使用 SPU 优化 PCA 算法 #259

Candicepan · 2023-07-21T07:03:36Z

此 ISSUE 为隐语开源共建计划（SecretFlow Open Source Contribution Plan，简称 SF OSCP）第二期任务 ISSUE，欢迎社区开发者参与共建～
若有感兴趣想要认领的任务，但还未报名，辛苦先完成报名进行哈～

任务介绍

任务名称：使用 SPU 优化 PCA 算法
技术方向：SPU/SML
任务难度：挑战🌟🌟🌟

详细要求

安全性（尽量少 reveal）
功能性：实现算法的基本功能，包括：
- 性能比原有的纯 power iteration 要高；或收敛速度更快(体现在能支持更大的特征维度)
- （推荐）不需要显示的计算协方差矩阵
收敛性：包含 simulator 跑出的实验数据并且证明收敛性/准确性（最好有与明文 sklearn 结果的对比，可详见参考 PR）
正确性：请确保提交的代码内容为可以直接运行的
代码规范：Python 代码需要使用 black+isort 进行格式化（流水线包含代码规范检查卡点）
提交说明：关联该 isuue 并提交代码至 https://github.com/secretflow/spu/tree/main/sml
特殊说明：若某个特性有特殊的限制，如需要 FM128，需要更多 fxp 等需要在注释文档中明确说明

能力要求

熟悉经典的机器学习算法
熟悉 JAX 或 NumPy，可以使用 NumPy 实现算法

操作说明

操作指引：https://www.secretflow.org.cn/docs/spu/latest/en-US/getting_started/tutorials/develop_your_first_mpc_application
可参考明文实现：decomposition/_pca
范例
参考 PR

开发须知

以下部分代码请必须增加代码注释，对对应代码模块进行说明，包括：

__init__函数的超参数含义
fit 的具体算法实现说明

tarantula-leo · 2023-08-01T04:40:01Z

tarantula-leo give it to me.

Candicepan · 2023-08-01T08:10:03Z

tarantula-leo give it to me.

感谢你的认领～为避免走弯路，请先提交你的设计思路，经确认后再进入开发阶段

设计思路说明：简单说明计划使用什么算法 or 什么优化器实现任务需求即可

tarantula-leo · 2023-08-01T08:58:10Z

使用QR分解实现

deadlywing · 2023-08-01T09:38:05Z

Hi，可以稍微详细一点么？

明确一下，我们希望这次的要求是PCA算法能不需要先计算cov matrix，您可以对照一下看自己的idea是否满足要求哈～

Thanks

tarantula-leo · 2023-08-01T10:10:48Z

考虑不计算协方差的出发点是什么？之前是准备用QR分解+QR迭代

deadlywing · 2023-08-01T10:49:59Z

因为计算协方差在n比较大的时候cost也比较大，如果再做特征值分解可能性能不一定能比当前的power iteration效率高～本期的任务会相对更注重性能，所以做了这样的要求。

当前，重点还是性能，如果QR迭代性能确实更高也是ok的～欢迎尝试

Thanks

tarantula-leo · 2023-08-02T08:16:05Z

“计算协方差在n比较大的时候cost也比较大”这是指矩阵乘法的cost么？
原PCA代码测试了下在特征10000时会自动退出，不报错，这边也会有这样的情况么？还是我自己电脑配置问题

deadlywing · 2023-08-02T08:19:05Z

是的，这是一个复杂度和n相关的矩阵乘
方便给一下复现的代码么，我试一下

tarantula-leo · 2023-08-02T08:33:57Z

使用的代码就是simple_pca.py和simple_pca——test.py，将data设置成：
A = random.normal(random.PRNGKey(0), (15, 10000))

deadlywing · 2023-08-02T08:49:11Z

hello，我运行了一下，不会自动退出(虽然结果是错误的，估计是精度不够了)

deadlywing · 2023-08-02T08:49:28Z

那估计是oom了？

tarantula-leo · 2023-08-02T08:52:30Z

能完整执行完transform么？我这边这步是没执行到的：
print("X_transformed_jax", result[0])

deadlywing · 2023-08-02T08:55:16Z

我这边有执行到，虽然结果是错误的

tarantula-leo · 2023-08-02T10:24:37Z

这看起来不是误差导致的，应该是算法逻辑问题，这里是会出现除0的：
vec /= jnp.linalg.norm(vec)

tarantula-leo · 2023-08-02T10:31:15Z

A = jnp.array([-0.3721109, 0.26423115, -0.18252768, -0.7368197, -0.44030377, -0.1521442, -0.67135346, -0.5908641, 0.73168886, 0.5673026])
A = A.reshape(2,5)

