This is the code for: Tackling Noisy Labels with Noise-Free Prototype Guided Representation Calibration (PyTorch implementation).
-
Ubuntu 22.04
-
Python 3.9
-
PyTorch, verion=2.0.0
-
CUDA, version=11.8
We verify the effectiveness of the proposed method on synthetic noisy datasets. In this repository, we provide the used datasets (the images and labels have been processed to .npy format). You should put the datasets in the folder “data” when you have downloaded them. Training example:
python main.py --dataset mnist --noise_type symmetric --noise_rate 0.4 --Lambda1 0.03 --Lambda2 0.01 --seed 1
Our pre-trained models by unsupervised contrastive learning for computing prototypes can be downloaded as following: Models
We referred to the official code of MoCo.