Official code for paper Meta Soft Label Generation for Noisy Labels accepted by ICPR 2020.

Illustration of the proposed MSLG algorithm

Requirements:

• torch
• torchvision
• scikit-learn
• matplotlib

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 -g gamma -s1 stage1 -s2 stage2 -k K -m metadata_num

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
• gamma: float gamma value
• stage1: integer epoch value for stage1
• stage2: integer epoch value for stage1
• K: integer K multiplier for label initialization
• metadata_num: number of meta-data

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

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