类似这样的数据应该也会出错

deadlywing · 2023-08-03T02:38:55Z

A = jnp.array([-0.3721109, 0.26423115, -0.18252768, -0.7368197, -0.44030377, -0.1521442, -0.67135346, -0.5908641, 0.73168886, 0.5673026])
A = A.reshape(2,5)

类似这样的数据应该也会出错

sorry，这个我没理解，是指输入数据用这个，最后也会出错么？

deadlywing · 2023-08-03T02:59:19Z

这看起来不是误差导致的，应该是算法逻辑问题，这里是会出现除0的：
vec /= jnp.linalg.norm(vec)

理论上，如果这里就除0的话，那说明cov matrix是半正定的，而且提取的主成分个数也太多了，我觉得这个就已经是属于input的问题～

deadlywing · 2023-08-03T03:04:32Z

我简单尝试了一下，应该是power iteration本身的算法精度问题，在特征维度p比较大的时候，可能需要非常大的迭代次数。

我测试到p=800左右，在默认配置下就有点飞了～

So，这样看来，如果您实现的算法能在收敛速度上优于power iteration，我感觉也是很有价值的～

tarantula-leo · 2023-08-03T08:01:29Z

用迭代的应该都会有精度问题吧比如皮尔逊里newton_inverse的实现方案会有类似问题么？是不是迭代的实现方案都无法支持大一些的数据量

deadlywing · 2023-08-03T08:07:52Z

也是类似的，随着矩阵维度和条件数的增大，newton_inverse的精确性也会变差，此时就需要增大迭代次数，所以这也算是一个magic number；

so，您也可以快速尝试一下QR，不过我也不太确定qr迭代的收敛速度比power iteration的收敛速度是否有优势；

tarantula-leo · 2023-08-03T09:39:28Z

单纯的qr迭代效率应该是不够的

import sys
import os
sys.path.append(os.path.join(os.path.dirname(__file__), '../../'))
import jax.numpy as jnp
from jax import random
from sklearn.datasets import load_iris
import spu.utils.simulation as spsim
import spu.spu_pb2 as spu_pb2
class PCA:
    def __init__(self, n_components):
        self.n_components = n_components
        self.eig_val = None
        self.eig_vec = None
        self.k_eigenvectors = None

    def fit(self, X):
        X = X - jnp.mean(X, axis=0)
        C = jnp.cov(X.T, ddof = 0)
        qr = []
        n = C.shape[0]
        Q = jnp.eye(n)

        for i in range(0, 100):
            qr = jnp.linalg.qr(C)
            Q = jnp.dot(Q,qr[0])
            C = jnp.dot(qr[1], qr[0])

        AK = jnp.dot(qr[0], qr[1])
        self.eig_val = jnp.diag(AK)
        klarge_index = jnp.argsort(self.eig_val)[::-1][:self.n_components]
        self.eig_vec = Q
        self.k_eigenvectors = (self.eig_vec.T)[klarge_index]
    
    def transform(self, X):
        return jnp.dot(X - jnp.mean(X, axis=0), self.k_eigenvectors.T)


iris = load_iris()
A = iris.data

# A = jnp.array([-1,-1,0,2,0,-2,0,0,1,1])
# A=A.reshape(2,5)

def fit_and_transform(X, n_components):
    pca = PCA(n_components)
    pca.fit(X)
    return pca.transform(X)

sim_aby = spsim.Simulator.simple(
    3, spu_pb2.ProtocolKind.ABY3, spu_pb2.FieldType.FM64
)
result = spsim.sim_jax(sim_aby, fit_and_transform,  static_argnums=(1,))(A, 2)
print(result)

deadlywing · 2023-08-03T11:13:12Z

确实，光看迭代部分，首先qr分解的cost就不小，然后还有两次矩阵乘；如果收敛速度没有决定性优势的话，感觉应该不如现在的power iteration

tarantula-leo · 2023-08-04T06:53:46Z

对于迭代计算方面有什么设计思路么？像PCA、Newton_inverse还有NMF都会涉及到迭代，目前基本是用固定迭代次数，spu没法用control_flow来实现，而且在数据高维的时候准确性会下降很多。

deadlywing · 2023-08-04T07:17:43Z

对大矩阵而言，很多迭代算法都有这种问题，当前只能通过手动调整迭代次数来达到更高的精度。但是这里我理解有两个精度问题：

迭代算法本身，这部分我理解可以通过查阅文献，理论分析算法收敛速度去判断；像power iteration，qr迭代都是典型的textbook algorhithm，本身比较简单，性能上应该不会最优，可能需要去参考当前成熟线性代数库的算法；（如eign, numpy,这也是为什么这个任务是挑战级别任务的原因哈）
MPC误差，这部分可以考虑环设置为FM128，同时提高fxp来解决

