Awesome

[NeurIPS-23] Class-Distribution-Aware Pseudo-Labeling for Semi-Supervised Multi-Label Learning

The implementation for the paper Class-Distribution-Aware Pseudo-Labeling for Semi-Supervised Multi-Label Learning (NeurIPS 2023).

See much more related works in Awesome Weakly Supervised Multi-Label Learning!

Preparing Data

See the README.md file in the data directory for instructions on downloading and preparing the datasets.

Training Model

To train and evaluate a model, the next two steps are required:

1. Firstly, we warm up the model with the labeled data. Run:

python run_warmup.py --loss_lb asl --lb_ratio 0.05 \
--warmup_epochs 12 --lr 1e-4 --net resnet50 \
--dataset_name coco --dataset_dir ./data

2. Secondly, we train the model with CAP method. Run:

python run_CAP.py --loss_lb asl --loss_ub asl --lb_ratio 0.05 \
--warmup_epochs 12 --warmup_batch_size 16 --lr 1e-4 --net resnet50 \
--dataset_name coco --dataset_dir ./data \
--init_pos_per 1.0 --init_neg_per 1.0

Training Logs

Download warmup models.

Hyper-Parameters

To generate different entries of the main table, modify the following parameters:

1. dataset_name: The dataset to use, e.g. 'coco', 'voc', 'nus'.
2. dataset_dir: The directory of all datasets.
3. batch_size: The batch size of samples (images).
4. net: The structure of model which is used to train, e.g. 'resnet50', 'mlder'(ML-Decoder with resnet50).
5. lb_ratio: The result of (the size of labeled samples) : (the size of all samples).
6. loss_lb: The loss used for labeled samples.
7. loss_ub: The loss used for unlabeled samples.