FFNet.md

PyTorch-FFNet

Environment Setup

Setup AI Model Efficiency Toolkit

Please install and setup AIMET before proceeding further. This model was tested with the torch_gpu variant of AIMET 1.23.

Additional Dependencies

Install skimage as follows

pip install scikit-image

Add AIMET Model Zoo to the pythonpath

export PYTHONPATH=$PYTHONPATH:<path to parent of aimet_model_zoo_path>

Dataset

The Cityscape Dataset can be downloaded from here:

• https://www.cityscapes-dataset.com/

---

Usage

To run evaluation with QuantSim in AIMET, use the following

python3  aimet_zoo_torch/ffnet/evaluators/ffnet_quanteval.py \
                --model-config <configuration to be tested> \
                --dataset-path <path to directory containing CityScapes> \
                --batch-size  <batch size as an integer value, defaults to 2> \

Available model configurations are:

• segmentation_ffnet40S_dBBB_mobile
• segmentation_ffnet54S_dBBB_mobile
• segmentation_ffnet78S_BCC_mobile_pre_down
• segmentation_ffnet78S_BCC_mobile_pre_down
• segmentation_ffnet122NS_CCC_mobile_pre_down

---

Model checkpoint and configuration

• The original prepared FFNet checkpoint can be downloaded from here:
  • https://github.com/quic/aimet-model-zoo/releases/tag/torch_segmentation_ffnet
• The Quantization Simulation (Quantsim) Configuration file can be downloaded from here: default_config_per_channel.json (Please see this page for more information on this file).

---

Quantization Configuration (INT8)

• Weight quantization: 8 bits, per channel symmetric quantization
• Bias parameters are not quantized
• Activation quantization: 8 bits, asymmetric quantization
• Model inputs are quantized
• TF-Enhanced was used as quantization scheme
• Cross layer equalization (CLE) has been applied on optimized checkpoint
• for low resolution models with pre_down suffix, the GaussianConv2D layer is disabled for quantization.

Results

Below are the mIoU results of the PyTorch FFNet model for the Cityscapes dataset:

Model Configuration	FP32 (%)	INT8 (%)
segmentation_ffnet78S_dBBB_mobile	81.3	80.7
segmentation_ffnet54S_dBBB_mobile	80.8	80.1
segmentation_ffnet40S_dBBB_mobile	79.2	78.9
segmentation_ffnet78S_BCC_mobile_pre_down	80.6	80.4
segmentation_ffnet122NS_CCC_mobile_pre_down	79.3	79.0