【Hackathon 5th No.18】Add Binomial and Poisson API -part #57856

NKNaN · 2023-09-29T10:42:11Z

PR types

New features

PR changes

APIs

Description

Add Binomial and Poisson API

rfc of Binomial: https://github.com/PaddlePaddle/community/blob/master/rfcs/APIs/20230926_api_design_for_binomial.md
rfc of Poisson: https://github.com/PaddlePaddle/community/blob/master/rfcs/APIs/20230926_api_design_for_poisson.md

paddle-bot · 2023-09-29T10:42:16Z

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.

paddle-ci-bot · 2023-10-17T03:10:41Z

Sorry to inform you that 4c1bfd4's CIs have passed for more than 7 days. To prevent PR conflicts, you need to re-run all CIs manually.

paddle-ci-bot · 2023-11-01T03:12:40Z

Sorry to inform you that 8a1f336's CIs have passed for more than 7 days. To prevent PR conflicts, you need to re-run all CIs manually.

NKNaN · 2023-11-16T14:11:57Z

可以麻烦两位老师看一下 CI-Coverage 和 Paddle-Doc-Preview 是哪里的问题吗？

luotao1 · 2023-11-17T02:37:28Z

Coverage可以点进去看覆盖率不够，Doc-Preview可不用管

NKNaN · 2023-11-20T02:39:43Z

请问两位老师还有什么需要修改的地方吗？如果没什么问题我就开始写中文文档。

luotao1 · 2023-11-21T09:56:36Z

如果没什么问题我就开始写中文文档。

可以先同步写中文文档

cxxly · 2023-11-22T06:43:51Z

python/paddle/distribution/binomial.py

+    """
+
+    def __init__(self, total_count, probability):
+        self.dtype = 'float32'


使用全局默认数据类型 https://www.paddlepaddle.org.cn/documentation/docs/zh/api/paddle/get_default_dtype_cn.html#get-default-dtype

cxxly · 2023-11-22T06:45:45Z

python/paddle/distribution/binomial.py

+        if isinstance(probability, np.ndarray):
+            probability = paddle.to_tensor(probability)
+        total_count = paddle.cast(total_count, dtype=self.dtype)
+        probability = paddle.cast(probability, dtype=self.dtype)


to_tensor 可以指定dtype参数，此处代码可以简化

cxxly · 2023-11-22T06:47:37Z

python/paddle/distribution/binomial.py

+        if isinstance(total_count, np.ndarray):
+            total_count = paddle.to_tensor(total_count)
+        if isinstance(probability, np.ndarray):
+            probability = paddle.to_tensor(probability)


上述文档中描述参数类型是 float|Tensor，实际支持类型和文档描述保持一致；此处建议去掉numpy类型支持

cxxly · 2023-11-22T06:54:02Z

python/paddle/distribution/binomial.py

+            Tensor: Sampled data with shape `sample_shape` + `batch_shape`.
+        """
+        if not isinstance(shape, Iterable):
+            raise TypeError('sample shape must be Iterable object.')


Iterable 和 Sequence 语义略有区别 https://stackoverflow.com/questions/72157296/what-is-the-difference-between-type-hinting-a-variable-as-an-iterable-versus-a-s ，此处应该是只需要传入Sequence

cxxly · 2023-11-22T06:57:40Z

python/paddle/distribution/binomial.py

+                self.probability, shape=output_shape
+            )
+            if in_dynamic_mode():
+                broadcast_func = np.frompyfunc(binomial_sample, nin=2, nout=1)


使用Paddle原生采样算法，numpy仅支持cpu采样； Paddle GPU底层基础采样算法使用curand进行加速，大规模数据采样性能远超过CPU

binomial的原生采样算法我目前没有找到直接对应哪个API，如果用 paddle.bernoulli() 来做的话，对于这种情况: Binomial( [ n_1, n_2 ], [ p_1, p_2 ] ) 的采样可能会比较麻烦，不知道这块要怎么做好一点

binomial可以通过均匀分布逆采样实现，参考pytorch/aten/src/ATen/native/Distributions.h sample_binomial；你的代码其实已经基于numpy实现了类似逻辑

这个题目涉及到新增c++ Kernel，难度会有所提升，已和组委会同学沟通 @luotao1

好的，我试着写一下

cxxly · 2023-11-22T06:59:27Z

python/paddle/distribution/binomial.py

+                break
+            u -= f
+            y += 1
+            f *= g / y - p_q


使用Paddle原生采样算法及逆变换

cxxly · 2023-11-22T06:59:42Z

python/paddle/distribution/binomial.py

+                    / 166320.0
+                )
+                if A <= tmp:
+                    return n - y if p > 0.5 else y


cxxly · 2023-11-22T07:03:27Z

python/paddle/distribution/poisson.py

+    """
+
+    def __init__(self, rate):
+        self.dtype = 'float32'


cxxly · 2023-11-22T07:04:01Z

python/paddle/distribution/poisson.py

+        super().__init__(batch_shape)
+
+    def _to_tensor(self, rate):
+        """Convert the input parameters into tensors with dtype = float32


with global default dtype.

cxxly · 2023-11-22T07:04:23Z

python/paddle/distribution/poisson.py

+            rate = paddle.to_tensor([rate], dtype=self.dtype)
+        if isinstance(rate, np.ndarray):
+            rate = paddle.to_tensor(rate)
+        rate = paddle.cast(rate, dtype=self.dtype)


cxxly · 2023-11-22T07:05:23Z

python/paddle/distribution/poisson.py

+                    x=output_rate,
+                    out=output,
+                )
+                return output


同上，使用paddle原生采样算法及逆变换实现

已修改，用 paddle.poisson()

cxxly · 2023-11-22T07:07:13Z

python/paddle/distribution/poisson.py

+                # step H
+                if c * abs(u) <= py * math.exp(px + E) - fy * math.exp(fx + E):
+                    break
+    return K


cxxly · 2023-11-22T07:07:20Z

python/paddle/distribution/poisson.py

+        np.ndarray, binomial samples.
+    """
+    broadcast_func = np.frompyfunc(poisson_sample, nin=1, nout=1)
+    return np.array(broadcast_func(rr), dtype='float32')


cxxly · 2023-11-22T07:10:46Z

test/distribution/test_distribution_binomial.py

+            self._np_mean(),
+            rtol=config.RTOL.get(str(self.probability.dtype)),
+            atol=config.ATOL.get(str(self.probability.dtype)),
+        )


使用 scipy.stats.binom.mean

cxxly · 2023-11-22T07:11:21Z

test/distribution/test_distribution_binomial.py

+            self._np_variance(),
+            rtol=config.RTOL.get(str(self.probability.dtype)),
+            atol=config.ATOL.get(str(self.probability.dtype)),
+        )


使用 scipy.stats.binom.var

NKNaN · 2023-12-19T10:42:49Z

请解决下冲突

已解决

cxxly

LGTM

jeff41404 · 2023-12-20T08:34:26Z

python/paddle/distribution/binomial.py

+            [2.94053698, 3.00781751, 2.51124287])
+    """
+
+    def __init__(self, total_count, probability):


use probs instead of probability will be better? consistent with other similar APIs like paddle.distribution.Bernoulli, paddle.distribution.Geometric and paddle.distribution.Multinomial?

rfc use probs which is inconsistent with the code

sunzhongkai588 · 2023-12-20T09:02:19Z

python/paddle/distribution/binomial.py

+
+    In the above equation:
+
+    * :math:`total_count = n`: is the size, meaning the total number of Bernoulli experiments.


Suggested change

* :math:`total_count = n`: is the size, meaning the total number of Bernoulli experiments.