话说回来，对PCA，原本是希望能实现svd算法，一方面能避免计算conv matric，另一方面也能给其他算法提供重要的基础算子。

tarantula-leo · 2023-08-09T04:31:48Z

用gram_schmidt来做qr应该会缓解溢出情况，不过还是需要去处理下scale。

tarantula-leo · 2023-08-09T04:48:33Z

精度损失应该主要在求特征向量上，不是qr分解这里
随着待求解svd矩阵大小增大（这里是和n_component相关），迭代轮数需要增大
另外，如果数据集大小增大需根据明文情况调整scale

import numpy as np
import jax.numpy as jnp
from jax import random

def eigh_power(A):
    n = A.shape[0]
    eig_val = []
    eig_vec = []
    for _ in range(n):
        vec = jnp.ones((n,))
        for _ in range(100):
            # Power iteration
            vec = jnp.dot(A, vec)
            vec /= jnp.linalg.norm(vec)

        eigval = jnp.dot(vec.T, jnp.dot(A, vec))
        eig_vec.append(vec)
        eig_val.append(eigval)
        A -= eigval * jnp.outer(vec, vec)
    eig_vecs = jnp.column_stack(eig_vec)
    eig_vals = jnp.array(eig_val)
    return eig_vals, eig_vecs
    
def svd(A):
    m, n = A.shape
    # sigma, U = jnp.linalg.eigh(jnp.dot(A, A.T))
    # print(jnp.dot(A, A.T))
    sigma, U = eigh_power(jnp.dot(A, A.T))
    # if use eigh_qr 
    # arg_sort = jnp.argsort(sigma)[::-1]
    # sigma = sigma[arg_sort]
    # U = U[:, arg_sort]
    sigma_sqrt = jnp.sqrt(sigma)
    # sigma_matrix = jnp.diag(sigma_sqrt)
    sigma_inv = jnp.diag((1/sigma_sqrt))
    V = jnp.dot(jnp.dot(sigma_inv, U.T), A)
    return (U, sigma_sqrt, V)

def gram_schmidt(A):
    Q=jnp.zeros_like(A)
    cnt = 0
    for a in A.T:
        u = jnp.copy(a)
        for i in range(0, cnt):
            u -= jnp.dot(jnp.dot(Q[:, i].T, a), Q[:, i])
        e = u / jnp.linalg.norm(u)
        Q = Q.at[:,cnt].set(e)
        cnt += 1
    return Q

def iteration(A, Omega, power_iter = 4):
    # iter = 7 if rank < 0.1 * min(A.shape) else 4
    Y = jnp.dot(A, Omega)

    for _ in range(power_iter):
        Y = jnp.dot(A, jnp.dot(A.T, Y))
        # print(Y)
    # Q, _ = jnp.linalg.qr(Y)
    Q = gram_schmidt(Y / 10000)
    return Q

def rsvd(A, Omega):
    Q = iteration(A, Omega)
    B = jnp.dot(Q.T, A)
    # u_tilde, s, v = jnp.linalg.svd(B, full_matrices=False)
    u_tilde, s, v = svd(B)
    u = jnp.dot(Q, u_tilde)
    return u, s, v

A = random.normal(random.PRNGKey(0), (1000,10))
A = A - jnp.mean(A, axis=0)
rank = 5
random_state = np.random.RandomState(0)
Omega = random_state.normal(size=(A.shape[1], rank)) / 10000000

def main(mat,Omega):
    u, s, v = rsvd(mat, Omega)
    v = v[:rank,:]
    return u,s,v

# test
print(main(A,Omega)[1])

# sklearn
from sklearn.utils.extmath import randomized_svd
u1,s1,v1=randomized_svd(A,rank,power_iteration_normalizer = 'QR',svd_lapack_driver = 'gesvd',n_oversamples=0, random_state=0)
print(s1)

# spu
import spu.utils.simulation as spsim
import spu.spu_pb2 as spu_pb2
config = spu_pb2.RuntimeConfig(
        protocol=spu_pb2.ProtocolKind.ABY3,
        field=spu_pb2.FieldType.FM128,
        fxp_fraction_bits=30,
    )
sim_aby = spsim.Simulator(3, config)
result = spsim.sim_jax(sim_aby, main)(A, Omega)
print(result[1])

deadlywing · 2023-08-09T06:57:41Z

精度损失应该主要在求特征向量上，不是qr分解这里
随着待求解svd矩阵大小增大（这里是和n_component相关），迭代轮数需要增大
另外，如果数据集大小增大需根据明文情况调整scale

import numpy as np
import jax.numpy as jnp
from jax import random

def eigh_power(A):
    n = A.shape[0]
    eig_val = []
    eig_vec = []
    for _ in range(n):
        vec = jnp.ones((n,))
        for _ in range(100):
            # Power iteration
            vec = jnp.dot(A, vec)
            vec /= jnp.linalg.norm(vec)

        eigval = jnp.dot(vec.T, jnp.dot(A, vec))
        eig_vec.append(vec)
        eig_val.append(eigval)
        A -= eigval * jnp.outer(vec, vec)
    eig_vecs = jnp.column_stack(eig_vec)
    eig_vals = jnp.array(eig_val)
    return eig_vals, eig_vecs
    
def svd(A):
    m, n = A.shape
    # sigma, U = jnp.linalg.eigh(jnp.dot(A, A.T))
    # print(jnp.dot(A, A.T))
    sigma, U = eigh_power(jnp.dot(A, A.T))
    # if use eigh_qr 
    # arg_sort = jnp.argsort(sigma)[::-1]
    # sigma = sigma[arg_sort]
    # U = U[:, arg_sort]
    sigma_sqrt = jnp.sqrt(sigma)
    # sigma_matrix = jnp.diag(sigma_sqrt)
    sigma_inv = jnp.diag((1/sigma_sqrt))
    V = jnp.dot(jnp.dot(sigma_inv, U.T), A)
    return (U, sigma_sqrt, V)

def gram_schmidt(A):
    Q=jnp.zeros_like(A)
    cnt = 0
    for a in A.T:
        u = jnp.copy(a)
        for i in range(0, cnt):
            u -= jnp.dot(jnp.dot(Q[:, i].T, a), Q[:, i])
        e = u / jnp.linalg.norm(u)
        Q = Q.at[:,cnt].set(e)
        cnt += 1
    return Q

def iteration(A, Omega, power_iter = 4):
    # iter = 7 if rank < 0.1 * min(A.shape) else 4
    Y = jnp.dot(A, Omega)

    for _ in range(power_iter):
        Y = jnp.dot(A, jnp.dot(A.T, Y))
        # print(Y)
    # Q, _ = jnp.linalg.qr(Y)
    Q = gram_schmidt(Y / 10000)
    return Q

def rsvd(A, Omega):
    Q = iteration(A, Omega)
    B = jnp.dot(Q.T, A)
    # u_tilde, s, v = jnp.linalg.svd(B, full_matrices=False)
    u_tilde, s, v = svd(B)
    u = jnp.dot(Q, u_tilde)
    return u, s, v

A = random.normal(random.PRNGKey(0), (1000,10))
A = A - jnp.mean(A, axis=0)
rank = 5
random_state = np.random.RandomState(0)
Omega = random_state.normal(size=(A.shape[1], rank)) / 10000000

def main(mat,Omega):
    u, s, v = rsvd(mat, Omega)
    v = v[:rank,:]
    return u,s,v

# test
print(main(A,Omega)[1])

# sklearn
from sklearn.utils.extmath import randomized_svd
u1,s1,v1=randomized_svd(A,rank,power_iteration_normalizer = 'QR',svd_lapack_driver = 'gesvd',n_oversamples=0, random_state=0)
print(s1)

# spu
import spu.utils.simulation as spsim
import spu.spu_pb2 as spu_pb2
config = spu_pb2.RuntimeConfig(
        protocol=spu_pb2.ProtocolKind.ABY3,
        field=spu_pb2.FieldType.FM128,
        fxp_fraction_bits=30,
    )
sim_aby = spsim.Simulator(3, config)
result = spsim.sim_jax(sim_aby, main)(A, Omega)
print(result[1])

我尝试了一下，特征向量求解应该也是一个bottleneck，但是qr也还是有影响的，用gram_schmidt替代jax的qr以后，能支持的维度大了一些
2,3 :我感觉都可以留一个超参数的口子；

btw，你说的迭代次数具体是指哪里？iteration函数里的还是eigh_power里的？或者两个地方的迭代次数都能放出来给用户调

tarantula-leo · 2023-08-09T07:12:36Z

迭代次数指的是eigh_power里的，simple_pca里的power_iteration迭代轮数和n_feature相关，这里的和n_component相关；iteration里的迭代次数一般就4或者7即可，根据n_component/n_feature比例来。
这个代码加上scale的超参是提PR么？考虑是不用pca路径，新建rsvd的路径。

deadlywing · 2023-08-09T07:38:22Z

迭代次数指的是eigh_power里的，simple_pca里的power_iteration迭代轮数和n_feature相关，这里的和n_component相关；iteration里的迭代次数一般就4或者7即可，根据n_component/n_feature比例来。
这个代码加上scale的超参是提PR么？考虑是不用pca路径，新建rsvd的路径。

