/curriculum_learning

Code implementing the experiments described in the paper "Statistical Measures For Defining Curriculum Scoring Function" by Sadasivan & Dasgupta (SubsetML Workshop @ ICML 2021)

Primary LanguagePython

Code references are provided in README files.

The codes run on Python 3.7.6 (Ubuntu 20.04)

Requires:
TensorFlow 1.15.0
PyTorch 1.5.1
Numpy 1.18.1
Skimage
autograd_hacks.py (External library for gradient computation. Reference: https://github.com/cybertronai)
CUDA 10.0 and GPU support.

The outputs are saved as pickle/numpy files in the directories as mentioned in each README.

Thanks to "On The Power of Curriculum Learning in Training Deep Networks" by Hacohen & Weinshall (ICML 2019).