Created by: yeyupiaoling

环境

• PaddlePaddle dev 版本

问题

话不多少，上一个demo，

import paddle.fluid as fluid
import numpy as np

BATCH_SIZE = 384
data1 = fluid.layers.data(name="input1", shape=[2], dtype="float32")
data2 = fluid.layers.data(name="input2", shape=[1], dtype="float32")


def cls_ohem(cls_prob, label):
    zeros = fluid.layers.zeros_like(label)
    ones = fluid.layers.zeros_like(label) + 1

    # 把小于等于0的都设置为0
    cond = fluid.layers.cast(fluid.layers.greater_than(label, zeros), dtype='int64')
    ie = fluid.layers.IfElse(cond)
    with ie.true_block():
        ie.output(ones)

    with ie.false_block():
        ie.output(zeros)
    label_filter_invalid = ie()[0]

    # 类别size[2*batch]
    num_cls_prob = fluid.layers.size(cls_prob)
    cls_prob_reshpae = fluid.layers.reshape(cls_prob, [num_cls_prob, -1])
    label_int = fluid.layers.cast(label_filter_invalid, np.int32)
    label_int = fluid.layers.squeeze(label_int, axes=[])

    # 获取batch数
    num_row = BATCH_SIZE
    # 对应某一batch而言，batch*2为非人类别概率，batch*2+1为人概率类别,indices为对应 cls_prob_reshpae
    # 应该的真实值，后续用交叉熵计算损失
    row = fluid.layers.range(0, num_row * 2, 2, np.int32)
    indices_ = row + label_int

    # 真实标签对应的概率
    label_prob = fluid.layers.squeeze(fluid.layers.gather(cls_prob_reshpae, indices_), axes=[])
    loss = 0 - fluid.layers.log(label_prob + 1e-10)
    zeros = fluid.layers.zeros_like(label_prob)

    # 统计 0 和 1 的数量
    valid_inds = fluid.layers.where(fluid.layers.greater_equal(label, zeros))
    num_valid = fluid.layers.reduce_sum(valid_inds)

    # 选取70%的数据
    keep_num = fluid.layers.cast(num_valid * 0.7, dtype=np.int32)

    # 只选取 0 和 1的70%损失
    loss1 = loss * valid_inds
    loss, _ = fluid.layers.topk(loss1, k=keep_num)
    return fluid.layers.reduce_mean(loss)


loss = cls_ohem(data1, data2)

place = fluid.CPUPlace()
exe = fluid.Executor(place)
exe.run(fluid.default_startup_program())

d1 = []
for i in range(BATCH_SIZE):
    r = np.random.rand()
    d1.append([r, 1 - r])

d1 = np.random.rand(BATCH_SIZE, 2).astype(np.float32)
d2 = np.array(np.random.randint(-2, 2, [BATCH_SIZE, 1])).astype('float32')


r = exe.run(program=fluid.default_main_program(),
            feed={"input1": d1, "input2": d2},
            fetch_list=[loss])

print(r)

错误信息为

C++ Callstacks: 
Tensor holds the wrong type, it holds int64_t, but desires to be bool at [/paddle/paddle/fluid/framework/tensor_impl.h:45]
PaddlePaddle Call Stacks: