Finetuning and Evaluating Vision Transformers and ResNets for Image Classification

Validating image classification benchmark results on ViTs, DeiT and ResNets BiT models

Hello everyone! This repository contains the code implementation for our ML701 project "Transformers Transforming Vision", in which we have revisted the baseline vision transformers such as ViTs, DeiT and compared them with the CNN models like ResNet BiT.

We have tried to validate the official published results of ViTs, DeiTs and ResNets on CIFAR-10 and CIFAR-100, when they are pretrained on different datasets like ImageNet and ImageNet-21k. Additionally, we have used CUB200 dataset and exploited the importance of having high resolution images specifically for vision transformers. Our approch is summarized below:

Comprehensive report for this project can be found here. Specifically, we provide finetuning and evalutation scripts for all ResNet-BiT, ViT and DeiT models which are supported by PyTorch library.

Requirements

To run the scripts, following packages needs to be installed (preferably on Ubuntu 18.04 LTS / 20.04 LTS):

Python (version 3.6 or greater)
Pytorch (version 1.10) and Torchvision (version 0.3.0)
Pytorch timm library (preferably version 0.4.12)

To install these, use pip package installer to install those one by one, or you can anaconda environment (highly recommended) using the provided env.yml file.

TO set-up environment using the anaconda environment please follow and execute the following commands in terminal:

Create new conda environment using the env.yml file (provided in repository).

$ conda env create --file env.yml

Now all required packages are installed in this environment, it can now be activated as follows:.

$ conda activate visiontimm

Command Line Parameters

The below table lists the command line parameters available for main.py which is used for both evaluation as well as finetuning of models.

`main.py`

Parameter	Description	Possible Values
--batch-size	batch size of dataloader	int: any integer value, e.g. 16, 24, etc.
--num-epochs	number of training epochs.	int: (any integer value)
--model-name	name of model to evaluate/finetune	string: ViTs, DeiTs and ResNets models supported in timm, eg: 'deit_base_patch16_224'
--img-size	resolution of input images	int: eg: 224
--optimizer	optimizer for training	string: possible values are 'ADAM' and 'SGD'
--lr	learning rate during finetuning	float: any float value, e.g 0.01
--training	finetune from checkpoint	boolean: (default: True), put False for evaluation
--weights-path	path where you want to save model weights after training.	string: (default: './saved_models')
--load-weights	path from where trained weights are to be loaded (only required for evaluating a pretrained model.	string: path of the model weights eg. path/to/my/model/weights
--figure-path	path for saving plots.	string: (default: './figures')
--data-path'	path where dataset is present.	string: (default: './data')
--dataset-name	Choice of dataset	string: can choose from 'CIFAR10', 'CIFAR100', 'CUB200'
--val-pct	validation split from the training set.	float: (default: 0.1)

NOTE: To use CUB200 dataset in the experiments, please manually download the CUB_200_2011.tgz file and paste it into the './data' directory.

Finetuning models

To finetune a timm model, we can run the main.py as follows:

$ python main.py --model-name 'timm model name' --dataset-name 'dataset name' --num-epochs 'epochs you want'

Sample input and output for finetuning:

To finetune a DeiT-Base model (deit_base_patch16_224) on CUB200 dataset for 1 epoch (with other parameters as default), we can execute as follows:

$ python main.py --model-name deit_base_patch16_224 --dataset-name CUB200 --num-epochs 1

Its expected output is shown below:

Evaluating models

To evalute a model already finetuned on, we can run the main.py as follows:

$ python main.py --training False --model-name 'timm model name' --dataset-name 'dataset name' --load-weights ./saved_model_weights_name

Sample input and output for evaluation:

To evaluate a finetuned/trained DeiT-Base model (deit_base_patch16_224) on CUB200 dataset with its given weights file, we can execute as follows:

$ python main.py --training False --model-name deit_base_patch16_224 --dataset-name CUB200 --load-weights ./saved_models/deit_224