* :math:`total\_count = n`: is the size, meaning the total number of Bernoulli experiments.

jeff41404

LGTM

sunzhongkai588

LGTM

paddle-bot bot added the contributor External developers label Sep 29, 2023

NKNaN changed the title ~~【Hackathon 5th】No.19: Add Binomial and Poisson API~~ 【Hackathon 5th No.19】Add Binomial and Poisson API Oct 1, 2023

Ligoml mentioned this pull request Oct 7, 2023

【PaddlePaddle Hackathon 5th】开源贡献个人挑战赛 #57262

Open

luotao1 added the PaddlePaddle Hackathon label Oct 9, 2023

luotao1 assigned luotao1 and cxxly Oct 9, 2023

NKNaN changed the title ~~【Hackathon 5th No.19】Add Binomial and Poisson API~~ 【Hackathon 5th No.18】Add Binomial and Poisson API Oct 11, 2023

NKNaN mentioned this pull request Nov 22, 2023

【Hackathon 5th No. 18】Add Binomial and Poisson API PaddlePaddle/docs#6325

Merged

cxxly reviewed Nov 22, 2023

View reviewed changes

python/paddle/distribution/binomial.py Outdated

/ 166320.0

)

if A <= tmp:

return n - y if p > 0.5 else y

Copy link

Contributor

cxxly Nov 22, 2023

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

同上

cxxly reviewed Nov 22, 2023

View reviewed changes

NKNaN added 16 commits December 19, 2023 14:38

fix test timeout

aac4888

fix test

69735d4

fix test

18b74ba

fix docs

f1629d6

fix test

49db8fb

fix test

0dfe211

fix test coverage

1150b6c

update docs

5bf911a

update poisson and poisson test

d686dde

update poisson

0e19d49

update poisson test

4afb38c

update binomial

b2f8549

update binomial api and tests

8bf89a3

update docs and tests

af9c127

update doc test

70d5abf

fix conflict

3fc4ad6

NKNaN force-pushed the ayase/develop branch from 9a63ead to 3fc4ad6 Compare December 19, 2023 06:42

cxxly previously approved these changes Dec 19, 2023

View reviewed changes

luotao1 assigned jeff41404 and sunzhongkai588 Dec 19, 2023

jeff41404 reviewed Dec 20, 2023

View reviewed changes

sunzhongkai588 reviewed Dec 20, 2023

View reviewed changes

update signiture

6f5e3df

NKNaN dismissed cxxly’s stale review via 6f5e3df December 20, 2023 15:09

jeff41404 approved these changes Dec 21, 2023

View reviewed changes

sunzhongkai588 approved these changes Dec 21, 2023

View reviewed changes

luotao1 changed the title ~~【Hackathon 5th No.18】Add Binomial and Poisson API~~ 【Hackathon 5th No.18】Add Binomial and Poisson API -part Dec 21, 2023

luotao1 merged commit ac69600 into PaddlePaddle:develop Dec 22, 2023
29 checks passed

NKNaN deleted the ayase/develop branch February 2, 2024 04:08


		In the above equation:

		* :math:`total_count = n`: is the size, meaning the total number of Bernoulli experiments.

	* :math:`total_count = n`: is the size, meaning the total number of Bernoulli experiments.
	* :math:`total\_count = n`: is the size, meaning the total number of Bernoulli experiments.

【Hackathon 5th No.18】Add Binomial and Poisson API -part #57856

【Hackathon 5th No.18】Add Binomial and Poisson API -part #57856

Conversation

NKNaN commented Sep 29, 2023 • edited

PR types

PR changes

Description

paddle-bot bot commented Sep 29, 2023

paddle-ci-bot bot commented Oct 17, 2023

paddle-ci-bot bot commented Nov 1, 2023

NKNaN commented Nov 16, 2023

luotao1 commented Nov 17, 2023

NKNaN commented Nov 20, 2023

luotao1 commented Nov 21, 2023

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

cxxly Nov 22, 2023 • edited

Choose a reason for hiding this comment

Choose a reason for hiding this comment

cxxly Nov 28, 2023 • edited

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

cxxly Nov 22, 2023 • edited

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

NKNaN commented Dec 19, 2023

cxxly left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

jeff41404 left a comment

Choose a reason for hiding this comment

sunzhongkai588 left a comment

Choose a reason for hiding this comment

NKNaN commented Sep 29, 2023 •

edited

cxxly Nov 22, 2023 •

edited

cxxly Nov 28, 2023 •

edited

cxxly Nov 22, 2023 •

edited