Involution

This repository contains re-Implementation of the paper https://ieeexplore.ieee.org/document/9577788

Involution Kernal

The authors of Involution: Inverting the Inherence of Convolution for Visual Recognition propose a novel involutional layers, which aims to enhance the representation power of convolutional neural networks by inverting the inherent properties of convolution operations. As such these kernals are channel agnostic and spatial specific.

Setup

pip install torch torchvision
pip install wandb
pip install lightning

Folders

models folder contains the main backbone implementations of models used as well as classification heads and lightning class for easy training and logging

slides contains presentation slides with results on Caltech-256

data contains the data module and custom dataset

Training

git clone https://github.com/thatblueboy/involution.git #clone the repo

Following model and training parameters can be configured in train.py by modifying the configs dictionary

Parameters

• model to specify which model you want to train. ResNetClassifier for Resnets and RedNetClassifier for RedNets containing involutions.

• ReDSnet_type to specify depth of the model. Can be one of 26, 38, 50, 101, 152

• batch_size is training batch size

• optimizer and optimizer_kwargs for learing optimizer. optimizer can be Adam or SGD

• num_workers is number of workers

• lr_scheduler for learing rate scheduler. One of ExponentialLR, CosineAnnealingLR, LinearLR, StepLR, PolynomialLR. Any changes to the lr_scheduler will require corresponding changes to lr_sceduler_kwargs

Note

We use a random split split on Caltech256. For uniformity we store this split in the data_module.pth and load it for every training run. This behaviour can be changed by setting the 'data_module_path' value in the config dict to None.

Switch from train to test

• To switch from training to testing mode, change the last line in the train.py from

trainer.fit(model, data_module)

to

trainer.test(model, data_module)

After making all the necessary changes:

wandb login
python train.py

Acknowledgements

Code was heavily inspired by the original papers code: https://github.com/d-li14/involution

Original paper can be found here: https://ieeexplore.ieee.org/document/9577788

This project was done as a partial fulfillment of the course CS F425: Deep Learning at BITS-Pilani