jianzhu / meta-pseudo-labels

Unofficial TF2 implementation of "Meta Pseudo Labels" with linked metrics, dataset and models

Meta Pseudo Labels

Unofficial TF2 implementation of "Meta Pseudo Labels" (official Paper and Code).

Results

	CIFAR-10-4K
	w/o finetune	w/ finetune
Paper	96.08%
Implementation	94.70%	94.71%

Usage

Setup Environment

python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
pre-commit install

Run Scripts

export PYTHONPATH=src

# download dataset
dvc pull -r origin

# train & evaluate
python src/main.py --data-dir data/cifar10 --config-name cifar10 --model-dir workdir/training --mpl-epochs 10000 --mpl-batch-size 64 --finetune-epochs 1000 --finetune-batch-size 512

# only evaluate
python src/evaluate.py --data-dir data/cifar10 --config-name cifar10 --saved-model-dir [workdir/training/finetune/model, saved_models/mpl, saved_models/finetune]