In this work, I replace standard convolutions with fixed separable convolutions in several popular neural networks. I study how these modifications influence performance of CNNs. Several different domains and networks are tried here.
- Image Classification with ResNets
- Image Generation with DCGAN
Note: Some of the code used here is taken from open source repositories or is influenced by work of other people. Complete list of references and credits will be added later. The main sources are:
- https://github.com/akamaster/pytorch_resnet_cifar10
- https://github.com/hysts/pytorch_image_classification
- https://github.com/pytorch/examples/tree/master/dcgan
The code is written in python and uses pytorch.
Dependencies can be installed via:
conda env create -f environment.yml
All files are provided under the terms of the Apache License, Version 2.0.