个人觉得，您可以在utils文件夹下建一个extmath.py，把我们这次讨论的一些内容放到util里，后续可以作为统一的工具函数，其中可以包括：

rsvd
eigh（qr 和 power iter）
qr分解 (Jax api(即householder) 和 gram_schmidt)

然后，我觉得您也可以在原来的pca里增加一个method=rsvd

不过也需要麻烦您描述一下我们当前这个方案的一些问题，如需要FM128，可能需要调整scale，以及迭代轮数要随着n_component增大而增大等信息（这些可以直接加在PCA那个类的注释doc里～）

tarantula-leo · 2023-08-09T08:24:42Z

大概这么整合：

extmath.py文件里直接使用def，不使用class
pca文件修改后，增加判断条件选择power_iteration或rsvd

deadlywing · 2023-08-09T08:27:20Z

大概这么整合：

extmath.py文件里直接使用def，不使用class

pca文件修改后，增加判断条件选择power_iteration或rsvd

嗯，，我觉得ok

extmath.py主要就是用来提供一些高阶的数学函数，以函数的形式就可以，没必要设计类
用method去区分就可以，fit里面增加一个rsvd的分支即可

tarantula-leo · 2023-08-09T12:14:10Z

调整simulation参数会导致之前的simple_pca出错么？跑了下好像有问题，n_component为5，用的数据集是：
X = random.normal(random.PRNGKey(0), (1000, 10))

        config = spu_pb2.RuntimeConfig(
                protocol=spu_pb2.ProtocolKind.ABY3,
                field=spu_pb2.FieldType.FM128,
                fxp_fraction_bits=30,
            )
        sim = spsim.Simulator(3, config)

transform里也需要修改，去除均值这一步，考虑也通过method区分

        X = X - self._mean
        return jnp.dot(X, self._components)

deadlywing · 2023-08-09T12:31:18Z

我本地跑了观察了一下结果，应该是数据大一点以后，精度没有这么高了所以导致assertion失败了
去均值应该都是需要的吧，即使是用svd

tarantula-leo · 2023-08-09T12:32:58Z

去均值用SVD的话fit需要，transform不用

tarantula-leo · 2023-08-09T12:34:09Z

先提交了extmath的PR，PCA的后面要抽空补充下原理

deadlywing · 2023-08-09T12:37:51Z

https://github.com/scikit-learn/scikit-learn/blob/7f9bad99d/sklearn/decomposition/_base.py#L101

sklearn里是有的，同样也是用的svd算法

deadlywing · 2023-08-09T12:40:08Z

另外，从定义出发，pca本质可以认为找方差最大的投影方向，如果没有去均值化，那么感觉应该容易被scale影响～

tarantula-leo · 2023-08-09T12:47:23Z

我用fit_transform对比结果的，看里面是没有用到：
https://github.com/scikit-learn/scikit-learn/blob/7f9bad99d6e0a3e8ddf92a7e5561245224dab102/sklearn/decomposition/_pca.py

deadlywing · 2023-08-10T01:58:57Z

我用fit_transform对比结果的，看里面是没有用到：
https://github.com/scikit-learn/scikit-learn/blob/7f9bad99d6e0a3e8ddf92a7e5561245224dab102/sklearn/decomposition/_pca.py

fit_transform是因为fit和transform一起做，在fit的时候本身就是去均值的，所以后面就不需要这步了；
但是transform的话一般是计算另外的数据集，所以是需要根据fit保留的均值去做标准化的，

