README.md

Code for "A novel adaptive learning rate scheduler for deep neural networks"

This repository contains the code for the paper

Yedida, Rahul, and Snehanshu Saha. "A novel adaptive learning rate scheduler for deep neural networks." arXiv preprint arXiv:1902.07399 (2019).

The main function is lr_schedule, that computes the adaptive learning rate for a given dataset and neural network. This method can be used with Keras' LRScheduler callback, as used in all the code files.

All results and code from the paper can be found here. Trained models and program outputs are also uploaded here. The top-level directories in this repository correspond to the dataset the experiments were run on. The paper directory contains the compiled PDF. The code uses the Keras deep learning library.

Directory Structure

At the cost of duplicate code, every directory has independently executable code, i.e., to run any experiment, only the files in that directory are required.

The Unconstrained or Nonconstrained directories under the DenseNet architecture are experiments that we re-did to make the comparison more fair. The only difference is that nb_filter is set to -1 in these experiments, as with all the other DenseNet experiments.

The directories are organized by dataset, algorithm, and other options, such as weight decay value, adaptive/baseline, etc. The full directory structure is below (output from tree -d)

.
├── CIFAR10
│   ├── Adam
│   │   ├── adam-resnet-1e3
│   │   │   ├── Adaptive
│   │   │   └── Baseline
│   │   └── densenet-1e4
│   │       ├── Adaptive
│   │       └── Baseline
│   ├── Momentum
│   │   ├── DenseNet
│   │   │   ├── Adaptive
│   │   │   └── Baseline
│   │   └── ResNet
│   │       ├── Adaptive
│   │       │   ├── 1e-2
│   │       │   └── 1e-3
│   │       └── Baseline
│   ├── RMSprop
│   │   ├── DenseNet
│   │   │   ├── Adaptive
│   │   │   └── Baseline
│   │   └── ResNet
│   │       ├── Adaptive
│   │       └── Baseline
│   └── SGD
│       ├── DenseNet
│       │   ├── Adaptive
│       │   └── Baseline
│       └── ResNet20
│           ├── Adaptive
│           │   ├── wd1e3-adaptive
│           │   └── wd1e3-fixed
│           └── Baseline
├── CIFAR100
│   ├── Adam
│   │   ├── DenseNet
│   │   │   ├── Adaptive
│   │   │   │   ├── Constrained filter size
│   │   │   │   │   ├── 1e4
│   │   │   │   │   └── 1e5
│   │   │   │   └── Unconstrained
│   │   │   └── Baseline
│   │   └── ResNet164
│   │       ├── Adaptive
│   │       └── Baseline
│   ├── Momentum
│   │   ├── DenseNet
│   │   │   ├── Adaptive
│   │   │   └── Baseline
│   │   └── ResNet164
│   │       ├── Adaptive
│   │       └── Baseline
│   ├── RMSprop
│   │   ├── DenseNet
│   │   │   ├── Adaptive
│   │   │   └── Baseline
│   │   └── ResNet
│   │       ├── Adaptive
│   │       └── Baseline
│   └── SGD
│       ├── DenseNet
│       │   ├── Adaptive
│       │   │   ├── Constrained Filters
│       │   │   └── Nonconstrained
│       │   └── Baseline
│       ├── ResNet162v2
│       │   ├── Adaptive
│       │   └── Baseline
│       └── ResNet56v2
├── MNIST
└── paper

Setup

The repository uses standard deep learning libraries in Python:

• keras
• sklearn
• numpy
• matplotlib
• pickle
• tqdm

Credits

This repository uses code from the following sources:

• ResNet implementation from Keras docs
• DenseNet implementation by titu1994
• AdaMo implementation based on Keras Adam implementation. Note that this URL is based on commit a139716.
• MNIST architecture from Kaggle Python notebook by Aditya Soni

Citation

If you find this work useful, please cite the paper:

@article{yedida2019novel,
  title={A novel adaptive learning rate scheduler for deep neural networks},
  author={Yedida, Rahul and Saha, Snehanshu},
  journal={arXiv preprint arXiv:1902.07399},
  year={2019}
}