sud0301 / my_uda_pytorch

Pytorch Implementation of Unsupervised Data Augmentation for Consistency Training

python main.py \
    --dataset CIFAR10 \
    --num-labeled  1000 \
    --lr-warm-up  \
    --warm-up-steps 10000 \
    --num-steps 100000 \
    --batch-size-lab 64 \
    --batch-size-unlab 320 \
    --confidence-mask 0.6 \
    --softmax-temp 0.5 \
    --seed 3 \
    --rot \
    --verbose \
    --use-ema 

python main.py \
    --dataset ImageNet \
    --num-classes 1000 \
    --percent-labeled  10 \
    --lr-warm-up  \
    --warm-up-steps 10000 \
    --num-steps 200000 \
    --batch-size-lab 128 \
    --batch-size-unlab 128 \
    --lr 0.3 \
    --wdecay 1e-4 \
    --confidence-mask 0.5 \
    --softmax-temp 0.4 \
    --gpu-id 0,1,2,3,4,5,6,7 \
    --seed 3 \
    --rot \
    --verbose \
    --use-ema 

Use rot flag for using rotation loss from S4L paper. Run for 400k iterations(num-steps) for improved results.

Results on CIFAR-10

# Labels	UDA Paper	This UDA Repo	with Rotation(rot)
250	8.76	11	8.2
500	6.68	9.1	6.4
1000	5.87	7.8	5.8
2000	5.51	6.9	5.7
4000	5.29	6.1	5.1

Model does not learn with higher learning rate lr=0.3