feat: improve accuracy of p_error simulation in virtual circuits

2026-02-08 19:44:57 -05:00 · 2023-02-07 12:10:13 +01:00
parent 840c0eba8c
commit 656761346a
6 changed files with 90 additions and 10 deletions
--- a/concrete/numpy/representation/graph.py
+++ b/concrete/numpy/representation/graph.py
@@ -2,17 +2,21 @@
 Declaration of `Graph` class.
 """

+import math
 import re
 from copy import deepcopy
 from typing import Any, Dict, Iterable, List, Optional, Tuple, Union

 import networkx as nx
 import numpy as np
+import scipy.special

 from ..dtypes import Float, Integer, UnsignedInteger
 from .node import Node
 from .operation import Operation

+P_ERROR_PER_ERROR_SIZE_CACHE: Dict[float, Dict[int, float]] = {}
+

 class Graph:
    """
@@ -81,6 +85,8 @@ class Graph:
                nodes and their values during computation
        """

+        # pylint: disable=no-member,too-many-nested-blocks
+
        if p_error is None:
            p_error = 0.0

@@ -106,16 +112,46 @@ class Graph:
                    if pred_node.operation != Operation.Input:
                        dtype = node.inputs[index].dtype
                        if isinstance(dtype, Integer):
-                            # this is not the real behavior of FHE
-                            # it's a simplified model, and it will be replaced at one point
+                            # see https://github.com/zama-ai/concrete-numpy/blob/main/docs/_static/p_error_simulation.pdf  # noqa: E501  # pylint: disable=line-too-long
+                            # to learn more about the distribution of error
+
+                            if p_error not in P_ERROR_PER_ERROR_SIZE_CACHE:
+                                std_score = math.sqrt(2) * scipy.special.erfcinv(p_error)
+                                p_error_per_error_size = {}
+
+                                error_size = 1
+                                last_p = 1 - p_error
+                                while last_p != 1.0 or error_size == 1:
+                                    new_std_score = (2 * error_size + 1) * std_score
+                                    new_p = scipy.special.erf(new_std_score / math.sqrt(2))
+
+                                    p_error_per_error_size[error_size] = new_p - last_p
+
+                                    last_p = new_p
+                                    error_size += 1
+
+                                # ordering of `p_error_per_error_size` is relied on
+                                # during the introduction of the error below
+                                # thus we explicitly sort it to make sure it's ordered
+                                p_error_per_error_size = dict(
+                                    sorted(p_error_per_error_size.items())
+                                )
+
+                                P_ERROR_PER_ERROR_SIZE_CACHE[p_error] = p_error_per_error_size
+                            else:  # pragma: no cover
+                                p_error_per_error_size = P_ERROR_PER_ERROR_SIZE_CACHE[p_error]

                            error = np.random.rand(*pred_results[index].shape)
-                            error = np.where(error < p_error**3, 3, error)
-                            error = np.where(error < p_error**2, 2, error)
-                            error = np.where(error < p_error, 1, np.where(error > 1, error, 0))
+
+                            accumulated_p_error = 0.0
+                            for error_size, p_error_for_size in p_error_per_error_size.items():
+                                accumulated_p_error += p_error_for_size
+                                error = np.where(error < accumulated_p_error, error_size, error)
+
+                            error = np.where(error < 1, 0, error).astype(np.int64)

                            error_sign = np.random.rand(*pred_results[index].shape)
-                            error_sign = np.where(error < 0.5, 1, -1)
+                            error_sign = np.where(error_sign < 0.5, 1, -1).astype(np.int64)

                            new_results = pred_results[index] + (error * error_sign)

--- a/docs/_static/p_error_simulation.pdf
+++ b/docs/_static/p_error_simulation.pdf
--- a/docs/linux.dependency.licenses.txt
+++ b/docs/linux.dependency.licenses.txt
@@ -15,6 +15,7 @@
 packaging                 23.0        Apache Software License; BSD License
 pyparsing                 3.0.9       MIT License
 python-dateutil           2.8.2       Apache Software License; BSD License
+ scipy                     1.7.3       BSD License
 six                       1.16.0      MIT License
 torch                     1.13.1      BSD License
 typing-extensions         3.10.0.2    Python Software Foundation License
--- a/poetry.lock
+++ b/poetry.lock
@@ -3274,6 +3274,48 @@ files = [
    {file = "ruff-0.0.191.tar.gz", hash = "sha256:d698c4d5e3b2963cbbb7c2728f404091d5c47cdf8d94db3eb2f335e2a93a6b1b"},
 ]

