Question

我有一个包含三个总和的损失函数如下：

sum for each j in words:

   sum for each i in documents:

        sum for each r in Tj

我实现这个问题的原因是Tj是一个列表列表，其中每个列表的长度不同。我试图将它作为一个numpy函数实现，然后使用tf.py_func，但现在我不知道如何优化它。

def loss_fun（COM，Wj，Wi，B，R，Tj）：

def loss_factor(j):

    L = R.shape[0]

    sum_container = []
    for l in range(L):     

       third_sum  = 0
       for i in range(len(Tj[j])):     

           entity_value = Tj[j][i]
           wi = np.reshape(Wi[entity_value], [1, 300])
           Rl = R[l,:,:]
           wj = np.reshape(Wj[j,:], [1, 300])

           wiRl = tf.matmul(wi, Rl)
           wjRl = tf.matmul(wj, Rl)

           wiRlxwjRl = np.matmul(wiRl, np.transpose(wjRl))
           pmi = COM[entity_value, j]
           bj = B[j,0]

           distance_expr = np.square( (wiRlxwjRl + bj - pmi) )
           third_sum = third_sum + distance_expr

       sum_container.append(third_sum)

    return(np.amin(sum_container) / len(Tj[j]))



L = R.shape[0]
M = B.shape[0]
p = -2

first_sum  = 0
second_sum = 0
third_sum  = 0

for j in range(M):

     sigma = loss_factor(j)
     sigma = 1/sigma

     second_sum = 0
     for l in range(L):    

       third_sum  = 0
       for i in range(len(Tj[j])):     

           entity_value = Tj[j][i]
           wi = np.reshape(Wi[entity_value], [1, 300])
           Rl = R[l,:,:]
           wj = np.reshape(Wj[j,:], [1, 300])

           wiRl = tf.matmul(wi, Rl)
           wjRl = tf.matmul(wj, Rl)

           wiRlxwjRl = np.matmul(wiRl, np.transpose(wjRl))
           pmi = COM[entity_value, j]
           bj = B[j,0]

           distance_expr = np.square( (wiRlxwjRl + bj - pmi) )
           third_sum = third_sum + distance_expr

        second_sum = second_sum + np.power(third_sum, p)

    first_sum = first_sum + sigma * second_sum


nuc_norm = 0
lambdaa = 0.1
for l in range(L):        
    Rl = R[l,:,:]
    nuc_norm = nuc_norm + np.linalg.norm(Rl, ord = 'nuc')

nuc_norm = lambdaa * nuc_norm

return (first_sum + nuc_norm)

######损失函数结束

K = 50

COM = tf.constant(co_occurrence_matrix, name = 'COM')
Tj  = tf.constant(t_j, name = 'Tj')
Wj  = tf.Variable(Wj, name = 'Wi')

Wi  = tf.Variable(tf.random_uniform([len(entities_list), 300], 1.0, -1.0), name = 'Wi')    
B   = tf.Variable(tf.random_uniform([len(features_list)-1, 1], 1.0, -1.0), name = 'B')
R   = tf.Variable(tf.random_normal(shape = (K, 300, 300)), name = 'R')


loss_val = tf.py_func(loss_fun, [COM, Wj, Wi, B, R, Tj], tf.float32)

optimizer = tf.train.AdagradOptimizer(0.01).minimize(loss_val)

我会欣赏任何想法

使用Tensorflow优化函数

0 个答案: