ResNet-50 with CBAM using PyTorch 1.8

Introduction

This repository contains the implementation of ResNet-50 with and without CBAM. Note that some parameters of the architecture may vary such as the kernel size or strides of convolutional layers. The implementation was tested on Intel's Image Classification dataset that can be found here.

Pretrained Weights

The trained models' weights are provided here.

How to Use

In order to train your own custom dataset with this implementation, place your dataset folder at the root directory. Make sure that your dataset is split into two subfolder. train and test where the former contains your training dataset and the latter contains your validation set. Refer to the folder named intel_dataset in the repo as an example.

If you wish to train the model without CBAM, you can do so with

python train.py --data_folder [YOUR DATASET FOLDER NAME] --gpus [NUMBER OF GPU]

To train the model with CBAM, run

python train.py --data_folder [YOUR DATASET FOLDER NAME] --gpus [NUMBER OF GPU] --use_cbam

There are more arguments that can be supplied to the command. Run

python train.py -h

for more information.

If you wish to visualize the final layer of feature maps produced by the trained model, create a folder in the root directory and simply place your images inside it and run

python visualize.py --data_folder [FOLDER NAME] 

Additionally, if the model was trained on CBAM architecture, then add --use_cbam at the end of the command above.

Performance

ResNet-50 with CBAM achieved an accuracy of 86.6% on the validation set while ResNet-50 without CBAM achieved an accuracy of 84.34% on the same validation set. The figures below show the improvement of the models over the epochs.

Visualization

The figures below are the feature maps of the final convolutional layer of both ResNet-50 without CBAM and ResNet-50 with CBAM.