Pytorch w8a8 SalsaNext quantsim model

Optimized w8a8 checkpoint, encoding and FP32 checkpoint for Pytorch SalsaNext model.

For w8a8 optimization:

• Batch norm folding followed by AdaRound in per channel mode have been applied on the original FP32 model.
• Percentile was used in per channel mode for quantsim.
• one operator activation output is enabled with 16-bit width.

PyTorch ResNext101 Model Checkpoints and Encodings

This releases ResNext101 w8a8 optimized quantized weight and encodings

PyTorch Resnet101 W8A8 for image classification

This release model artifacts of ResNet101 w8a8 model artifacts and encodings for image classification task.

Pytorch w4a8 RangeNet++ quantsim model

Optimized w4a8 checkpoint, encoding and FP32 checkpoint for Pytorch RangeNet++ model.

For w4a8 optimization:

• BatchNorm folding followed by AdaRound in per channel mode have been applied on the original FP32 model.
• Percentile was used in per channel mode for quantsim.

Pytorch w4a8 HRNET-PoseNet quantsim model

Optimized w4a8 checkpoint, encoding and FP32 checkpoint for Pytorch HRNET-PoseNet model.

For w4a8 optimization:

• Batch norm folding followed by AdaRound in per channel mode have been applied on the original FP32 instance of HRNET-PoseNet model.
• TF quantization scheme was used in per channel mode for quantsim.

PyTorch GPUNet0 model

Optimized w8a8 checkpoint, encoding and FP32 checkpoint for Pytorch GPUNet0 model.

For w8a8 optimization:

• Adaround followed by bn_fold_to_scale in per channel mode have been applied on the original FP32 model.
• Percentile was used in per channel mode for quantsim.

AIMET Model Zoo February 2023 release (1.1.0)

Release summary

• Added code, documentation and artifacts for the following new PyTorch models: SalsaNext
• Added Pytorch w4a8 support for HRNET-Posenet, QuickSRNet and RangeNet++ models
• Model refactoring and code updates for a few models (HRNet, classification models, super resolution models, ViT, MobileViT, gpt2)
• Cleaned up more code to minimize errors from static analysis (pylint)
• Adding pylint capability via cmake and Jenkins pull request mechanism
• Released second package release of aimet model zoo (installable wheel file binaries)

Artifacts summary

• AIMET model zoo python wheel file for PyTorch and TensorFlow models and evaluators

Package Installation

• Please click here for instructions to install the pre-built AIMET model zoo python package.

Phase 2 Model Artifacts - February 2023

Artifacts for:

1. ABPN
2. SESR
3. XLSR
4. ResNet
5. RegNet
6. HRNet-w48 Semantic Segmentation
7. QuickSRNet-Large-4x

Pytorch RangeNet++ model (W8A8)

Optimized W8A8 checkpoint, encoding and FP32 checkpoint for Pytorch RangeNet++ model.
For W8A8 optimization:

• Batchnorm folding followed by AdaRound in per channel mode have been applied on the original FP32 model.
• Percentile was used in per channel mode for quantsim.

AIMET Model Zoo January 2023 release (1.0.0)

Release summary

1. Cleaned up code to minimize errors from static analysis (pylint)
2. Refactored files and folders for 8 models into evaluators, models, dataloaders, etc to make it more self-contained and automated (phase 2 update) - DeepLabV3, EfficientNetLite0, FFNet, HRNET-Posenet, Inverseform, MobileNetV2, QuickSRNet, SSD-MobileNetV2
3. Added package generation capability via cmake and TOML
4. Added the PyTorch GPT2 and RangetNet++ models (w8a8 only)

Artifacts summary

• AIMET model zoo python wheel file for PyTorch and TensorFlow models and evaluators

Package Installation

• Please click here for instructions to install the pre-built AIMET model zoo python package.