tensorflow编程: Layers (contrib)

Higher level ops for building neural network layers

tf.contrib.layers.batch_norm

  添加一个 Batch Normalization 层 。

tf.contrib.layers.batch_norm (inputs, decay=0.999, updates_collections=tf.GraphKeys.UPDATE_OPS, is_training=True, data_format=DATA_FORMAT_NHWC)

  可用作conv2d和fully_connected的归一化函数。

tf.nn.conv2d_transpose

   conv2d 的转置

tf.conv2d_transpose (value, filter, output_shape, strides, padding=’SAME’, data_format=’NHWC’, name=None)

# -*- coding: utf-8 -*-import tensorflow as tfdef func(in_put, in_channel, out_channel):    weights = tf.get_variable(name="weights", shape=[2, 2, in_channel, out_channel],                              initializer=tf.contrib.layers.xavier_initializer_conv2d())    convolution = tf.nn.conv2d(input=in_put, filter=weights, strides=[1, 1, 1, 1], padding='VALID')    conv_shape = convolution.get_shape().as_list()    deconv_shape = [conv_shape[0], conv_shape[1]*2, conv_shape[2]*2, conv_shape[3]]    deconvolution = tf.nn.conv2d_transpose(value=convolution, filter=weights, output_shape=deconv_shape, strides=[1, 2, 2, 1], padding='VALID')    return in_put, convolution, deconvolutiondef main():    with tf.Graph().as_default():        input_x = tf.placeholder(dtype=tf.float32, shape=[1, 4, 4, 1])        in_put, convolution, deconvolution = func(input_x, 1, 1)        with tf.Session() as sess:            sess.run(tf.global_variables_initializer())            import numpy as np            _in_put, _convolution, _deconvolution = sess.run([in_put, convolution, deconvolution], feed_dict={input_x:np.random.uniform(low=0, high=255, size=[1, 4, 4, 1])})            print '\nin_put:'            print in_put            # print _in_put            print '\nconvolution:'            print convolution            # print _convolution            print '\ndeconvolution:'            print deconvolution            # print _deconvolutionif __name__ == "__main__":    main()

2017-09-29 09:51:41.472842: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1052] Creating TensorFlow device (/device:GPU:0) -> (device: 0, name: GeForce GTX 1070, pci bus id: 0000:01:00.0, compute capability: 6.1)in_put:Tensor("Placeholder:0", shape=(1, 4, 4, 1), dtype=float32)convolution:Tensor("Conv2D:0", shape=(1, 3, 3, 1), dtype=float32)deconvolution:Tensor("conv2d_transpose:0", shape=(1, 6, 6, 1), dtype=float32)Process finished with exit code 0

tf.nn.dropout

tf.nn.dropout (x, keep_prob, noise_shape=None, seed=None, name=None)

# coding=utf-8import tensorflow as tfdef main():    with tf.Graph().as_default():        import numpy as np        input_x = np.random.uniform(0, 255, [3, 3])        print input_x        drop = [0, 0, 0]        for i, keep_prob in enumerate([0.1, 0.5, 1.0]):            drop[i] = tf.nn.dropout(x=input_x, keep_prob=keep_prob)        with tf.Session() as sess:            sess.run(tf.global_variables_initializer())            for drop_i in drop:                _drop_i = sess.run(drop_i)                print '\n----------\n'                print _drop_iif __name__ == "__main__":    main()

# 初始输入[[  16.46278229  253.27597997  246.33614039] [ 130.45261984  227.85971767  142.72621045] [ 173.23025953  165.99906514  180.13238617]]2017-09-29 11:02:29.146976: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1052] Creating TensorFlow device (/device:GPU:0) -> (device: 0, name: GeForce GTX 1070, pci bus id: 0000:01:00.0, compute capability: 6.1)----------# keep_prob = 0.1[[    0.            0.            0.       ] [    0.            0.            0.       ] [    0.            0.         1801.3238617]]----------# keep_prob = 0.5[[  32.92556457  506.55195994    0.        ] [ 260.90523969  455.71943533    0.        ] [ 346.46051906    0.          360.26477234]]----------# keep_prob = 1.0[[  16.46278229  253.27597997  246.33614039] [ 130.45261984  227.85971767  142.72621045] [ 173.23025953  165.99906514  180.13238617]]

tf.contrib.layers.fully_connected

tf.contrib.layers.fully_connected (inputs, num_outputs, activation_fn=tf.nn.relu)

• 其中默认进行了 convolution 和 activation 。
• convolution 中，只对输入的最后一维求平均，且阶数不变。即 ‘weights:0’.shape=[inputs.shape[-1], num_outputs]）。
• ‘weights:0’.shape 永远是二维的。
• num_outputs 是 ‘weights:0’ 第二维（即第-1维）的参数值；经过fn计算后，也变成了 结果输出的tensor 的最后一维（即第-1维）的参数值。
• 如果设置 activation_fn=None，则输出结果 不经过激活，依然可能包含负数值。

# coding=utf-8import tensorflow as tffrom tensorflow.contrib.layers import fully_connecteddef main():    with tf.Graph().as_default():        import numpy as np        input_x = np.random.uniform(0, 10, [3, 3])        print input_x        fn = fully_connected(inputs=input_x, num_outputs=1)        with tf.Session() as sess:            sess.run(tf.global_variables_initializer())            _fn = sess.run(fn)            print _fn            print '\n----------\n'            for (x, y) in zip(tf.global_variables(), sess.run(tf.global_variables())):                print '\n', x, '\n', yif __name__ == "__main__":    main()

# 原始输入矩阵[[ 7.73305319  0.2780667   7.27101124] [ 0.84666041  0.92980727  6.83676724] [ 1.02844109  5.51824496  1.78840816]]2017-09-29 11:33:02.500942: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1052] Creating TensorFlow device (/device:GPU:0) -> (device: 0, name: GeForce GTX 1070, pci bus id: 0000:01:00.0, compute capability: 6.1)# fn后输出的结果tensor[[ 2.32549239] [ 1.69284669] [ 0.        ]]----------<tf.Variable 'fully_connected/weights:0' shape=(3, 1) dtype=float64_ref> [[-0.01048241] [-0.83954232] [ 0.36308597]]<tf.Variable 'fully_connected/biases:0' shape=(1,) dtype=float64_ref> [ 0.]Process finished with exit code 0

# coding=utf-8import tensorflow as tffrom tensorflow.contrib.layers import fully_connecteddef main():    with tf.Graph().as_default():        import numpy as np        input_x = np.random.uniform(0, 10, [2, 4, 4, 3])        print np.shape(input_x)        fn = fully_connected(inputs=input_x, num_outputs=1)        with tf.Session() as sess:            sess.run(tf.global_variables_initializer())            _fn = sess.run(fn)            print np.shape(_fn)            print '\n----------\n'            for i in tf.global_variables():                print '\n', iif __name__ == "__main__":    main()

(2, 4, 4, 3)2017-09-29 11:46:17.248114: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1052] Creating TensorFlow device (/device:GPU:0) -> (device: 0, name: GeForce GTX 1070, pci bus id: 0000:01:00.0, compute capability: 6.1)# convolution 中，只对输入的最后一维求平均，且阶数不变。即 'weights:0'.shape=[inputs.shape[-1], num_outputs]）。(2, 4, 4, 1)----------<tf.Variable 'fully_connected/weights:0' shape=(3, 1) dtype=float64_ref><tf.Variable 'fully_connected/biases:0' shape=(1,) dtype=float64_ref>

tf.nn.relu

max(features, 0)

tf.nn.relu(features, name=None)

tf.nn.relu6

min(max(features, 0), 6)。即对 tf.nn.relu 的优化，防止 relu过后 某些 极端值 依然 大于6

tf.nn.relu6(features, name=None)

tf.nn.softmax

计算公式： softmax = exp(logits) / reduce_sum(exp(logits), dim)

tf.nn.softmax (logits, dim=-1, name=None)

