This repo is the result of the work done regarding the study of the adaptive gradient clipping (AGC) on the ResNet architectures It's made out of four folders.
- The folder 'comparison_study' contains the networks we used to make a comparison between the AGC, a normalizer-free, and a vanilla ResNet50. Those nets were trained on the CIFAR-10 dataset.
- Then, the folder 'different_netwrok_size' holds the file of different sizes of ResNet architecture implemented with the AGC. Those nets were trained with the CIFAR-10 dataset.
- Following is the folder 'plantclef_nets' contains the same nets as the 'comparison_study' one but this time the nets are trained on the PlantCLEF 2013 dataset.
- Finally, the last folder 'plantclef_preprocess' contains the scripts to transform the data from PlantCLEF 2013 into a usable dataset.
This work is based on the paper High-Performance Large-Scale Image Recognition Without Normalization and was implemented using Keras thanks to the work of P. Sayak [ 1 ] [ 2 ].
[ 1 ] A. Brock, S. De, S. Smith, K. Simonyan (2021). High-Performance Large-Scale Image Recognition Without Normalization.
[ 2 ] P. Sayak, Adaptive Gradient Clipping Repo:https://github.com/sayakpaul/Adaptive-Gradient-Clipping