This repository has been archived by the owner on Apr 29, 2024. It is now read-only.
v0.9.0
·
2 commits
to release/0.9.x
since this release
Summary
The new version of Concrete NumPy comes with improved performance, 16-bit integer support, a new compilation API, and lot's of other improvements!
Links
Docker Image: zamafhe/concrete-numpy:v0.9.0
PyPI Package: https://pypi.org/project/concrete-numpy/0.9.0
v0.9.0
Feature
- Add error probability properties to circuits (
b842bb6) - Make p_error simulation in virtual circuits faster and a bit more realistic (
39958cf) - Make both p error and global p error optional (
1472c8f) - Bump compiler to v0.23.1 (
df8d34a) - Allow fine-grained control on verbosity settings (
a25ba9b) - Support up to 16-bit signed integers with table lookups (
23337e9) - Bump compiler to v0.22.1 (
6376c8e) - Consider p_error in virtual circuits (
702375f) - Introduce tag extension, create integer range getter for graphs, allow filtering in integer bit width getter (
25b9a59) - Create auto adjust rounders configuration option (
2b84def) - Introduce auto rounders (
ccd3f9a) - Introduce round bit pattern extension for virtual circuits (
c552a95) - Introduce maxpool extension for virtual circuits (
cb9cbb0) - Support multiplication between two encrypted values in virtual circuits (
3de24a8) - Improve fusing related error messages (
f0014dc) - Support axes kwarg for np.transpose (
a262408) - Bump compiler to v0.21.0 (
821c61e) - Add support for np.expand_dims (
d6ba763) - Introduce 'global_p_error' configuration (
ea34aeb) - Support up to 16-bit table lookups (
21a0923) - Bump compiler to 0.20.0 (
191150b) - Introduce circuit decorator to directly define circuits (
66c707c)
Fix
- Allow explicitly setting axis to None in np.sum (
416ca69) - Use proper dtype for bias during convolution (
58689d5) - Correctly transfer sign information in direct definition (
0827cb9) - Use builtin math instead of numpy for log2 and ceil to avoid overflow issues (
bac9792) - Use pads arg for torch evaluation (
5f07a72) - Adhere to ONNX spec of kernel_shape (
7901d88)