Activation functions for CNN

Activation functions for convolutional neural networks: proposals and experimental study

Algorithms included

This repo contains the code to run experiments with different activation functions that have been used recently for convolutional network models. These are:

• ReLU
• LReLU
• RTReLU
• RTPReLU
• PairedReLU
• EReLU
• EPReLU
• SQRT
• RReLU
• ELU
• SlopedReLU
• PELU
• PTELU
• MPELU
• s₊
• s₊₊
• s₊₂
• s₊₂L
• ELUs₊₂
• ELUs₊₂L

Installation

Dependencies

This repo basically requires:

• Python (>= 3.6.8)
• click (>=6.7)
• h5py (>=2.9.0)
• Keras (==2.2.4)
• matplotlib (>=3.1.1)
• numpy (>=1.17.2)
• opencv-python (>=4.1.2)
• pandas (>=0.23.4)
• Pillow (>=5.2.0)
• prettytable (>=0.7.2)
• scikit-image (>=0.15.0)
• scikit-learn (>=0.21.3)
• tensorflow (==1.13.1)

Compilation

To install the requirements, use:

Install for CPU pip install -r requirements.txt

Install for GPU pip install -r requirements_gpu.txt

Development

Contributions are welcome. Pull requests are encouraged to be formatted according to PEP8, e.g., using yapf.

Usage

You can run all the experiments with CIFAR-10, CIFAR-100, CINIC-10, MNIST and Fashion-MNIST by running the following lines:

python main_experiment.py experiment -f exp/activations_cifar10.json
python main_experiment.py experiment -f exp/activations_cifar100.json
python main_experiment.py experiment -f exp/activations_cinic10.json
python main_experiment.py experiment -f exp/activations_mnist.json
python main_experiment.py experiment -f exp/activations_fashion.json

Note that the CINIC dataset must be stored under ../datasets/CINIC.

The .json files contain all the details about the experiments settings.

After running the experiments, you can use tools.py to watch the results:

python tools.py

Citation

The paper titled "Activation functions for convolutional neural networks: proposals and experimental study" has been submitted to IEEE Transactions on Neural Networks and Learning Systems (IEEE TNNLS).

Contributors

Activation functions for convolutional neural networks: proposals and experimental study

• Víctor Manuel Vargas (@victormvy)
• Pedro Antonio Gutiérrez (@pagutierrez)
• César Hervás-Martínez ([email protected])