+[[package]]
+name = "scipy"
+version = "1.7.3"
+description = "SciPy: Scientific Library for Python"
+category = "main"
+optional = false
+python-versions = ">=3.7,<3.11"
+files = [
+    {file = "scipy-1.7.3-1-cp310-cp310-macosx_12_0_arm64.whl", hash = "sha256:c9e04d7e9b03a8a6ac2045f7c5ef741be86727d8f49c45db45f244bdd2bcff17"},
+    {file = "scipy-1.7.3-1-cp38-cp38-macosx_12_0_arm64.whl", hash = "sha256:b0e0aeb061a1d7dcd2ed59ea57ee56c9b23dd60100825f98238c06ee5cc4467e"},
+    {file = "scipy-1.7.3-1-cp39-cp39-macosx_12_0_arm64.whl", hash = "sha256:b78a35c5c74d336f42f44106174b9851c783184a85a3fe3e68857259b37b9ffb"},
+    {file = "scipy-1.7.3-cp310-cp310-macosx_10_9_x86_64.whl", hash = "sha256:173308efba2270dcd61cd45a30dfded6ec0085b4b6eb33b5eb11ab443005e088"},
+    {file = "scipy-1.7.3-cp310-cp310-macosx_12_0_arm64.whl", hash = "sha256:21b66200cf44b1c3e86495e3a436fc7a26608f92b8d43d344457c54f1c024cbc"},
+    {file = "scipy-1.7.3-cp310-cp310-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:ceebc3c4f6a109777c0053dfa0282fddb8893eddfb0d598574acfb734a926168"},
+    {file = "scipy-1.7.3-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:f7eaea089345a35130bc9a39b89ec1ff69c208efa97b3f8b25ea5d4c41d88094"},
+    {file = "scipy-1.7.3-cp310-cp310-win_amd64.whl", hash = "sha256:304dfaa7146cffdb75fbf6bb7c190fd7688795389ad060b970269c8576d038e9"},
+    {file = "scipy-1.7.3-cp37-cp37m-macosx_10_9_x86_64.whl", hash = "sha256:033ce76ed4e9f62923e1f8124f7e2b0800db533828c853b402c7eec6e9465d80"},
+    {file = "scipy-1.7.3-cp37-cp37m-manylinux_2_12_i686.manylinux2010_i686.whl", hash = "sha256:4d242d13206ca4302d83d8a6388c9dfce49fc48fdd3c20efad89ba12f785bf9e"},
+    {file = "scipy-1.7.3-cp37-cp37m-manylinux_2_12_x86_64.manylinux2010_x86_64.whl", hash = "sha256:8499d9dd1459dc0d0fe68db0832c3d5fc1361ae8e13d05e6849b358dc3f2c279"},
+    {file = "scipy-1.7.3-cp37-cp37m-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:ca36e7d9430f7481fc7d11e015ae16fbd5575615a8e9060538104778be84addf"},
+    {file = "scipy-1.7.3-cp37-cp37m-win32.whl", hash = "sha256:e2c036492e673aad1b7b0d0ccdc0cb30a968353d2c4bf92ac8e73509e1bf212c"},
+    {file = "scipy-1.7.3-cp37-cp37m-win_amd64.whl", hash = "sha256:866ada14a95b083dd727a845a764cf95dd13ba3dc69a16b99038001b05439709"},
+    {file = "scipy-1.7.3-cp38-cp38-macosx_10_9_x86_64.whl", hash = "sha256:65bd52bf55f9a1071398557394203d881384d27b9c2cad7df9a027170aeaef93"},
+    {file = "scipy-1.7.3-cp38-cp38-macosx_12_0_arm64.whl", hash = "sha256:f99d206db1f1ae735a8192ab93bd6028f3a42f6fa08467d37a14eb96c9dd34a3"},
+    {file = "scipy-1.7.3-cp38-cp38-manylinux_2_12_i686.manylinux2010_i686.whl", hash = "sha256:5f2cfc359379c56b3a41b17ebd024109b2049f878badc1e454f31418c3a18436"},
+    {file = "scipy-1.7.3-cp38-cp38-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:eb7ae2c4dbdb3c9247e07acc532f91077ae6dbc40ad5bd5dca0bb5a176ee9bda"},
+    {file = "scipy-1.7.3-cp38-cp38-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:95c2d250074cfa76715d58830579c64dff7354484b284c2b8b87e5a38321672c"},
+    {file = "scipy-1.7.3-cp38-cp38-win32.whl", hash = "sha256:87069cf875f0262a6e3187ab0f419f5b4280d3dcf4811ef9613c605f6e4dca95"},
+    {file = "scipy-1.7.3-cp38-cp38-win_amd64.whl", hash = "sha256:7edd9a311299a61e9919ea4192dd477395b50c014cdc1a1ac572d7c27e2207fa"},
+    {file = "scipy-1.7.3-cp39-cp39-macosx_10_9_x86_64.whl", hash = "sha256:eef93a446114ac0193a7b714ce67659db80caf940f3232bad63f4c7a81bc18df"},
+    {file = "scipy-1.7.3-cp39-cp39-macosx_12_0_arm64.whl", hash = "sha256:eb326658f9b73c07081300daba90a8746543b5ea177184daed26528273157294"},
+    {file = "scipy-1.7.3-cp39-cp39-manylinux_2_12_i686.manylinux2010_i686.whl", hash = "sha256:93378f3d14fff07572392ce6a6a2ceb3a1f237733bd6dcb9eb6a2b29b0d19085"},
+    {file = "scipy-1.7.3-cp39-cp39-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:edad1cf5b2ce1912c4d8ddad20e11d333165552aba262c882e28c78bbc09dbf6"},
+    {file = "scipy-1.7.3-cp39-cp39-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:5d1cc2c19afe3b5a546ede7e6a44ce1ff52e443d12b231823268019f608b9b12"},
+    {file = "scipy-1.7.3-cp39-cp39-win32.whl", hash = "sha256:2c56b820d304dffcadbbb6cbfbc2e2c79ee46ea291db17e288e73cd3c64fefa9"},
+    {file = "scipy-1.7.3-cp39-cp39-win_amd64.whl", hash = "sha256:3f78181a153fa21c018d346f595edd648344751d7f03ab94b398be2ad083ed3e"},
+    {file = "scipy-1.7.3.tar.gz", hash = "sha256:ab5875facfdef77e0a47d5fd39ea178b58e60e454a4c85aa1e52fcb80db7babf"},
+]
+
+[package.dependencies]
+numpy = ">=1.16.5,<1.23.0"
+
 [[package]]
 name = "secretstorage"
 version = "3.3.3"
@@ -3846,4 +3888,4 @@ testing = ["flake8 (<5)", "func-timeout", "jaraco.functools", "jaraco.itertools"
 [metadata]
 lock-version = "2.0"
 python-versions = ">=3.7,<3.11"
-content-hash = "1ba82f6d92e8cfd1636e23c6ad7374408f27cf86bb3d80c77ae9c2b109e8e8b8"
+content-hash = "34cb8b9ae8d4245e0d2f98b37cd5c11288e55e4968fa05313b853510019b0ef6"
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -44,6 +44,7 @@ numpy = "^1.21.0"
 Pillow = "^9.0.0"
 concrete-compiler = "^0.23.4"
 torch = "^1.13.1"
+scipy = "1.7.3"

