/COMP551_P4

Primary LanguagePython

Acknowledgements

This project is for the fourth project of COMP551 at McGill University in fall 2021. Here we bid thanks to Yuyan Chen, Ing Tian, and Zijun Zhao, without whom this project cannot come real.

Brief

In this project, we investigate the effects of a family of activation functions ACON proposed in this paper. Concretely, we explore the effects of ACON and Meta-ACON with respect to ReLU in various experimental setups. For example, we have tested the performance of ACON, Meta ACON, and ReLU for variants of VGG16 on the CIFAR-100 datset.

Project Structure

.
├── README.md
├── acon.py
├── classifier
│   ├── __init__.py
│   ├── metric.py
│   ├── network
│   │   ├── __init__.py
│   │   ├── alex_net.py
│   │   ├── resnet_acon.py
│   │   ├── resnet_metaacon.py
│   │   ├── resnet_relu.py
│   │   ├── shuffle.py
│   │   ├── shuffle_acon.py
│   │   ├── shuffle_metaacon.py
│   │   ├── vgg16_6_acon.py
│   │   ├── vgg16_6_metaacon.py
│   │   ├── vgg16_6_relu.py
│   │   ├── vgg16_acon.py
│   │   ├── vgg16_metaacon.py
│   │   └── vgg16_relu.py
│   └── plugin.py
├── data
│   └── __init__.py
├── main.py
└── p4.ipynb

acon.py contains the activation functions ACON and MetaACON we wish to investigate in this experiment. Inside the classifier folder, we have defined various models spanning from VGG, AlexNet, ShuffleNet, and ResNet. Also, some common utils pertaining to these models are defined in classifier/__init__.py, classifier/metric.py, and classifier/plugin.py. Next, the data folder contains utils related to dataset processing. Finally. we have provided a sample p4.ipynb to run our codes in Colab.

How to run codes in Colab?

Though Colab is convenient, we suffer from frequent disconnections. Hence, we often run our script locally via main.py . For your convenience, we have provided a sample ipynb file such that you can replicate our results in Colab. The procedure is simple.

  1. Name the project folder as COMP551_P4.
  2. Zip the project folder into COMP551_P4.zip.
  3. Open the p4.ipynb in Colab and connect to a GPU runtime.
  4. Upload COMP551_P4.zip to the Colab runtime under the /content folder.
  5. Run the Jupyter notebook.