Awesome

T2T

Code for the paper: "Trash to Treasure: Harvesting OOD Data with Cross-Modal Matching for Open-Set Semi-Supervised Learning" by Junkai Huang, Chaowei Fang, Weikai Chen, Zhenhua Chai, Xiaolin Wei, Pengxu Wei, Liang Lin, and Guanbin Li.

Usage

Dataset Preparation

Download CIFAR-10 dataset and place cifar-10-batches-py under the directory data.

Train

Train the model by 1000 labeled data of CIFAR-10 dataset and use TinyImageNet as OOD data source:

# stage 1
python train_t2t_stage1.py \
--dataset cifar10 \
--ood-dataset TIN \
--num-labeled 1000 \
--arch wideresnet \
--batch-size 64 \
--expand-labels \
--seed 5 \
--total-steps 50000 \
--eval-step 500 \
--out results/cifar10_TIN@1000.5/stage1

# stage 2
python train_t2t_stage2.py \
--dataset cifar10 \
--ood-dataset TIN \
--num-labeled 1000 \
--arch wideresnet \
--batch-size 64 \
--mu 5 \
--expand-labels \
--seed 5 \
--total-steps 200000 \
--eval-step 1000 \
--resume results/cifar10_TIN@1000.5/stage1/checkpoint.pth.tar \
--out results/cifar10_TIN@1000.5/stage2

Use UDA as baseline:

python train_uda.py \
--dataset cifar10 \
--ood-dataset TIN \
--num-labeled 1000 \
--arch wideresnet \
--batch-size 64 \
--mu 5 \
--expand-labels \
--seed 5 \
--total-steps 250000 \
--eval-step 1000 \
--out results/cifar10_TIN@1000.5/uda