anyway，一个最简单的方式就是检查一下结果，不管用svd还是power iter，transform的结果都是一样的～

tarantula-leo · 2023-08-10T02:16:08Z

之前是用下面的fit_transform来验证结果的，看结果不太一样，如果分开用fit和transform结果是一致的

            # X_new = X * V = U * S * Vt * V = U * S
            U *= S[: self.n_components_]

tarantula-leo · 2023-08-10T08:49:07Z

想问下用emul报错的原因是什么？另外，simulation时需要修改config，在emul里需要怎么进行设置？

[2023-08-10 16:49:29,458] Start multiprocess cluster...
[2023-08-10 16:49:29,819] [MainProcess] Starting grpc server at 127.0.0.1:9920
[2023-08-10 16:49:30,194] [MainProcess] Starting grpc server at 127.0.0.1:9921
[2023-08-10 16:49:30,501] [MainProcess] Run : builtin_spu_init at node:0
[2023-08-10 16:49:30,507] [MainProcess] Run : builtin_spu_init at node:1
I0810 16:49:30.510710 30898 external/com_github_brpc_brpc/src/brpc/server.cpp:1113] Server[yacl::link::internal::ReceiverServiceImpl] is serving on port=9930.
I0810 16:49:30.511318 30898 external/com_github_brpc_brpc/src/brpc/server.cpp:1116] Check out http://node2:9930 in web browser.
I0810 16:49:30.515334 30912 external/com_github_brpc_brpc/src/brpc/server.cpp:1113] Server[yacl::link::internal::ReceiverServiceImpl] is serving on port=9931.
I0810 16:49:30.515905 30912 external/com_github_brpc_brpc/src/brpc/server.cpp:1116] Check out http://node2:9931 in web browser.
[2023-08-10 16:49:30,597] [MainProcess] Starting grpc server at 127.0.0.1:9922
I0810 16:49:30.611969 30924 external/com_github_brpc_brpc/src/brpc/socket.cpp:2345] Checking Socket{id=0 addr=127.0.0.1:9931} (0x7f14600e37c0)
I0810 16:49:30.612149 30924 external/com_github_brpc_brpc/src/brpc/socket.cpp:2405] Revived Socket{id=0 addr=127.0.0.1:9931} (0x7f14600e37c0) (Connectable)
I0810 16:49:30.617444 30934 external/com_github_brpc_brpc/src/brpc/socket.cpp:2345] Checking Socket{id=1 addr=127.0.0.1:9932} (0x7f14600e35c0)
[2023-08-10 16:49:30,999] [MainProcess] Starting grpc server at 127.0.0.1:9923
[2023-08-10 16:49:31,371] [MainProcess] Starting grpc server at 127.0.0.1:9924
I0810 16:49:31.629143 30932 external/com_github_brpc_brpc/src/brpc/socket.cpp:2345] Checking Socket{id=1 addr=127.0.0.1:9932} (0x7f14600e3a00)
I0810 16:49:39.534662 30912 external/com_github_brpc_brpc/src/brpc/server.cpp:1173] Server[yacl::link::internal::ReceiverServiceImpl] is going to quit
[2023-08-10 16:49:39,536] [MainProcess] Traceback (most recent call last):
  File "/root/training/Python3.8/lib/python3.8/site-packages/spu/utils/distributed.py", line 324, in Run
    ret_objs = fn(self, *args, **kwargs)
  File "/root/training/Python3.8/lib/python3.8/site-packages/spu/utils/distributed.py", line 541, in builtin_spu_init
    link = libspu.link.create_brpc(desc, my_rank)
RuntimeError: what: 
        [external/yacl/yacl/link/context.cc:167] connect to mesh failed, failed to setup connection to rank=2
stacktrace: 
#0 pybind11::cpp_function::initialize<>()::{lambda()#3}::_FUN()+0x7f14756d9fed
#1 pybind11::cpp_function::dispatcher()+0x7f14756c1956
#2 PyCFunction_Call+0x43bdca



I0810 16:49:39.735126 30932 external/com_github_brpc_brpc/src/brpc/socket.cpp:2345] Checking Socket{id=0 addr=127.0.0.1:9931} (0x7f14600e37c0)
I0810 16:49:40.530438 30898 external/com_github_brpc_brpc/src/brpc/server.cpp:1173] Server[yacl::link::internal::ReceiverServiceImpl] is going to quit
[2023-08-10 16:49:40,532] [MainProcess] Traceback (most recent call last):
  File "/root/training/Python3.8/lib/python3.8/site-packages/spu/utils/distributed.py", line 324, in Run
    ret_objs = fn(self, *args, **kwargs)
  File "/root/training/Python3.8/lib/python3.8/site-packages/spu/utils/distributed.py", line 541, in builtin_spu_init
    link = libspu.link.create_brpc(desc, my_rank)
RuntimeError: what: 
        [external/yacl/yacl/link/context.cc:167] connect to mesh failed, failed to setup connection to rank=2
stacktrace: 
#0 pybind11::cpp_function::initialize<>()::{lambda()#3}::_FUN()+0x7f14756d9fed
#1 pybind11::cpp_function::dispatcher()+0x7f14756c1956
#2 PyCFunction_Call+0x43bdca



[2023-08-10 16:49:40,534] Shutdown multiprocess cluster...

