Skip to content
Implementation of paper: Making Deep Neural Network Robust to Label Noise: a Loss Correction Approach.
Branch: master
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Permalink
Type Name Latest commit message Commit time
Failed to load latest commit information.
cfg/experiments Add cfg files. Apr 4, 2018
helper Fix a bug in sampling. May 6, 2018
network Corrected a bug Apr 4, 2018
script Now result folder will be created. Apr 4, 2018
trainer 1. Added some assertion to prevent underlying error. Apr 4, 2018
README.md
__init__.py Initial commit Mar 16, 2018
experiment_mnist.py
results.png Modified code structure, so that new networks will be easier to be in… May 6, 2018
show_result_of_mnist_experiment.py

README.md

label_noise_correction

Implementation of paper: Making Deep Neural Network Robust to Label Noise: a Loss Correction Approach.

Requirements:

  • Python 2.7
  • TensorFlow 1.4
  • Matplotlb
  • Numpy

Usage

  • Train all models and evaluate all the tests with: python experiment_mnist.py, or with bash script/run_experiment_mnist for faster training and testing. When this is finished, 4 files named backward.npy, backward_t.npy, cross_entropy.npy, forward.npy, forward_t.npy should have been created under the path ./result/mnist/.
  • Show the result with: python show_result_of_mnist_experiment.py.

Result

This is the result of Fully connected network on MNIST. Notice that when N=0.5, the parametric matrix T is singular. results.png

You can’t perform that action at this time.