thu-pacman / pet Goto Github PK

在 NVIDIA GPU 环境下克隆此工程并编译运行。

首先使用 PET_DCU/benchmark/models/scripts/bert_onnx.py 导出了bert的onnx文件，然后编译了DCU之后在build下执行：

./onnx_origin ../benchmark/models/scripts/mybert-new.onnx

此时会尝试导入ONNX到PET的Graph中，这个时候发生了报错：

onnx_origin: /home/PET_DCU/include/operator.h:1795: tpm::ReshapeOp::ReshapeOp(tpm::Tensor *, tpm::Tensor *): Assertion `input->size() == output->size()' failed.
Aborted (core dumped)

可以看到挂在Reshape Op的形状检查，我打印了一下input->size() 和 output->size() ，分别为：

589824 37748736

所以输出Tensor的size是输入的64倍，这个64是导出ONNX时的batch_size，希望可以解答这个错误的原因是什么

Hello,

I am trying to understand the generator as in the PET paper.

To my understanding, the threshold parameter in the Generator::run() function defines the number of random test point that a legal mutant has to satisfy, as computed by the Generator::approx_equal() function.

The default threshold=0.7 results in, for example, sometimes a Matmul has 20 mutants, and sometimes has 28 mutants. Is this randomness intended?

If a legal mutant passes this test, how does the correction kernel generated? Is the correction kernel generated by the Generator?

Many Thanks,

Hello, I'm trying to PET, but when I configure according to the usage document, the "onnx" folder under "build" is not automatically generated. What should I do? And how do I use PET in Python?