案例符号回归 Transformer4SR在Nvidia A100精度结果与官方差距大

[yuchunliu-metax]

bug 描述 bug description

简述：
案例transformer4sr 在A100平台验证srsd-feynman_easy模式精度指标与官方差距大
1）A100 精度结果 Hit rate: 0/30 ,与官网精度指标8/30差距较大
2）公式预测结果与官方差距大
A100:

The ground truth is: 25*x1 + x2*log(x1)
The prediction is: -C*x1 + 1/log(x2)

官方：

The ground truth is: 25*x1 + x2*log(x1)
The prediction is: -C*x1 + x2*log(x1)

软件版本：paddle：2.6.2 , PaddleScience：commit ab417c8
测试案例：PaddleScience/examples/transformer4sr
测试环境：NVIDIA A100-PCIE-40GB

问题复现：
测试命令：
1） 训练：python transformer4sr.py
2） 评估：python transformer4sr.py mode=eval EVAL.pretrained_model_path=./best_model.pdparams

日志：
eval.log
train.log
best_model.zip

精度验证：
1、A100 精度结果:

=> Mean ZSS distance: 0.677 +- 0.230
=> Hit rate: 0/30
target C*x1*x2 pred: C/(x1 + x2 + x3**2)**2
target C*x1/x2**2 pred: C*x1*log(sin(1/x2)**2)
target C*x1*x2 pred: C*x1*sin(x2)**2
target C*x1*x2 pred: C*x1*sin(x2)**2
target C*x1*x2**2 pred: 1/(x1**2*cos(exp(-C*x2)))
target C*x1*x2*sin(x3) pred: C*x2**2*x3*log(x1)
target C*x1*x2*x3*sin(x4) pred: C*x1*x2*log(x3)
target C*x1/x2 pred: C*x1*log(cos(x1))
target C*sin(x1)/sin(x2) pred: x2 + exp(-C*x2 + x1)
target C*x1*x3/(x1*x2 + x3) pred: sin(C*x2**2)**2/x1**2
The ground truth is: 25*x1 + x2*log(x1)
The prediction is: -C*x1 + 1/log(x2)

2、官方 精度结果:

Mean ZSS distance(srsd-feynman_easy): 0.658 +- 0.390
Hit rate(srsd-feynman_easy): 8/30

其他补充信息 Additional Supplementary Information

No response