This repository contains the code for the following paper:
A. Iscen, G. Tolias, Y. Avrithis, O. Chum. "Label Propagation for Deep Semi-supervised Learning", CVPR 2019
This code generally follows Mean Teacher Pytorch implementation. Python 3 and FAISS are required.
As in the original Mean Teacher repository, run the following command:
>> ./data-local/bin/prepare_cifar10.sh
Run the similar command for CIFAR-100, taken from the fastswa-semi-sup repository:
>> ./data-local/bin/prepare_cifar100.sh
We took the Mini-Imagenet dataset hosted in this repository and pre-processed it.
Download train.tar.gz and test.tar.gz, and extract them in the following directory:
>> ./data-local/images/miniimagenet/
There are two stages to our method.
Stage 1 consists of training a network with known labels only. Following command can be run to reproduce this experiment:
>> python train_stage1.py --exclude-unlabeled=True --num-labeled=$NOLABELS --gpu-id=$GPUID --label-split=$SPLITID --isMT=False --isL2=True --dataset=$DATASET
where $NOLABELS is the number of labeled points, $GPUID is the GPU to be used (0 by default), $SPLITID is the ID of the split to be used (10-19 for cifar*, 10-12 for miniimagenet), and $DATASET is the name of the dataset (cifar10, cifar100, or miniimagenet).
After the training for Stage 1 is completed, run the following command for Stage 2, which resumes the training from the model trained in Stage 1, but this time with pseudo-labels using the entire dataset:
>> python train_stage2.py --labeled-batch-size=50 --num-labeled=$NOLABELS --gpu-id=$GPUID --label-split=$SPLITID --isMT=False --isL2=True --dataset=$DATASET
Use the following commands for combining Mean Teacher with our method:
Stage 1:
>> python train_stage1.py --labeled-batch-size=50 --num-labeled=$NOLABELS --gpu-id=$GPUID --label-split=$SPLITID --isMT=True --isL2=False --dataset=$DATASET
Stage 2:
>> python train_stage2.py --labeled-batch-size=50 --num-labeled=$NOLABELS --gpu-id=$GPUID --label-split=$SPLITID --isMT=True --isL2=False --dataset=$DATASET