文本生成中的 mask CrossEntropy -Tensorflow

文本生成中的真实标签为 index，shape = [n_batch,]
输出的为概率分布，shape = [n_batch, num_decoder_tokens]
计算 loss 时应 mask 序列 padding部分

def categorical_CE_generation(max_len_char,n_decoder_tokens,idx_pad=0):
    def mask_categorical_crossentropy(y_true,y_pred):
        epsilon = K.epsilon()
        y_pred = K.clip(y_pred, epsilon, 1. - epsilon)#避免 loss 为nan
        # print (y_true,y_pred)
        y_true = tf.cast(tf.reshape(y_true,[-1,max_len_char]),tf.int32)
        y_pred = tf.reshape(y_pred,[-1,max_len_char,n_decoder_tokens])
        
        loss_ = tf.nn.sparse_softmax_cross_entropy_with_logits( labels = y_true,logits = y_pred)
        mask = tf.constant(idx_pad,dtype = tf.int32)
        loss_ = tf.boolean_mask(tensor = loss_, mask = tf.not_equal(y_true,mask)) 
        loss_CE = tf.reduce_sum(loss_)/tf.cast(tf.shape(y_true)[0],dtype=tf.float32)
        return loss_CE
    return mask_categorical_crossentropy

Example:

sess = tf.Session()

y_label_onehot = tf.convert_to_tensor([