None,
                bias_constraint=None,
                trainable=True,    # 参数（权重和偏置项）是否可以被训练。
                name=None,         # 名字
                reuse=None):       # 参数是否重用
            """
    shape = inputs.get_shape()
    with tf.variable_scope(name):
        weights = tf.get_variable(
            'w', shape=[shape[-1], units],
            initializer=tf.truncated_normal_initializer(stddev=0.1),
            trainable=trainable
        )
        fc_out = tf.matmul(inputs, weights)
        if use_bias:
            bias = tf.get_variable(
                'b', shape=[units],
                initializer=tf.zeros_initializer(),
                trainable=trainable
            )
            fc_out = fc_out + bias
        if activation:
            return activation(fc_out)
        return fc_out


def dense_func():
    with tf.Graph().as_default():
        hidden_layer_weights = tf.truncated_normal(shape=[4, 3], stddev=1.0)
        output_weights = tf.truncated_normal(shape=[3, 2], stddev=1.0)
        weights = [
            tf.get_variable('w1', initializer=hidden_layer_weights),
            tf.get_variable('w2', initializer=output_weights)
        ]
        biases = [
            tf.get_variable('b1', initializer=tf.zeros(3)),
            tf.get_variable('b2', initializer=tf.zeros(2))
        ]
        input_x = tf.constant(value=[[1, 2, 3, 4],
                                     [-1, -23, -32, -12],
                                     [11, 12, 2, 3]], dtype=tf.float32)
        hidden_input = tf.matmul(input_x, weights[0]) + biases[0]
        hidden_out = tf.nn.relu(hidden_input)
        """
        tf.layers.dense(
            inputs,      # 输入的tensor
            units,       # 隐藏层的节点数量（神经元的个数。）
            activation=None,  # 是否使用激活函数，None 代表不用激活函数
            use_bias=True,    # 是否创建偏置项
            kernel_initializer=None,  # 权重的初始化器
            bias_initializer=init_ops.zeros_initializer(),  # 偏置项初始化器
            kernel_regularizer=None,
            bias_regularizer=None,
            activity_regularizer=None,
            kernel_constraint=None,
            bias_constraint=None,
            trainable=True,    # 参数（权重和偏置项）是否可以被训练。
            name=None,         # 名字
            reuse=None):       # 参数是否重用
        """
        hidden_out1 = tf.layers.dense(
            input_x, units=3, use_bias=True, activation=tf.nn.relu
        )
        hidden_out2 = dense(
            input_x, units=3, name='dense', activation=None,
            use_bias=True, trainable=True)

        logits = tf.matmul(hidden_out, weights[1]) + biases[1]

        with tf.Session() as sess:
            sess.run(tf.global_variables_initializer())
            hidden_out_, hidden_out1_, hidden_out2_, final_output_ = sess.run(
                [hidden_out, hidden_out1, hidden_out2, logits]
            )
            print(hidden_out_, hidden_out1_)


def drop_out(x, keep_prob):
    """
    自己实现dropout函数。
    :param x:
    :param keep_prob:
    :return:
    """
    x_shape = x.get_shape()
    k_prob = tf.cast(tf.random_uniform(shape=x_shape, maxval=1) < keep_prob, tf.float32) / keep_prob
    x = x * k_prob
    return x


def dropout_func():
    with tf.Graph().as_default():
        hidden_layer_weights = tf.truncated_normal(shape=[4, 3], stddev=1.0)
        output_weights = tf.truncated_normal(shape=[3, 2], stddev=1.0)
        weights = [
            tf.get_variable('w1', initializer=hidden_layer_weights),
            tf.get_variable('w2', initializer=output_weights)
        ]
        biases = [
            tf.get_variable('b1', initializer=tf.zeros(3)),
            tf.get_variable('b2', initializer=tf.zeros(2))
        ]
        input_x = tf.constant(value=[[1, 2, 3, 4],
                                     [-1, -23, -32, -12],
                                     [11, 12, 2, 3]], dtype=tf.float32)
        hidden_input = tf.matmul(input_x, weights[0]) + biases[0]

        # tf.nn.leaky_relu()   渗漏relu激活函数的api。
        hidden_output = tf.nn.leaky_relu(hidden_input)
        # fixme dropout要在激活函数后面使用。
        hidden_output_drop = tf.nn.dropout(hidden_output, keep_prob=0.6)

        # 调用自己实现的函数。
        hidden_output_drop1 = drop_out(hidden_output, keep_prob=0.6)

        final_output = tf.matmul(hidden_output_drop, weights[1]) + biases[1]

        with tf.Session() as sess:
            sess.run(tf.global_variables_initializer())
            hidden_output_, hidden_output_drop_, hidden_output_drop1_ = sess.run(
                [hidden_output, hidden_output_drop, hidden_output_drop1]
            )
            print('leaky relu激活值为:{}'.format(hidden_output_),
                  '\n',
                  'tf.nn.dropout之后的值为:{}'.format(hidden_output_drop_),
                  '\n',
                  '自己实现的dropout之后的值为:{}'.format(hidden_output_drop1_))


if __name__ == '__main__':
    activation_func()
    dense_func()
    dropout_func()

激活之前的值为:[[  2.859401    2.0733218  -2.0179286]
 [ -5.016388   -0.7504692  -1.0032654]
 [ 19.776606  -11.05093   -22.942028 ]] 
 relu 激活值为:[[ 2.859401   2.0733218  0.       ]
 [ 0.         0.         0.       ]
 [19.776606   0.         0.       ]] 
 relu6 激活值为:[[2.859401  2.0733218 0.       ]
 [0.        0.        0.       ]
 [6.        0.        0.       ]] 
 leaky relu激活值为:[[ 2.859401    2.0733218  -0.40358573]
 [-1.0032777  -0.15009384 -0.20065308]
 [19.776606   -2.210186   -4.5884056 ]]
[[ 2.6600227 11.10044    0.       ]
 [13.155464   0.        28.872334 ]
 [ 0.        16.034973   0.       ]] [[ 0.         3.531556   0.4266281]
 [ 3.4427366  0.        18.767883 ]
 [ 0.         7.7806087  0.       ]]
leaky relu激活值为:[[ 3.4195879 -0.524014   2.676589 ]
 [ 6.741766   8.814518  11.51157  ]
 [-4.762424   4.847961   8.176558 ]] 
 tf.nn.dropout之后的值为:[[ 5.699313   -0.87335664  4.4609814 ]
 [11.236277   14.690863   19.18595   ]
 [-7.937373    8.079935   13.627596  ]] 
 自己实现的dropout之后的值为:[[ 5.699313   -0.87335664  0.        ]
 [11.236277   14.690863   19.18595   ]
 [-0.          0.         13.627596  ]]

Process finished with exit code 0