 [tool.poetry.dev-dependencies]
 isort = "^5.10.1"
--- a/tests/compilation/test_circuit.py
+++ b/tests/compilation/test_circuit.py
@@ -302,9 +302,9 @@ def test_bad_server_save(helpers):
    assert str(excinfo.value) == "Just-in-Time compilation cannot be saved"


-@pytest.mark.parametrize("p_error", [0.5, 0.1, 0.01])
+@pytest.mark.parametrize("p_error", [0.75, 0.5, 0.4, 0.25, 0.2, 0.1, 0.01, 0.001])
@pytest.mark.parametrize("bit_width", [10])
-@pytest.mark.parametrize("sample_size", [100_000])
+@pytest.mark.parametrize("sample_size", [1_000_000])
@pytest.mark.parametrize("tolerance", [0.075])
 def test_virtual_p_error(p_error, bit_width, sample_size, tolerance, helpers):
    """
@@ -333,7 +333,7 @@ def test_virtual_p_error(p_error, bit_width, sample_size, tolerance, helpers):
        expected_number_of_errors_on_average - (expected_number_of_errors_on_average * tolerance),
        expected_number_of_errors_on_average + (expected_number_of_errors_on_average * tolerance),
    ]
-    assert acceptable_number_of_errors[0] < errors < acceptable_number_of_errors[1]
+    assert acceptable_number_of_errors[0] <= errors <= acceptable_number_of_errors[1]


 def test_circuit_run_with_unused_arg(helpers):