关注公众号“从机器学习到深度学习那些事”获取更多最新资料
写在前面
学习建议：以下学习过程中有不理解可以简单查找下资料，但不必纠结（比如非得深究某一个函数等），尽量快速的学一遍，不求甚解无妨。多实操代码，不能只复制代码，或者感觉懂了就只看。熟能生巧，我亦无他，唯手熟尔
今天介绍一些基础函数及其用法，基本全是代码，一些解释都放在代码的注释里了。直接看代码吧，记得在你本地跑一下看哦
代码1
#tensor.get_shape() 获取tensor的shape，就是维度这些 #tensor.get_shape().as_list()，把shape转换成列表形式 # 上面这两个函数不常直接用，在理解TensorFlow的用法等有帮助，就先列在这里了 # a = tf.constant([[1, 2, 3], [4, 5, 6]], dtype=tf.float32, name='a') with tf.Session() as sess: print(a.eval()) print("shape: ", a.get_shape(), ",type: ", type(a.get_shape())) print("shape: ", a.get_shape().as_list(), ",type: ", type(a.get_shape().as_list())) ''' 输出如下 [[1. 2. 3.] [4. 5. 6.]] shape: (2, 3) ,type: <class 'tensorflow.python.framework.tensor_shape.TensorShape'> shape: [2, 3] ,type: <class 'list'> '''
代码2
#tf.argmax #tf.argmax(input, dimension, name=None) returns the index with the largest value across dimensions of a tensor. # 上面注释是英文的，翻译下就是，tf.argmax()这个函数输入的张量中，沿着指定维度中最大的一个值的索引，可以这样理解，0就是行，1就是列，3等更大的就是沿着更好维度算，下面的例子理解下。 # 注意，返回的是索引号 a = tf.constant([[1, 6, 5], [2, 3, 4]]) with tf.Session() as sess: print(a.eval()) print("argmax over axis 0") print(tf.argmax(a, 0).eval()) print("argmax over axis 1") print(tf.argmax(a, 1).eval()) ''' [[1 6 5] [2 3 4]] argmax over axis 0 [1 0 0] argmax over axis 1 [1 2] '''
代码3
#tf.reduce_sum #tf.reduce_sum(input_tensor, reduction_indices=None, keep_dims=False, name=None) computes the sum of elements across dimensions of a tensor. Unless keep_dims is true, the rank of the tensor is reduced by 1 for each entry in reduction_indices. If keep_dims is true, the reduced dimensions are retained with length 1. If reduction_indices has no entries, all dimensions are reduced, and a tensor with a single element is returned # 大概翻译下，就是求和，计算张量的各维度上的元素的和。还有两个参数， keep_dims和reduction_indices，下面代码执行并理解下 a = tf.constant([[1, 1, 1], [2, 2, 2]]) with tf.Session() as sess: print(a.eval()) print("reduce_sum over entire matrix") print(tf.reduce_sum(a).eval()) print("reduce_sum over axis 0") print(tf.reduce_sum(a, 0).eval()) print("reduce_sum over axis 0 + keep dimensions") print(tf.reduce_sum(a, 0, keep_dims=True).eval()) print("reduce_sum over axis 1") print(tf.reduce_sum(a, 1).eval()) print("reduce_sum over axis 1 + keep dimensions") print(tf.reduce_sum(a, 1, keep_dims=True).eval()) ''' 输出： [[1 1 1] [2 2 2]] reduce_sum over entire matrix 9 reduce_sum over axis 0 [3 3 3] reduce_sum over axis 0 + keep dimensions [[3 3 3]] reduce_sum over axis 1 [3 6] reduce_sum over axis 1 + keep dimensions [[3] [6]] '''
代码4
''' tf.get_variable,是用来获取或者创建一个变量的函数，它需要用一个初始化函数，根据给的shape初始化相同shape的tensor。初始化函数有很多种，前面一篇中也有介绍过，各种随机初始化，常数初始化等， # 看下面例子 tf.get_variable(name, shape=None, dtype=None, initializer=None, trainable=True) is used to get or create a variable instead of a direct call to tf.Variable. It uses an initializer instead of passing the value directly, as in tf.Variable. An initializer is a function that takes the shape and provides a tensor with that shape. Here are some initializers available in TensorFlow: •tf.constant_initializer(value) initializes everything to the provided value, •tf.random_uniform_initializer(a, b) initializes uniformly from [a, b], •tf.random_normal_initializer(mean, stddev) initializes from the normal distribution with the given mean and standard deviation. ''' my_initializer = tf.random_normal_initializer(mean=0, stddev=0.1) v = tf.get_variable('v', shape=[2, 3], initializer=my_initializer) with tf.Session() as sess: tf.initialize_all_variables().run() print(v.eval()) """ [[ 0.14729649 -0.07507571 -0.00038549] [-0.02985961 -0.01537443 0.14321376]] """
代码5
my_initializer = tf.random_normal_initializer(mean=0, stddev=0.1) v = tf.get_variable('v', shape=[2, 3], initializer=my_initializer) # tf.variable_scope 这个函数是管理变量的命名空间的，这个也是方便tensorboard中能可视化，使模型逻辑流程等更容易理解等 # tf.variable_scope(scope_name) manages namespaces for names passed to tf.get_variable. with tf.variable_scope('layer1'): w = tf.get_variable('v', shape=[2, 3], initializer=my_initializer) print(w.name) with tf.variable_scope('layer2'): w = tf.get_variable('v', shape=[2, 3], initializer=my_initializer) print(w.name) ''' layer1/v:0 layer2/v:0 '''
代码6
''' reuse_variables 上面的代码只能运行一次，就是因为这个reuse_variables， 用这个scope.reuse_variables()来获取（重用）之前创建的变量，而不是再创建一个新的 Note that you should run the cell above only once. If you run the code above more than once, an error message will be printed out: "ValueError: Variable layer1/v already exists, disallowed.". This is because we used tf.get_variable above, and this function doesn't allow creating variables with the existing names. We can solve this problem by using scope.reuse_variables() to get preivously created variables instead of creating new ones. ''' my_initializer = tf.random_normal_initializer(mean=0, stddev=0.1) v = tf.get_variable('v', shape=[2, 3], initializer=my_initializer) with tf.variable_scope('layer1'): w = tf.get_variable('v', shape=[2, 3], initializer=my_initializer) # print(w.name) with tf.variable_scope('layer2'): w = tf.get_variable('v', shape=[2, 3], initializer=my_initializer) # print(w.name) with tf.variable_scope('layer1', reuse=True): w = tf.get_variable('v') # Unlike above, we don't need to specify shape and initializer print(w.name) # or with tf.variable_scope('layer1') as scope: scope.reuse_variables() w = tf.get_variable('v') print(w.name)
小结
今天主要是贴代码了，没有过多的文字解释，一些解释也尽量放在代码注释中，这样在本地自己电脑上运行也方便理解。初学的话代码还是要自己敲一遍，自己敲的过程中肯定会遇到一些问题，这就是反复学习的过程，还是不要省。
今天代码中的英文依然没有抹去，尽量看一下，顺便学下英语也不错。主要是因为有些英文文档不好翻译过来反而不好理解，相信这种情况以后也会遇到。所以可以先学点。
本篇完
