Created by: mzchtx

环境

• system: centos6u3
• python: 2.7.14
• paddle: 1.5.1.post97
• nccl: 2.2.13

问题

使用 FP16 训练 ResNet50，如果添加 label_smooth 则会报错，代码如下：

out = ResNet50().net(input=image, class_dim=class_dim)

epsilon = 0.1
one_hot_label = fluid.layers.one_hot(input=label, depth=class_dim)
smooth_label = fluid.layers.label_smooth(label=one_hot_label, epsilon=epsilon)
cost = fluid.layers.softmax_with_cross_entropy(out, smooth_label, soft_label=True)
avg_cost = fluid.layers.mean(x=cost)

错误信息：

去掉 label_smooth 正常：

out = ResNet50().net(input=image, class_dim=class_dim)
cost = fluid.layers.softmax_with_cross_entropy(out, label, soft_label=True)
avg_cost = fluid.layers.mean(x=cost)