Requirements

conda create -n introdl python=3.9
conda activate introdl
- if mac user:
    conda install pytorch torchvision torchaudio -c pytorch
- else:
    conda install pytorch torchvision torchaudio pytorch-cuda=11.3 -c pytorch -c nvidia
pip install -r requirements.txt

To download fMRI datasets, run

PYTHONPATH=./ python scripts/download_datasets.py

To generate synthetics datasets, run

PYTHONPATH=./ python scripts/generate_datasets.py

Examples

Training

for model in lstm mean_lstm transformer mean_transformer dice
do
    for dataset in cobre abide synth1 synth2
    do
        PYTHONPATH=./ python scripts/run_experiments.py --mode tune --model $model --ds $dataset --prefix test
        PYTHONPATH=./ python scripts/run_experiments.py --mode exp --model $model --ds $dataset --prefix test
    done
done

PYTHONPATH=./ python scripts/run_experiments.py --mode tune --model lstm --ds abide --prefix test
PYTHONPATH=./ python scripts/run_experiments.py --mode exp --model lstm --ds abide --prefix test

Introspection

For model introspection to work you need to move (or copy) the exp folder of the desired model from the assets/logs to assets/trained_models. Choose the same --prefix as the prefix of your desired model's folder. We uploaded some tained weights; use introdl prefix for them.

for model in lstm mean_lstm transformer mean_transformer dice
do
    for dataset in cobre abide synth1 synth2
    do
        PYTHONPATH=./ python scripts/run_introspection.py --model $model --ds $dataset --methods saliency --prefix introdl
    done
done

PYTHONPATH=./ python scripts/run_introspection.py --model lstm --ds abide --methods saliency --prefix introdl

Options for `scripts/run_experiments.py`

Required

--mode:
- tune - tune mode: run multiple experiments with different hyperparams
- exp - experiment mode: run experiments with best hyperparams found in the tune mode
- resume - see below
--model: some of the working models; check the sourse code for more info
- lstm
- mean_lstm
- transformer
- mean_transformer
- dice
--ds: dataset for the experiments
- cobre
- abide
- synth1
- synth2

Optional

--prefix: custom prefix for the project
- default prefix is UTC time
- appears in the name of logs directory and the name of WandB project
- tune->exp experiments will use the same prefix (unless it is default)
- don't use - character in the prefix
- don't use resume or tune as a prefix
--multiclass: some datasets have multiple classes (default: False); pass --multiclass if you want to load all classes
--zscore: whether dataset should be zscored over time direction (default: False); pass --zscore if you want zscore the data
--filter-indices: whehter ICA components in real-world fMRI data should be filtered (default: True); pass --no-filter-indices if you want to load all ICA components
--max-epochs - max epochs to use (default: 200)
--n-splits - number of splits for StratifiedKFold cross-validation (default=5):
- the ds dataset is split in num-splits equally sized folds
--n-trials - number of trials for each test fold (default=50 for tune and 10 for exp):
- in tune it equals to number of different hyperparams sets tested on each fold
- in exp mode, for each trial, a new seed for train_test_split is used for splitting train-val dataset into train and val datasets
- important note: if you provide the same num-splits and num-trials for different experiments on the same dataset, datasets splits will be the same
--batch-size - batch size (default: 64)
--patience - patience for early stopping (default: 30)

Required for `resume` mode

--mode:
- resume - resume mode: for resuming interrupted experiment
--path:
- path to the interrupted experiment (e.g., /Users/user/intro-dl-project/assets/logs/prefix-mode-model-ds)

paavalipopov/intro-dl-project