Learning to Generate Soft-Labels from Noisy Labels
Official code for paper MetaLabelNet: Learning to Generate Soft-Labels from Noisy-Labels.
Illustration of the proposed MetaLabelNet algorithm
Requirements:
- torch
- torchvision
- scikit-learn
- matplotlib
Running Proposed Algorithm
Code can be run as follows:
python main.py -d dataset_name -n noise_type -r noise_ratio -s batch_size -a alpha -b beta -s1 stage1 -s2 stage2 -m meta_data_num -u unlabeled_data_num -v verbose
where options for input arguments are as follows
- dataset_name: cifar10, clothing1M, food101N
- noise_type: feature-dependent, symmetric (valid only for cifar10 dataset for synthetic noise)
- noise_ratio: integer value between 0-100 representing noise percentage (valid only for cifar10 dataset for synthetic noise)
- batch_size: any integer value
- alpha: float alpha value
- beta: float beta value
- stage1: integer epoch value for stage1
- stage2: integer epoch value for stage1
- meta_data_num: number of meta-data
- unlabeled_data_num: number of unlabeled-data
- verbose: integer value of: 0 (silent), 1(print at each epoch), 2(print at each batch)
Any of the input parameters can be skipped to use the default value. For example, to run with default values for all parameters:
python main.py -d clothing1M
Running Baseline Methods
Baseline methods can be run as follows:
python baselines.py -d dataset_name -n noise_type -r noise_ratio -m model_name
where baseline model can be one of the followings:
- model_name: cross_entropy, symmetric_crossentropy, generalized_crossentropy, bootstrap_soft, forwardloss, joint_optimization, pencil, coteaching, mwnet, mlnt