DeepGuiser Tutorial

General Flow

NAT-D Flow

Pretrain supernet
Train transformer by training integrated model

Predictor Flow

Pretrain supernet
Build Transbench
Train Predictor on Transbench
Train transformer with predictor

Detailed Tutorial

pretrain_supernet

cd supernet/

To train one supernet, run

python train_supernet.py --gpu [gpuid] --seed [seed_num]

After training complete, copy supernet and twin supernet to ${workspaceFolder}/supernet/selected_supernet and rename them to supernet.pt and supernet_twin.pt

train gates integrated model

cd integrated_models/nat_disguiser

And run,

python net_disguiser.py --gpu [gpuid]

build transbench

cd transbench

If you want to build transbench yourself,

cd high_fidelity

And run,

python build_high_fidelity_transbench.py

The already build transbench written in yaml is in data/high_fidelity

train predictor

cd predictors

If you want to train high fidelity predictor,

cd high_fidelity

python finetune_predictor.py

The config file of predictor and the config file of optimization detail is in the two yamls.

train predictor based tansformer

The code is in

integrated_models/predictor_based

final test