IDRND Anti-spoofing Challenge Solution

NN pipeline

# prepare data & train 5 folds
python run_nn.py prepare-folds \
	--in-dir=./data/train \
	--out-csv=./data/train/dataset.csv \
	--holdout-csv=./data/train/holdout.csv \
	--n-folds=5 \
	--holdout-size=0.2
bash run.sh --dataset train \
	--model resnet18 \
	--n-epochs 30 \
	--batch-size 256 \
	--n-workers 4 \
	--fast
python run_nn.py distil-model \
	--model=resnet18 \
	--in-weights=./models/easy_gold.pth \
	--out-weights=./models/easy_gold.pth

# infer
python run_nn.py infer \
	--in-csv=$PATH_INPUT/meta.csv \
	--in-dir=$PATH_INPUT \
	--out-csv=$PATH_OUTPUT/solution.csv \
	--model=resnet18 \
	--weights-path=./models/easy_gold.pth \
	--batch-size=256 \
	--n-workers=4

LBP pipeline

# prepare data & train model
python run_lbp.py prepare-cutout-datasets \
	--in-dir=./data/train \
	--out-dir-crops=./data/crops \
	--out-dir-cutout=./data/cutout \
	--verbose=True
python run_lbp.py prepare-lbp-dataset \
	--dirpath=./data/crops \
	--features-npy=./data/crops/features.npy \
	--targets-csv=./data/crops/dataset.csv \
	--verbose=True
python run_lbp.py train \
	--features-npy=./data/crops/features.npy \
	--targets-csv=./data/crops/dataset.csv \
	--n-splits=5 \
	--n-repeats=10 \
	--logdir=./logs/lbp

# infer
python run_lbp.py infer \
	--in-csv=$PATH_INPUT/meta.csv \
	--in-dir=$PATH_INPUT \
	--out-csv=$PATH_OUTPUT/solution.csv \
	--weights-path=./logs/lbp/model.pkl

Copyright and License

© Copyright 2019-present Yauheni Kachan. Licensed under the MIT License.