Fix numpy 1.20+ deprecation warnings

[zlsh80826]

PR types

Bug fixes

PR changes

Others

Describe

np.bool, np.int, np.float, np.complex, np.object, np.str, np.long, np.unicode are deprecated. There are tons of warnings during Paddle runs, those warnings are very distracting. This PR makes following changes

• Replace np.bool/np.bool8 with np.bool_

• Replace np.object with np.object_

• Replace np.complex with np.complex128

• Replace np.float with np.float64

• Replace np.int with np.int_

• See https://numpy.org/devdocs/release/1.20.0-notes.html#deprecations for more details.

• Tensorflow alredy adopted the change last year

• Related issue: paddle2.2.2创建张量时遇到NumPy弃用警告：DeprecationWarning #41745

• Besides this PR, we should notify all developers that not to use the deprecated aliases

[paddle-bot-old]

你的PR提交成功，感谢你对开源项目的贡献!
请关注后续CI自动化测试结果，详情请参考Paddle-CI手册。
Your PR has been submitted. Thanks for your contribution!
Please wait for the result of CI firstly. See Paddle CI Manual for details.

[zlsh80826]

@qili93 There are many of CI failures after changing np.float to np.float32. It seems the variance should be np.float64, could you check it again?

[qili93]

Got it, I will double confirm the failure CI test cases.

[qili93]

hi @zlsh80826

Double confirmed Paddle code, the root cause is :

1. the type: <class 'numpy.ndarray'> coverted to std::vector<int> by protobuf.cc https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/fluid/pybind/protobuf.cc#L289
2. the attribute.h always got failure when it try to convert std::vector<int> to std::vector<float> https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/fluid/framework/attribute.h#L47

And the reason np.float64 works is : std::vector<double> defined how to convert int array to double https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/fluid/framework/attribute.h#L184

So propose to keep all the numpy array data type to np.float64 in python code, i.e. change the box_coder_op / density_prior_box_op / prior_box_op back to np.float64.

I have tried to add a similar function to handle std::vector<float> as std::vector<double> does, while it will result in accuracy loss. As described in the PR here #30126.

Apologize for my previous comment is not correct.

Thanks!
Qi

[paddle-bot-old]

Sorry to inform you that eb4d04d's CIs have passed for more than 7 days. To prevent PR conflicts, you need to re-run all CIs manually.

[qili93]

hi, @zlsh80826

Seems some code style check failed in CI, propose to use this docker image to run git commit command, it will use clang-format installed in docker to format the code style automatically.

nvidia-docker run -it --name paddle-dev -v `pwd`:/workspace \
  --network=host --shm-size=128G --workdir=/workspace \
  --cap-add=SYS_PTRACE --security-opt seccomp=unconfined \
  registry.baidubce.com/paddlepaddle/paddle:latest-dev-cuda11.2-cudnn8-gcc82 /bin/bash

# install pre-commit inside container
pip3 install pre-commit

# then clang-format will format the code style automatically
git commit -m "......"

[zlsh80826]

Hi @qili93,
I already rerun the formatter. Could you help me to rerun the PR-CI-Coverage? It seems there is a network issue.

[qili93]

LGTM

[XiaoguangHu01]

LGTM