Awesome
Pytorch-Tiny-ImageNet
Installation
If you are familiar with poetry, you can install dependencies with poetry install. Otherwise, you can install dependencies with requirements.txt
Trouble shooting with OpenCV here
Dataset
python prepare_dataset.py will download and preprocess tiny-imagenet dataset.
In the original dataset, there are 200 classes, and each class has 500 images.
However, in test dataset there are no labels, so I split the validation dataset into validation and test dataset. (25 per class)
Probably not the best train(500), val(25), test(25) splitting method, but I think it's good enough for this project to evaluate transfer learning.
The dataset is then resized from 64x64 to 224x224.
If you don't want to prepare dataset, you can download dataset
Summary
Goal of this project is to evaluate transfer learning on tiny-imagenet dataset.
Tiny-ImageNet dataset has images of size 64x64, but ImageNet dataset is trained on 224x224 images. To match the input size, I resized tiny-imagenet dataset to 224x224 and trained on pretrained weight from ImageNet.
Finetune few layers, and use pretrained weight from 224x224 trained model to retrain 64x64 image on ResNet18
Test Result
| Model | Test Result | Input size | pretrained weight |
|---|---|---|---|
| AlexNet | 35.88% | 64x64 | ImageNet |
| ResNet18 | 53.58% | 64x64 | ImageNet |
| ResNet18 | 69.62% | 224x224 | ImageNet |
| ResNet18 | 69.80% | 64x64 | Model Above |
Updates
This project was done as part of Udacity Machine Learning Engineer Nanodegree Capstone Project in 2018.
I haven't updated the code since then, but I decided to update this project thanks to a lot of interest and stars from the community.
Since then, I have been working as Data Engineer, and I have improved my python programming skills.
This version includes the following changes:
- Poetry for dependency management
- Pytorch with M1 Mac GPU support
- Dataset download and preprocessing with python