tpppppub · 2023-08-10T09:26:15Z

想问下用emul报错的原因是什么？另外，simulation时需要修改config，在emul里需要怎么进行设置？

[2023-08-10 16:49:29,458] Start multiprocess cluster...
[2023-08-10 16:49:29,819] [MainProcess] Starting grpc server at 127.0.0.1:9920
[2023-08-10 16:49:30,194] [MainProcess] Starting grpc server at 127.0.0.1:9921
[2023-08-10 16:49:30,501] [MainProcess] Run : builtin_spu_init at node:0
[2023-08-10 16:49:30,507] [MainProcess] Run : builtin_spu_init at node:1
I0810 16:49:30.510710 30898 external/com_github_brpc_brpc/src/brpc/server.cpp:1113] Server[yacl::link::internal::ReceiverServiceImpl] is serving on port=9930.
I0810 16:49:30.511318 30898 external/com_github_brpc_brpc/src/brpc/server.cpp:1116] Check out http://node2:9930 in web browser.
I0810 16:49:30.515334 30912 external/com_github_brpc_brpc/src/brpc/server.cpp:1113] Server[yacl::link::internal::ReceiverServiceImpl] is serving on port=9931.
I0810 16:49:30.515905 30912 external/com_github_brpc_brpc/src/brpc/server.cpp:1116] Check out http://node2:9931 in web browser.
[2023-08-10 16:49:30,597] [MainProcess] Starting grpc server at 127.0.0.1:9922
I0810 16:49:30.611969 30924 external/com_github_brpc_brpc/src/brpc/socket.cpp:2345] Checking Socket{id=0 addr=127.0.0.1:9931} (0x7f14600e37c0)
I0810 16:49:30.612149 30924 external/com_github_brpc_brpc/src/brpc/socket.cpp:2405] Revived Socket{id=0 addr=127.0.0.1:9931} (0x7f14600e37c0) (Connectable)
I0810 16:49:30.617444 30934 external/com_github_brpc_brpc/src/brpc/socket.cpp:2345] Checking Socket{id=1 addr=127.0.0.1:9932} (0x7f14600e35c0)
[2023-08-10 16:49:30,999] [MainProcess] Starting grpc server at 127.0.0.1:9923
[2023-08-10 16:49:31,371] [MainProcess] Starting grpc server at 127.0.0.1:9924
I0810 16:49:31.629143 30932 external/com_github_brpc_brpc/src/brpc/socket.cpp:2345] Checking Socket{id=1 addr=127.0.0.1:9932} (0x7f14600e3a00)
I0810 16:49:39.534662 30912 external/com_github_brpc_brpc/src/brpc/server.cpp:1173] Server[yacl::link::internal::ReceiverServiceImpl] is going to quit
[2023-08-10 16:49:39,536] [MainProcess] Traceback (most recent call last):
  File "/root/training/Python3.8/lib/python3.8/site-packages/spu/utils/distributed.py", line 324, in Run
    ret_objs = fn(self, *args, **kwargs)
  File "/root/training/Python3.8/lib/python3.8/site-packages/spu/utils/distributed.py", line 541, in builtin_spu_init
    link = libspu.link.create_brpc(desc, my_rank)
RuntimeError: what: 
        [external/yacl/yacl/link/context.cc:167] connect to mesh failed, failed to setup connection to rank=2
stacktrace: 
#0 pybind11::cpp_function::initialize<>()::{lambda()#3}::_FUN()+0x7f14756d9fed
#1 pybind11::cpp_function::dispatcher()+0x7f14756c1956
#2 PyCFunction_Call+0x43bdca



I0810 16:49:39.735126 30932 external/com_github_brpc_brpc/src/brpc/socket.cpp:2345] Checking Socket{id=0 addr=127.0.0.1:9931} (0x7f14600e37c0)
I0810 16:49:40.530438 30898 external/com_github_brpc_brpc/src/brpc/server.cpp:1173] Server[yacl::link::internal::ReceiverServiceImpl] is going to quit
[2023-08-10 16:49:40,532] [MainProcess] Traceback (most recent call last):
  File "/root/training/Python3.8/lib/python3.8/site-packages/spu/utils/distributed.py", line 324, in Run
    ret_objs = fn(self, *args, **kwargs)
  File "/root/training/Python3.8/lib/python3.8/site-packages/spu/utils/distributed.py", line 541, in builtin_spu_init
    link = libspu.link.create_brpc(desc, my_rank)
RuntimeError: what: 
        [external/yacl/yacl/link/context.cc:167] connect to mesh failed, failed to setup connection to rank=2
stacktrace: 
#0 pybind11::cpp_function::initialize<>()::{lambda()#3}::_FUN()+0x7f14756d9fed
#1 pybind11::cpp_function::dispatcher()+0x7f14756c1956
#2 PyCFunction_Call+0x43bdca



[2023-08-10 16:49:40,534] Shutdown multiprocess cluster...

