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.

Sign up for GitHub

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

Jump to bottom

Test impact of symmetric vs. affine quantization #4

Open

etrommer opened this issue Sep 20, 2023 · 2 comments

Assignees

Labels

research

Owner

etrommer commented Sep 20, 2023

TorchApprox has been extended to use PyTorch's internal affine quantization: https://github.com/etrommer/torch-approx/tree/feature/torchquant

It needs to be benchmarked in order to assess whether this improves inference accuracy.

Experimental setups:

8-bit affine vs. 8-bit symmetric for a choice of networks
lower bitwidth (4-bit?) as it is expected to produce more pronounced results

etrommer added the research label

etrommer self-assigned this

Owner Author

etrommer commented Oct 13, 2023 •

edited

Loading

231013_affine_vs_per_tensor_lenet5.csv

Results for LeNet5/MNIST
2,3,4-Bit EvoApprox multipliers
Default gradient_clip_val = 0.5
3-Bit multipliers rerun with gradient_clip_val = 1.0

Owner Author

etrommer commented Oct 16, 2023

231016_affine_vs_per_tensor_resnet8.csv

Results for ResNet8/CIFAR10
8x4-Bit EvoApprox multipliers
Default gradient_clip_val = 0.5

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment