DL-DIY potential project ideas

• try out out-of-distribution detection on CamVid-OOD semantic segmentation [original train set] [OOD test set]
• for image classification assess the impact of mixing techniques (mixup, cutmix, etc.) on performance
• train improved model and evaluate performance on distribution shift settings, e.g., CIFAR-C

---

NeuBoots

This repository contains an official implementation of the Neural Bootstrapper (NeurIPS 2021).

Prerequsites

• python==3.7
• pytorch==1.2
• torchvision
• tqdm
• PIL

Configuration script

Before running main.py, please make an ini script file under the script folder for configurations.

Example of ini_file:

[default]

dataset = cifar100
output_dir = outs
num_epoch = 200
dist = False
phase = train
cpus = 4
gpus = '2'
model = resnet34
is_nbs = True
num_classes = 100
lr = 0.1
weight_decay = 0.0005
optim = sgd
batch_size = 128
n_a = 400
num_bs = 100
dropout_rate = 0.
scheduler = cosine
epoch_th = 30

Run

CIFAR-10

➜ python main.py cutout/cifar10

CIFAR-100

➜ python main.py cutout/cifar100

SVHN

➜ python main.py cutout/svhn

Citation

@inproceedings{neuboots2021,
  title={Neural Bootstrapper},
  author={Shin, Minsuk and Cho, Hyungjoo and Min, Hyun-seok and Lim, Sungbin},
  booktitle={Thirty-Fifth Conference on Neural Information Processing Systems},
  year={2021}
}