We implemented quantization methods for visual transformers and analyzed tradeoffs between model performance and accuracy (Visage Technologies internship)
To reproduce results, consult the scripts/ directory.
The general workflow is:
- Finetune the models (
start_finetune.sh small finetune-smalletc.) - Run the experiments (
start_experiments.sh)