chingisooinar/self-supervised-skku

Final Project for Deep Learning Course at Sungkyunkwan University (2021 Spring). Placed 1st at Kaggle's in-class competition (0.95957 accuracy).

Hyperparameters

Parameter	Meaning
low-dim	categories
lr	Learning rate
b	Batch size
model	Network name
Instance	Run title
fixmatch	Whether to use Fixmatch
threshold	Threshold for pseudo labels
T	temperature
r	Checkpoint name
test-only	Whether to run inference only
uda	Whether to use UDA
ensemble	Whether to run Ensemble inference

Competition results: https://www.kaggle.com/c/dnn2021ssl/leaderboard

Example for training using FixMatch

train.py --low-dim 10 --lr 1e-4 --b 64 --model vgg19 --instance NAME --fixmatch True --threshold 0.95 --T 0.5

Example for training using UDA

train.py --low-dim 10 --lr 1e-4 --b 64 --model vgg19 --NAME --uda True

Example for training Supervised only

train.py --low-dim 10 --lr 1e-4 --b 64 --model vgg19 --NAME

Example for inference

train.py --low-dim 10 --lr 1e-4 --b 64 --model vgg16 --instance NAME -r NAME.t7 --test-only

Example for Ensemble inference

python3 train.py --low-dim 10 --lr 1e-4 --b 64 --model vgg19,vgg16 --instance ensemble -r NAME.t7,NAME2.t7 --test-only --ensemble True