Question

我正在尝试在keras中实现一个丢失函数，例如下面的伪代码

for i in range(N):
    for j in range(N):
        sum += some_calculations

但我读到张量流不支持循环，因此我从here了解了while_loop（cond，body，loop_vars）函数

我理解了while循环的基本工作，因此我实现了以下代码：

def body1(i):
    global data
    N = len(data)*positive_samples     //Some length
    j = tf.constant(0)    //iterators
    condition2 = lambda j, i :tf.less(j, N)   //one condition only j should be less than N
    tf.add(i, 1)   //increment previous index i
    result = 0

    def body2(j, i):
        global similarity_matrix, U, V
        result = (tf.transpose(U[:, i])*V[:, j])   //U and V are 2-d tensor Variables and here only a column is extracted and their final product is a single value
        return result

    tf.while_loop(condition2, body2, loop_vars=[j, i])
    return result


def loss_function(x):
    global data
    N = len(data)*positive_samples
    i = tf.constant(0)
    condition1 =  lambda i : tf.less(i, N)
    return tf.while_loop(condition1, body1, [i])

但是当我运行此代码时，我收到错误

ValueError: The two structures don't have the same number of elements. First structure: [<tf.Tensor 'lambda_1/while/while/Identity:0' shape=() dtype=int32>, <tf.Tensor 'lambda_1/while/while/Identity_1:0' shape=() dtype=int32>], second structure: [0]

Answer 1

tf.while_loop使用起来可能很棘手，请务必仔细阅读文档。正文的返回值必须与循环变量具有相同的结构，tf.while_loop操作的返回值是变量的最终值。为了进行计算，您应该传递一个额外的循环变量来存储部分结果。你可以这样做：

def body1(i, result):
    global data
    N = len(data) * positive_samples
    j = tf.constant(0)
    condition2 = lambda j, i, result: tf.less(j, N)
    result = 0

    def body2(j, i, result):
        global similarity_matrix, U, V
        result_j = (tf.transpose(U[:, i]) * V[:, j])
        return j + 1, i, result + result_j

    j, i, result = tf.while_loop(condition2, body2, loop_vars=[j, i, result])
    return i + 1, result

def loss_function(x):
    global data
    N = len(data)*positive_samples
    i = tf.constant(0)
    result = tf.constant(0, dtype=tf.float32)
    condition1 = lambda i, result: tf.less(i, N)
    i, result = tf.while_loop(condition1, body1, [i, result])
    return result

您的代码中不清楚要使用x。但是，在这种情况下，操作的结果应该等于：

result = tf.reduce_sum(tf.transpose(U) @ V)
# Equivalent
result = tf.reduce_sum(tf.matmul(tf.transpose(U), V))

哪个也会快得多。

Answer 2

这是TensorFlow中使用tf.while_loop的嵌套循环的另一个示例。在这一张图中，张量x的第i个元素通过张量v的第i个元素给出的次数进行迭代连接。

import tensorflow as tf
x = tf.Variable([[1,1],[2,2],[3,3]])
v = tf.constant([1,2,3])
i = tf.constant(0)
a_combined = tf.zeros([0, 2], dtype=tf.int32)


def body(x,v,i,a_combined):
    x_slice = tf.slice(x,[i,0], [1, x.shape[1]])
    v_slice = tf.slice(v,[i],[1])
    j = tf.constant(0)
    b_combined = tf.zeros([0, 2], dtype=tf.int32)
    
    print("i: ", i)
    
    def body_supp(x_slice,v_slice,j, b_combined):
        
        print("j: ", j)
        
        j = tf.add(j,1)
        b_combined = tf.concat([b_combined,x_slice],0)
        return x_slice, v_slice, j, b_combined 
    
    while_condition_supp = lambda x_slice, v_slice, j, b_combined: tf.less(j, v_slice)
    
    x_slice, v_slice, j, b_combined = tf.while_loop(while_condition_supp, body_supp, [x_slice, v_slice, j, b_combined])
    
    i = tf.add(i,1)

    a_combined = tf.concat([a_combined,b_combined],0)
    return x, v, i, a_combined

while_condition = lambda x, v, i, a_combined: i < v.shape[0]  

x, v, i, a_combined = tf.while_loop(while_condition, body, [x, v, i, a_combined])

a_combined

输出看起来像这样：

<tf.Tensor: shape=(6, 2), dtype=int32, numpy=
array([[1, 1],
       [2, 2],
       [2, 2],
       [3, 3],
       [3, 3],
       [3, 3]], dtype=int32)>

在tensorflow中嵌套while循环

2 个答案: