smartdolphin / semi-self-supervised

Semi-Self-Supervised Learning

semi-self-supervised

This is an official PyTorch implementation of Wafer BIN Map Defect Image Classification with Semi-Self-Supervised Learning.

This code is only available in Semi-Self Supervised Learning. Now only experiments on WM-811K is available.

Requirements

• Python 3.6+
• PyTorch 1.4.0
• torchvision 0.5
• tensorboard
• tqdm
• numpy

Usage

Train

Train the model by 225, 450, 900 labeled data of WM-811K dataset:

python train.py --dataset wm811k --num-labeled 900 --arch wideresnet --batch-size 64 --lr 0.03 --out wm-811k@900 (TODO)

Train the model by 900 labeled data of Wm-811K dataset by using DistributedDataParallel:

python -m torch.distributed.launch --nproc_per_node 4 ./train.py --dataset wm811k --num-labeled 900 --arch wideresnet --batch-size 16 --lr 0.03 --out wm811k@900

Monitoring training progress

tensorboard --logdir=<your out_dir>

Results (Accuracy)

WM-811K

#Labels	225	450	900
Supervised	23.51 ± 2.36	19.39 ± 4.66	10.26 ± 2.98
Pseudo-Label	7.23 ± 1.35	5.98 ± 0.21	4.94 ± 0.07
PI-Model	7.42 ± 0.99	6.04 ± 0.41	5.08 ± 0.02
Mean-Teacher	7.64 ± 0.86	7.49 ± 1.19	5.13 ± 0.18
VAT	8.43 ± 1.1	5.97 ± 0.3	8.3 ± 0.53
VAT + EntMin.	8.6 ± 1.98	6.1 ± 0.31	5.98 ± 0.73
FixMatch (RA)	13.14 ± 1.87	8.99 ± 1	5.49 ± 0.72
Paper (our)	6.03 ± 0.14	5.35 ± 0.17	4.66 ± 0.13

CIFAR10

#Labels	40	250	4000
Supervised	-	-	-
Pseudo-Label	-	-	-
PI-Model	-	-	-
Mean-Teacher	-	-	-
VAT	-	-	-
VAT + EntMin.	-	-	-
FixMatch (RA)	86.19 ± 3.37	94.93 ± 0.65	95.74 ± 0.05
Paper (our)

* Results of this code were evaluated on 1 run.