Re-implementation of ConvNets on CIFAR-100 with PyTorch

Contact email: imdchan@yahoo.com

Introduction

Here are some re-implementations of Convolutional Networks on CIFAR-100 dataset.

Note that the training set that consists of 50k training images was divided into 45k/5k train/val split. So I first made stratefied 10-fold split, resulting in the 'train_folds.csv'.

Requirements

• A single TITAN RTX (24G memory) is used.

• Python 3.7+

• PyTorch 1.0+

Usage

1. Clone this repository

    git clone https://github.com/longrootchen/cifar100-pytorch.git
   

2. Train a model, taking resnext29_16x64d as an example

    python -u train.py --work-dir ./experiments/resnext29_16x64d --resume ./experiments/resnext29_16x64d/checkpoints/last_checkpoint.pth
   

3. Evaluate a model, taking resnext29_16x64d as an example

    python -u eval.py --work-dir ./experiments/resnext29_16x64d --ckpt-name last_checkpoint.pth
   

Results

Error Rate (%)	original paper	re-implementation
ResNeXt-29, 8x64d	17.77 [1]	19.06
ResNeXt-29, 16x64d	17.31 [1]	18.75
DenseNet-100-BC, k=12	22.27 [2]
DenseNet-250-BC, k=24	17.60 [2]
DenseNet-190-BC, k=40	17.18 [2]
SE-ResNet-110	23.85 [3]
SE-ResNet-164	21.31 [3]

References

[1] Saining Xie, Ross Girshick, Piotr Dollár, Zhouwen Tu, Kaiming He. Aggregated Residual Transformations for Deep Neural Networks. In CVPR, 2017.

[2] Gao Huang, Zhuang Liu, Laurens van der Maaten, Kilian Q. Weinberger. Densely Connected Convolutional Networks. In CVPR, 2017.

[3] Jie Hu, Li Shen, Samuel Albanie, Gang Sun, Enhua Wu. Squeeze-and-Excitation Networks. In CVPR, 2018.