Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

【Hackathon No.7】add rfc: api design for cdist #37

Merged
merged 3 commits into from
Mar 29, 2022
Merged
Changes from 1 commit
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
260 changes: 260 additions & 0 deletions rfcs/APIs/20220316_api_design_for_cdist.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,260 @@
# paddle.cdist 设计文档

|API名称 | paddle.cdist |
|---|---|
|提交作者<input type="checkbox" class="rowselector hidden"> | 汪昕([GreatV](https://github.com/GreatV)) |
|提交时间<input type="checkbox" class="rowselector hidden"> | 2022-03-16 |
|版本号 | V1.0 |
|依赖飞桨版本<input type="checkbox" class="rowselector hidden"> | develop |
|文件名 | 20200316_api_design_for_cdist.md |


# 一、概述
## 1、相关背景

Issue: [【PaddlePaddle Hackathon 2】7、为 Paddle 新增 cdist API](https://github.com/PaddlePaddle/Paddle/issues/40328)

## 2、功能目标

为 Paddle 新增 cdist API。dist API 用于计算两个输入 Tensor 的 p 范数(p-norm),计算结果为形状为 [1] 的 Tensor,而 cdist API 则用于计算两个输入 Tensor 的所有行向量对的 p 范数(p-norm),输出结果的形状和两个 Tensor 乘积的形状一致。

## 3、意义

为 Paddle 新增 cdist API,提供距离计算功能。

# 二、飞桨现状

对飞桨框架目前不支持此功能,可用其他API组合实现的此功能,对于距离度量为欧氏距离(p=2),代码如下;

```Python
import paddle


def pairwise_dist (A, B):
A2 = (A ** 2).sum(axis=1).reshape((-1, 1))
B2 = (B ** 2).sum(axis=1)
D = A2 + B2 - 2 * A.mm(B.transpose((1, 0)))
D = D.sqrt()
return D

a = paddle.to_tensor([[0.9041, 0.0196], [-0.3108, -2.4423], [-0.4821, 1.059]])
b = paddle.to_tensor([[-2.1763, -0.4713], [-0.6986, 1.3702]])
pairwise_dist(a, b)


Tensor(shape=[3, 2], dtype=float32, place=CPUPlace, stop_gradient=True,
[[3.11927032, 2.09589314],
[2.71384096, 3.83217263],
[2.28300953, 0.37910157]])
```

# 三、业内方案调研

## 1. Scipy

在 Scipy 中使用的 API 格式如下:

`scipy.spatial.distance.cdist(XA, XB, metric='euclidean', *, out=None, **kwargs)`

上述函数参数中,`metric` 表示距离度量方式。Scipy 支持丰富的距离度量方式,如:'braycurtis', 'canberra', 'chebyshev', 'cityblock', 'correlation','cosine', 'dice', 'euclidean', 'hamming', 'jaccard', 'jensenshannon','kulsinski', 'kulczynski1', 'mahalanobis', 'matching', 'minkowski','rogerstanimoto', 'russellrao', 'seuclidean', 'sokalmichener','sokalsneath', 'sqeuclidean', 'yule' 等。

实现的伪代码如下:

```Python
def cdist(XA, XB, metric, **kwargs):
mA = XA.shape[0]
mB = XB.shape[0]
for i in range(mA):
for j in range(mB):
dm[i, j] = metric(XA[i], XB[j], **kwargs)
return dm
```

## 2. MatLab

在 MatLab 中使用的 API 格式如下:

`D = pdist2(X,Y,Distance)`

在 MatLab 版中,`Distance` 表示距离度量方式。MatLab 同样支持丰富的距离度量方式,如:'euclidean','squaredeuclidean', 'seuclidean', 'mahalanobis', 'cityblock', 'minkowski' 等等。

实现 `minkowski` 距离的伪代码如下:

```matlab
function [D,I] = pdist2(X,Y,dist,varargin)
additionalArg = varargin{1};
outClass = 'double';

[nx,p] = size(X);
[ny,py] = size(Y);

expon = additionalArg;
for i = 1:ny
dpow = zeros(nx,1,outClass);
for q = 1:p
dpow = dpow + abs(X(:,q) - Y(i,q)).^expon;
end
dpow = dpow .^ (1./expon);
```
## 3. Pytorch

在 Pytorch 中使用的 API 格式如下:

`torch.cdist(x1, x2, p=2.0, compute_mode='use_mm_for_euclid_dist_if_necessary')`

在 Pytorch 中, 如果 p∈[1, ∞),则函数与 `scipy.spatial.distance.cdist(input,'minkowski', p=p)` 等价;如果 p=0,则函数与 `scipy.spatial.distance.cdist(input, 'hamming') * M` 等价;如果 p=∞,则函数与 `scipy.spatial.distance.cdist(xn, lambda x, y: np.abs(x - y).max())` 意义最接近。

## 4. Tensorflow

没有对应的API, 可以用其他 API 组合实现。[#issues 30659](https://github.com/tensorflow/tensorflow/issues/30659)

```Python
def pairwise_dist (A, B):
"""
Computes pairwise distances between each elements of A and each elements of B.
Args:
A, [m,d] matrix
B, [n,d] matrix
Returns:
D, [m,n] matrix of pairwise distances
"""
# squared norms of each row in A and B
na = tf.reduce_sum(tf.square(A), 1)
nb = tf.reduce_sum(tf.square(B), 1)

# na as a row and nb as a co"lumn vectors
na = tf.reshape(na, [-1, 1])
nb = tf.reshape(nb, [1, -1])

# return pairwise euclidean difference matrix
D = tf.sqrt(tf.maximum(na - 2*tf.matmul(A, B, False, True) + nb, 0.0))
return D
```

