MLP-NAS

• Clone using:

git clone --recurse-submodules https://github.com/TomasGadea/MLP-NAS.git
cd MLP-NAS

• Create python3.9.13 environ (python3 version of scar):

python3 -m venv environ
source environ/bin/activate
pip install -r requirements.txt

• Ask @TomasGadea for config files: config.json and add your wandb API key (this last thing is optional)

Train Search

(Open a tmux session rooted in MLP-NAS)

sh execute.sh

To see all available params check main.py.

• --use-amp stores True when added and uses torch.cuda.amp.autocast and torch.cuda.amp.GradScaler.
• --wandb stores True and logs into your wandb account using your API (optional).

Retrain Fixed architecture found with Train Search

(Open a tmux session rooted in MLP-NAS)

sh fixed_execute.sh

To see all available params check fixed_main.py.

• --path-to-supernet is the output path of any past experiment of execute.sh. Check the example in fixed_execute.sh.

Output Files

Output files for Train Search are:

• flops_table.txt: string formatted table of n_params and flops of model.
• log.csv: metrics such as acc, F, mmc, along epochs.
• params.json: parameters that include all arguments called in execute.sh and other extra info.
• W.pt: Last version of the model saved in PyTorch format after all training epochs.
• W_test.pt: Best version of the model saved in PyTorch format after all training epochs.

Output files for Retrain Fixed are:

• flops_table.txt: string formatted table of n_params and flops of model.
• log.csv: metrics such as acc, mmc, along epochs.
• params.json: parameters that include all arguments called in fixed_execute.sh and other extra info.
• W.pt: Last version of the model saved in PyTorch format after all training epochs.

Retrain Fixed files are stored in out/retrain/ dir, unlike Train Search that are directly into out/ directoy. They can be modified however using the --output arg if desired.