请提供下代码和复现步骤
创建 emulator 的时候直接使用你修改后的 config.json 路径即可

tarantula-leo · 2023-08-10T10:25:59Z

emulation的config.json文件里没看到这个参数，只有调整field的：
fxp_fraction_bits=30
在simulation里是这样的：

        config = spu_pb2.RuntimeConfig(
                protocol=spu_pb2.ProtocolKind.ABY3,
                field=spu_pb2.FieldType.FM128,
                fxp_fraction_bits=30,
            )
        sim = spsim.Simulator(3, config)

可以直接在config.json里添加这个参数么？

tpppppub · 2023-08-10T11:00:46Z

emulation的config.json文件里没看到这个参数，只有调整field的： fxp_fraction_bits=30 在simulation里是这样的：
        config = spu_pb2.RuntimeConfig(
                protocol=spu_pb2.ProtocolKind.ABY3,
                field=spu_pb2.FieldType.FM128,
                fxp_fraction_bits=30,
            )
        sim = spsim.Simulator(3, config)
可以直接在config.json里添加这个参数么？

runtime_config 里加上 "fxp_fraction_bits": 30

deadlywing · 2023-08-10T11:14:16Z

util那个pr已经merge，pca麻烦重新发一个pr

tarantula-leo · 2023-08-10T12:36:21Z

配置问题是要新建一个config.json的吗？

tpppppub · 2023-08-10T12:39:07Z

你在 PCA 子目录下先建一个吧

tarantula-leo · 2023-08-11T02:18:27Z

util那个pr已经merge，pca麻烦重新发一个pr

已经提交，需要看下emul部分是否会报错。

# Pull Request ## What problem does this PR solve? 使用 SPU 优化 PCA 算法 Issue Number: Fixed #259 ## Possible side effects? - Performance: 1. 收敛速度更快(体现在能支持更大的特征维度) 2. 不需要显示的计算原数据集的协方差矩阵 - Backward compatibility:

Candicepan added this to OSCP Phase 2 Jul 21, 2023

Candicepan converted this from a draft issue Jul 21, 2023

Candicepan added OSCP SecretFlow Open Source Contribution Plan challenge labels Jul 21, 2023

Candicepan assigned tarantula-leo Aug 1, 2023

Candicepan moved this from Needs Triage to In Progress in OSCP Phase 2 Aug 1, 2023

tarantula-leo mentioned this issue Aug 11, 2023

[Add Randomized SVD in PCA] #298

Closed

Candicepan moved this from In Progress to In Review in OSCP Phase 2 Aug 11, 2023

tarantula-leo mentioned this issue Aug 11, 2023

[Add Randomized SVD in PCA] #300

Merged

deadlywing closed this as completed in #300 Aug 15, 2023

github-project-automation bot moved this from In Review to Done in OSCP Phase 2 Aug 15, 2023

使用 SPU 优化 PCA 算法 #259

使用 SPU 优化 PCA 算法 #259

Comments

Candicepan commented Jul 21, 2023 • edited Loading

任务介绍

详细要求

能力要求

操作说明

开发须知

tarantula-leo commented Aug 1, 2023

Candicepan commented Aug 1, 2023

tarantula-leo commented Aug 1, 2023

deadlywing commented Aug 1, 2023

tarantula-leo commented Aug 1, 2023

deadlywing commented Aug 1, 2023

tarantula-leo commented Aug 2, 2023

deadlywing commented Aug 2, 2023

tarantula-leo commented Aug 2, 2023

deadlywing commented Aug 2, 2023

deadlywing commented Aug 2, 2023

tarantula-leo commented Aug 2, 2023

deadlywing commented Aug 2, 2023

tarantula-leo commented Aug 2, 2023

tarantula-leo commented Aug 2, 2023

deadlywing commented Aug 3, 2023

deadlywing commented Aug 3, 2023

deadlywing commented Aug 3, 2023

tarantula-leo commented Aug 3, 2023

deadlywing commented Aug 3, 2023

tarantula-leo commented Aug 3, 2023

deadlywing commented Aug 3, 2023

tarantula-leo commented Aug 4, 2023

deadlywing commented Aug 4, 2023

tarantula-leo commented Aug 9, 2023

tarantula-leo commented Aug 9, 2023

deadlywing commented Aug 9, 2023

tarantula-leo commented Aug 9, 2023

deadlywing commented Aug 9, 2023

tarantula-leo commented Aug 9, 2023

deadlywing commented Aug 9, 2023

tarantula-leo commented Aug 9, 2023

deadlywing commented Aug 9, 2023

tarantula-leo commented Aug 9, 2023

tarantula-leo commented Aug 9, 2023

deadlywing commented Aug 9, 2023

deadlywing commented Aug 9, 2023

tarantula-leo commented Aug 9, 2023

deadlywing commented Aug 10, 2023

tarantula-leo commented Aug 10, 2023

tarantula-leo commented Aug 10, 2023

tpppppub commented Aug 10, 2023

tarantula-leo commented Aug 10, 2023

tpppppub commented Aug 10, 2023

deadlywing commented Aug 10, 2023

tarantula-leo commented Aug 10, 2023

tpppppub commented Aug 10, 2023

tarantula-leo commented Aug 11, 2023

Candicepan commented Jul 21, 2023 •

edited

Loading