Awesome CIFAR Zoo

This repository contains the pytorch code for multiple CNN architectures and improve methods based on the following papers, hope the implementation and results will helpful for your research!!

• Architecure

• (lenet) LeNet-5, convolutional neural networks

• (alexnet) ImageNet Classification with Deep Convolutional Neural Networks

• (vgg) Very Deep Convolutional Networks for Large-Scale Image Recognition

• (resnet) Deep Residual Learning for Image Recognition

• (preresnet) Identity Mappings in Deep Residual Networks

• (resnext) Aggregated Residual Transformations for Deep Neural Networks

• (densenet) Densely Connected Convolutional Networks

• (senet) Squeeze-and-Excitation Networks

• (bam) BAM: Bottleneck Attention Module

• (cbam) CBAM: Convolutional Block Attention Module

• (genet) Gather-Excite: Exploiting Feature Context in Convolutional Neural Networks

• (sknet) SKNet: Selective Kernel Networks

• Regularization

• (shake-shake) Shake-Shake regularization

• (cutout) Improved Regularization of Convolutional Neural Networks with Cutout

• (mixup) mixup: Beyond Empirical Risk Minimization

• Learning Rate Scheduler

• (cos_lr) SGDR: Stochastic Gradient Descent with Warm Restarts

• (htd_lr) Stochastic Gradient Descent with Hyperbolic-Tangent Decay on Classification

Requirements and Usage

Requirements

• Python (>=3.6)

• PyTorch (>=1.1.0)

• Tensorboard(>=1.4.0) (for visualization)

• Other dependencies (pyyaml, easydict)

PS: for TensorboardX version, check tag pt1.0

pip install -r requirements.txt

Usage

simply run the cmd for the training:

## 1 GPU for lenetCUDA_VISIBLE_DEVICES=0 python -u train.py --work-path ./experiments/cifar10/lenet## resume from ckptCUDA_VISIBLE_DEVICES=0 python -u train.py --work-path ./experiments/cifar10/lenet --resume## 2 GPUs for resnet1202CUDA_VISIBLE_DEVICES=0,1 python -u train.py --work-path ./experiments/cifar10/preresnet1202## 4 GPUs for densenet190bcCUDA_VISIBLE_DEVICES=0,1,2,3 python -u train.py --work-path ./experiments/cifar10/densenet190bc

We use yaml file config.yaml to save the parameters, check any files in ./experimets for more details.
You can see the training curve via tensorboard, tensorboard --logdir path-to-event --port your-port.
The training log will be dumped via logging, check log.txt in your work path.

Results on CIFAR

Vanilla architectures

With additional regularization

PS: the default data augmentation methods are RandomCrop + RandomHorizontalFlip + Normalize,
and the √ means which additional method be used.