paddle softmax_with_cross_entropy 的用法

import numpy as np
import paddle.fluid as fluid
import paddle

logit_y = np.array([[1.23, 2.33, 3.33, 2.11], \
                    [5.23, 2.33, 3.33, 2.11], \
                    [1.23, 8.33, 3.33, 2.11], \
                    [1.23, 2.33, 3.33, 2.11]]).astype(np.float32)

output_y1 = np.array([[3], [0], [1], [1]]).astype(np.int64)
output_y2 = np.array([[0,0,0,1], [1,0,0,0], [0,1,0,0], [0,1,0,0]]).astype(np.float32)

logit_y = paddle.to_tensor(logit_y)
output_y1 = paddle.to_tensor(output_y1)
output_y2 = paddle.to_tensor(output_y2)

#print(logit_y.shape, output_y1.shape)
#print(logit_y.numpy(), output_y1.numpy())

loss1 = fluid.layers.softmax_with_cross_entropy(logit_y, output_y1)
print(loss1)
print(loss1.numpy()[0])

loss2 = fluid.layers.softmax_with_cross_entropy(logit_y, output_y2, soft_label=True)
print(loss2)
print(loss2.numpy()[0])


# -------------

# loss_i = -1.0 * y_i * np.log(p_i), where y_i = 1 !!!

fluid.layers.softmax(logit_y)
"""
Tensor(shape=[4, 4], dtype=float32, place=CPUPlace, stop_gradient=True,
       [[0.06858130, 0.20602959, 0.56004649, 0.16534261],
        [0.80080152, 0.04406267, 0.11977477, 0.03536112],
        [0.00081730, 0.99053812, 0.00667419, 0.00197042],
        [0.06858130, 0.20602959, 0.56004649, 0.16534261]])
"""

print(-1.0 * np.log(0.16534261)) # 1.79973555

官方文档：softmax_with_cross_entropy-API文档-PaddlePaddle深度学习平台该OP实现了softmax交叉熵损失函数。该函数会将softmax操作、交叉熵损失函数的计算过程进行合并，从而提供了数值上更稳定的梯度值。 因为该运算对 logits 的 axis 维执行softmahttps://www.paddlepaddle.org.cn/documentation/docs/zh/1.8/api_cn/layers_cn/softmax_with_cross_entropy_cn.html