Please install and setup AIMET before proceeding further.
This model was tested with the torch_gpu variant of AIMET 1.25.
python -m pip install gdownPlease follow the steps from open-mmlab/mmaction2 install guide to install mmaction2 as dependency. The package versions we used for open-mmlab are:
- mmaction2 1.0.0
- mmengine 0.7.3
- mmcv 2.0.0
Append the repo location to your PYTHONPATH with the following:
export PYTHONPATH=$PYTHONPATH:<path to aimet-model-zoo>Instructions to prepare ActivityNet can be found at:
- https://github.com/open-mmlab/mmaction2/blob/main/tools/data/activitynet/README.md Note that option 1 was used for this model
After downloading and processing the dataset, please change the data path to point to your download location in aimet_zoo_torch/mmaction2/model/configs/localization/bmn/bmn_2xb8-400x100-9e_activitynet-feature.py
Before running the evaluation script, set your config path in the model cards via replacing with your own path in the "config" field. The model cards are .json files located under model/model_cards/
To run evaluation with QuantSim in AIMET, use the following
python aimet_zoo_torch/mmaction2/evaluators/mmaction2_quanteval.py --model-config <configuration to be tested> --use-cudaAvailable model configurations are:
- bmn_w8a8
- Weight quantization: 8 bits, per tensor symmetric quantization
- Bias parameters are not quantized
- Activation quantization: 8 bits, asymmetric quantization
- Model inputs are quantized
- TF enhanced was used as quantization scheme