PaCKD: Pattern-Clustered Knowledge Distillation for Compressing Memory Access Prediction Models
Cluster options:
- Past Block Addresses ('a')
- Past Block Address Deltas ('d')
- Past IPs ('i')
Models:
- LSTM ('l')
- MLPMixer ('m')
- ResNet ('r')
To run:
- Import conda env inside of
PaCKD.yaml - Change directories inside of
params.yaml
To preprocess:
src/preprocess.py {app} {cluster option} {gpu}
To train and validate teachers:
src/train_tchs.py {app} {cluster option} {model 1} ... {model k} {gpu}src/validate_tchs.py {app} {cluster option} {model 1} ... {model k} {gpu}
To train and validate students:
src/train_stu.py {app} {cluster option} {alpha tch 1} ... {alpha tch k} {stu model} {tch model 1} ... {tch model k} {gpu}