# coding=utf-8import tensorflow as tfimport numpy as npinput_x = tf.constant(np.random.uniform(0, 5, [2, 3]))softmax = tf.nn.softmax(logits=input_x)# 自己写的softmax接口def my_softmax(logits, dim=-1):    my_softmax = tf.div(tf.exp(logits), tf.reduce_sum(tf.exp(logits), dim, keep_dims=True))    return my_softmaxwith tf.Session() as sess:    sess.run(tf.global_variables_initializer())    my_softmax = my_softmax(logits=input_x)    _input_x, _softmax, _my_softmax = sess.run([input_x, softmax, my_softmax])    print 'input:', np.shape(input_x), input_x, '\n', _input_x    print '\n----------\n'    print 'softmax:', np.shape(_softmax), softmax, '\n', _softmax    print '\n----------\n'    print 'my_softmax:', np.shape(_my_softmax), my_softmax, '\n', _my_softmax    print '\n----------\n'    for i in tf.global_variables():        print '\n', i

# softmax输入类型为tensor型input: (2, 3) Tensor("Const:0", shape=(2, 3), dtype=float64) [[ 3.88517858  3.69402461  3.07837121] [ 1.27162028  2.12622856  4.34646188]]----------# softmax后，shape保持不变，返回tensor型softmax: (2, 3) Tensor("Softmax:0", shape=(2, 3), dtype=float64) [[ 0.44008545  0.36351295  0.1964016 ] [ 0.04000495  0.09402978  0.86596527]]----------# 根据计算公式：softmax = exp(logits) / reduce_sum(exp(logits), dim) 得到了相同的接口输出my_softmax: (2, 3) Tensor("div:0", shape=(2, 3), dtype=float64) [[ 0.44008545  0.36351295  0.1964016 ] [ 0.04000495  0.09402978  0.86596527]]----------# 内存中无参数保存Process finished with exit code 0

Regularizers

tf.contrib.layers.l1_regularizer

tf.contrib.layers.l1_regularizer (scale, scope=None)

tf.contrib.layers.l2_regularizer

tf.contrib.layers.l2_regularizer (scale, scope=None)

Initializers

tf.contrib.layers.xavier_initializer

执行“Xavier”初始化的初始化程序。

tf.contrib.layers.xavier_initializer (uniform=True, seed=None, dtype=tf.float32)

# coding=utf-8import tensorflow as tfxavier = tf.get_variable(name="weights",                         shape=[2, 2],                         initializer=tf.contrib.layers.xavier_initializer())constant = tf.get_variable(name='biases',                         shape=[2],                         initializer=tf.constant_initializer())with tf.Session() as sess:    sess.run(tf.global_variables_initializer())    _xavier, _constant = sess.run([xavier, constant])    print '\n\nxavier:'    print xavier    print _xavier    print '\n\nconstant:'    print constant    print _constant

xavier:<tf.Variable 'weights:0' shape=(2, 2) dtype=float32_ref>[[ 1.20015538  0.34742999] [ 0.39075291  0.60076308]]constant:<tf.Variable 'biases:0' shape=(2,) dtype=float32_ref>[ 0.  0.]

import tensorflow as tfprint '\n\ntf.contrib.layers.xavier_initializer_conv2d() :\n', tf.contrib.layers.xavier_initializer_conv2d()print '\n\ntf.constant_initializer() :\n', tf.constant_initializer()print '\n\ntf.global_variables_initializer() :\n', tf.global_variables_initializer()

tf.contrib.layers.xavier_initializer_conv2d() :<function _initializer at 0x7fe5133da578>tf.constant_initializer() :<tensorflow.python.ops.init_ops.Constant object at 0x7fe528bbdfd0>tf.global_variables_initializer() :name: "init"op: "NoOp"

Optimization

Summaries

Feature columns

---

---