# 四、对比分析

## 1. 不同框架API使用方式

### 1. Scipy

```Python
from scipy.spatial import distance
import numpy as np

a = np.array([[0.9041, 0.0196], [-0.3108, -2.4423], [-0.4821, 1.059]])
b = np.array([[-2.1763, -0.4713], [-0.6986, 1.3702]])

distance.cdist(a, b, "euclidean")


array([[3.11927026, 2.09589304],
[2.71384068, 3.83217237],
[2.28300936, 0.37910116]])
```

### 2. MatLab

```matlab
a = [0.9041, 0.0196; -0.3108, -2.4423; -0.4821, 1.059];
b = [-2.1763, -0.4713; -0.6986, 1.3702];
c = pdist2(a, b)


c =

3.1193 2.0959
2.7138 3.8322
2.2830 0.3791
```

### 3. PyTorch

```Python
import torch
a = torch.tensor([[0.9041, 0.0196], [-0.3108, -2.4423], [-0.4821, 1.059]])
b = torch.tensor([[-2.1763, -0.4713], [-0.6986, 1.3702]])
torch.cdist(a, b, p=2)
tensor([[3.1193, 2.0959],
[2.7138, 3.8322],
[2.2830, 0.3791]])
```

### 4. Tensorflow

```Python
import tensorflow as tf
def pairwise_dist (A, B):
"""
Computes pairwise distances between each elements of A and each elements of B.
Args:
A, [m,d] matrix
B, [n,d] matrix
Returns:
D, [m,n] matrix of pairwise distances
"""
# squared norms of each row in A and B
na = tf.reduce_sum(tf.square(A), 1)
nb = tf.reduce_sum(tf.square(B), 1)
# na as a row and nb as a co"lumn vectors
na = tf.reshape(na, [-1, 1])
nb = tf.reshape(nb, [1, -1])
# return pairwise euclidean difference matrix
D = tf.sqrt(tf.maximum(na - 2*tf.matmul(A, B, False, True) + nb, 0.0))
return D
a = tf.constant([[0.9041, 0.0196], [-0.3108, -2.4423], [-0.4821, 1.059]])
b = tf.constant([[-2.1763, -0.4713], [-0.6986, 1.3702]])
pairwise_dist(a, b)
<tf.Tensor: shape=(3, 2), dtype=float32, numpy=
array([[3.1192703 , 2.0958931 ],
[2.7138407 , 3.8321724 ],
[2.2830095 , 0.37910157]], dtype=float32)>
```

上述框架从使用体验来说,**Scipy** 支持的距离度量方式最多,调用API方便,易于使用,但不支持 Tensor。**MatLab** 同 **Scipy** 也存在相同的问题。**PyTorch** 则支持 Tensor,函数与 `scipy.spatial.distance.cdist(input,'minkowski', p=p)` 等价。**Tensorflow** 则支持 Tensor,但没有此API,需要手动实现,使用较为繁琐。

# 五、设计思路与实现方案

## 命名与参数设计

<!-- 参考:[飞桨API 设计及命名规范](https://www.paddlepaddle.org.cn/documentation/docs/zh/develop/dev_guides/api_contributing_guides/api_design_guidelines_standard_cn.html) -->

API设计为 `paddle.cdist(x1, x2, p=2.0)`。其中 `x1` 为 `B X P X M` 张量,`x2` 为 `B X R X M` 张量。`p` 为 p-范数对应的 p 值,p ∈[0,∞]。输出张量的形状为 `B X P X R`。

GreatV marked this conversation as resolved.
Show resolved Hide resolved
这里与 `torch.cdist(x1, x2, p=2.0, compute_mode='use_mm_for_euclid_dist_if_necessary')` 的设计不同之处是去除了 `compute_mode` 参数,其作用是使用矩阵乘法加速欧氏距离(p=2)的计算。在 PyTorch 中,参数 `compute_mode='use_mm_for_euclid_dist_if_necessary'`,当 P > 25 或 R > 25 时,则使用矩阵乘法加速计算,否则使用普通的欧氏距离计算。对于 PaddlePaddle 的实现这应该成为默认的优化计算行为,无需做成用户可选的行为。

## 底层OP设计

参考 PyTorch [cdist_impl](https://github.com/pytorch/pytorch/blob/master/aten/src/ATen/native/Distance.cpp#L49) 进行实现,实现位置为 Paddle repo 的 `paddle/phi/kernels` 目录。

## API实现方案

参考 PyTorch 进行实现,实现位置为 Paddle repo `python/paddle/tensor/linalg.py` 目录。

# 六、测试和验收的考量

<!-- 参考:[新增API 测试及验收规范](https://www.paddlepaddle.org.cn/documentation/docs/zh/develop/dev_guides/api_contributing_guides/api_accpetance_criteria_cn.html) -->

可考虑一下场景:

1. 当`x1`、`X2` 为空张量,输出为空张量,且输出张量形状正确;
2. 结果一致性,和 SciPy 以及 PyTorch 结果的数值的一致性, `paddle.cdist(x1, x2, p=2.0)` , `scipy.spatial.distance.cdist(input,'minkowski', p=2)` 和 `torch.cdist(x1, x2, p=2.0, compute_mode='use_mm_for_euclid_dist_if_necessary')` 结果是否一致;
3. 当 p < 1 时, 确保反向传播不会产生 NaNs;
4. 异常测试,对于参数异常值输入,应该有友好的报错信息及异常反馈,需要有相关测试Case验证。

# 七、可行性分析和排期规划

本 API 主要参考 PyTorch 实现,难度适中,工期上能满足要求。

# 八、影响面

为独立新增API,对其他模块没有影响。

# 名词解释

# 附件及参考资料