etrommer / agn-approx Goto Github PK

I can't find the source code for the class EvoApproxLib() in the Multiplier Assignment section of the 1_mnist_ipynb file, could you please help me? Thank you very much.

Evaluate accuracy of HTP models on several architecture:

• Evaluation for all 8-Bit EvoApprox Multipliers
• MobileNetV2 & TinyImageNet
• VGG16 & TinyImageNet
• Resnet8 & CIFAR10

Models to evaluate (all evaluation with behavioral simulation):

• Accurate Retraining
• Behavioral Simulation
• HTP Model
• Noise

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

Sphinx documentation currently needs to be built locally due to non-public dependencies. Move to CI as soon as torchapprox is published.

MLFlow logging is currently disabled in the trainer instance if the log_mlflow parameter is not set. The logging instance however still creates an experiment. Import and usage of MLFlow could be made fully dependent on an environment variable instead.