Introduction

The source code and models for our paper PNP: Robust Learning from Noisy Labels by Probabilistic Noise Prediction

Framework

Installation

After creating a virtual environment of python 3.6, run pip install -r requirements.txt to install all dependencies

How to use

The code is currently tested only on GPU.

• Data preparation

  Created a folder Datasets and download cifar100/web-aircraft/web-bird/web-car/food101n dataset into this folder.

• Source code

  • If you want to train the whole model from beginning using the source code, please follow subsequent steps:
    • Prepare data
    • Modify GPU device in the corresponding train script xxx.sh in scripts folder
    • Activate virtual environment (e.g. conda) and then run

    bash scripts/xxx.sh
    

• Demo

  • If you just want to do a quick test on the model, please follow subsequent steps:
    • Prepare data
    • Download one of the following trained models

      wget https://web-pnp.oss-cn-shanghai.aliyuncs.com/pnp_hard-f101n-r50-87.3109.pth
      wget https://web-pnp.oss-cn-shanghai.aliyuncs.com/pnp_soft-f101n-r50-87.5010.pth
      wget https://web-pnp.oss-cn-shanghai.aliyuncs.com/pnp_hard-web_car-r50_89.9266.pth
      wget https://web-pnp.oss-cn-shanghai.aliyuncs.com/pnp_soft-web-car_r50_90.1132.pth
      

    • Modify GPU, MODEL, DATASET, and NCLASSES accordingly in the demo script demo.sh in scripts folder
    • Activate virtual environment (e.g. conda) and then run

      bash scripts/demo.sh