Homeworks and the final project of Deep Learning, NYU 2021 Spring.
You can find class repo here: https://github.com/Atcold/NYU-DLSP21
This final project is an image classification task with 25,600 labeld train images, 25,600 labeled validation images and 512,000 unlabeled images. There are a total of 800 classes.
We used MoCo as our semi-supervised learning framework and tried to combine design choices from other models. Most of the codes is from https://github.com/facebookresearch/moco.
We reached 30.48% accuracy with 250 epchos of pretrain + 100 epochs frozen-features classifier + 100 epochs unfrozen-features classfier.