YOLOX.md

PyTorch-YOLOX

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 1.25.

Additional Dependencies

Install necessary dependencies as follows:

pip install pycocotools

Add AIMET Model Zoo to the pythonpath

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

Dataset

The MS-COCO 2017 validation dataset can be downloaded from here:

• http://images.cocodataset.org/zips/val2017.zip

The dataset's root folder (which you pass an as arg), should have 2 subfolders: annotations, and images. The annotations folder should contain only json files. The images folder should contain three subfolders: train2017, val2017, test2017.

---

Usage

To run evaluation with QuantSim in AIMET, use the following

python3 aimet_zoo_torch/yolox/evaluators/yolox_quanteval.py \
                --model-config <configuration to be tested> \
                --dataset-path <Path to MS-COCO 2017> \
                --batch-size <Number of images per batch, default is 64>

Available model configurations are:

• yolox_s
• yolox_l

---

Model checkpoint and configuration

• The original prepared YOLOX checkpoint can be downloaded from here:
  • https://github.com/quic/aimet-model-zoo/releases/download/torch_yolox_int8
• 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
• Percentile was used as quantization scheme
  • percentile value is set as 99.9942 by searching for YOLOX-s
  • percentile value is set as 99.99608 by searching for YOLOX-l
• BatchNorm Folding (BNF) has been applied on optimized checkpoint

Results

Below are the [email protected]:0.95 results of the PyTorch YOLOX model for the MS-COCO2017 dataset:

Model Configuration	FP32 (%)	INT8 (%)
YOLOX-s	40.5	39.7
YOLOX-l	49.7	48.8