MPC Quantized Machine Learning- Jacobi SVD #952
Conversation
MPC Quantized Machine Learning- Jacobi SVD
|
I have read the CLA Document and I hereby sign the CLA You can retrigger this bot by commenting recheck in this Pull Request. Posted by the CLA Assistant Lite bot. |
|
Stale pull request message. Please comment to remove stale tag. Otherwise this pr will be closed soon. |
| S = (A + A.T) / 2 | ||
| return S | ||
|
|
||
| n = 10 |
There was a problem hiding this comment.
算法实现的文件中不要出现这种全局的测试，你可以在tests/jacobi_test.py中测试明文实现 或者 密态精度
| A, _, final_iterations = lax.fori_loop(0, max_iter, body_fun, (A, max_off_diag, iterations)) | ||
|
|
||
| singular_values = jnp.abs(jnp.diag(A)) | ||
| idx = jnp.argsort(-singular_values) |
There was a problem hiding this comment.
我理解你就只是要把特征值从大到小排序，直接sort就行啊？
singular_values[idx]还会做一次secret index，没有必要
| idx = jnp.argsort(-singular_values) | ||
| singular_values = singular_values[idx] | ||
|
|
||
| return singular_values |
| return A | ||
|
|
||
| @jit | ||
| def jacobi_svd(A, tol=1e-10, max_iter=5): |
There was a problem hiding this comment.
麻烦增加一些必要的注释信息，起码需要包括：
- 明文下原始算法名称（最好包括ref）
- (optional)若有与明文算法不同处，可以简单概括并说明一下
- 涉及的参数介绍 以及 返回值介绍
| print("Start Jacobi SVD emulation.") | ||
|
|
||
| def proc_transform(A): | ||
| singular_values = jacobi_svd(A) |
There was a problem hiding this comment.
tips：默认配置下，FM64的fxp_bits = 18, 故最小的非0正数为2**-18，如果有稳定性问题，可以试试把tol改大一些？
There was a problem hiding this comment.
请尽量把算法实现定义在jacobi_svd.py中，在emul文件中仅需要import即可
|
另外，请先sign CLA： 复制一下上图红框里的话comment一下即可 |
|
PLUS: SPU重构了构建系统，python代码的构建推荐使用wrapper函数 |
|
|
||
| n = 10 | ||
|
|
||
| A_jax = generate_symmetric_matrix(n) |
There was a problem hiding this comment.
这个算法只适用于对称矩阵么？如果是的话也可以直接在docstring里注明一下。。
MPC Quantized Machine Learning- Jacobi SVD
Pull Request
What problem does this PR solve?
Issue Number: Fixed #
Possible side effects?
Performance:
Backward compatibility: