Pytorch-cifar100

practice on cifar100 using pytorch

Requirements

This is my experiment eviroument, pytorch0.4 should also be fine

• python3.5

• pytorch1.0

• tensorflow1.5(optional)

• cuda8.0

• cudnnv5

• tensorboardX1.6(optional)

Usage

1. enter directory

$ cd pytorch-cifar100

2. dataset

I will use cifar100 dataset from torchvision since it's more convenient, but I also kept the sample code for writing your own dataset module in dataset folder, as an example for people don't know how to write it.

3. run tensorbard(optional)

Install tensorboardX (a tensorboard wrapper for pytorch)

$ pip install tensorboardX
$ mkdir runs
Run tensorboard
$ tensorboard --logdir='runs' --port=6006 --host='localhost'

4. train the model

Train all the model on a Tesla P40(22912MB)

You need to specify the net you want to train using arg -net

$ python train.py -net vgg16

sometimes, you might want to use warmup training by set -warm to 1 or 2, to prevent network diverge during early training phase.

The supported net args are:

squeezenet
mobilenet
mobilenetv2
shufflenet
shufflenetv2
vgg11
vgg13
vgg16
vgg19
densenet121
densenet161
densenet201
googlenet
inceptionv3
inceptionv4
inceptionresnetv2
xception
resnet18
resnet34
resnet50
resnet101
resnet152
preactresnet18
preactresnet34
preactresnet50
preactresnet101
preactresnet152
resnext50
resnext101
resnext152
attention56
attention92
seresnet18
seresnet34
seresnet50
seresnet101
seresnet152
nasnet

Normally, the weights file with the best accuracy would be written to the disk with name suffix 'best'(default in checkpoint folder).

5. test the model

Test the model using test.py

$ python test.py -net vgg16 -weights path_to_vgg16_weights_file

Implementated NetWork

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

• googlenet Going Deeper with Convolutions

• inceptionv3 Rethinking the Inception Architecture for Computer Vision

• inceptionv4, inception_resnet_v2 Inception-v4, Inception-ResNet and the Impact of Residual Connections on Learning

• xception Xception: Deep Learning with Depthwise Separable Convolutions

• resnet Deep Residual Learning for Image Recognition

• resnext Aggregated Residual Transformations for Deep Neural Networks

• resnet in resnet Resnet in Resnet: Generalizing Residual Architectures

• densenet Densely Connected Convolutional Networks

• shufflenet ShuffleNet: An Extremely Efficient Convolutional Neural Network for Mobile Devices

• shufflenetv2 ShuffleNet V2: Practical Guidelines for Efficient CNN Architecture Design

• mobilenet MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications

• mobilenetv2 MobileNetV2: Inverted Residuals and Linear Bottlenecks

• residual attention network Residual Attention Network for Image Classification

• senet Squeeze-and-Excitation Networks

• squeezenet SqueezeNet: AlexNet-level accuracy with 50x fewer parameters and <0.5MB model size

• nasnet Learning Transferable Architectures for Scalable Image Recognition

Training Details

I didn't use any training tricks to improve accuray, if you want to learn more about training tricks, please refer to my another repo, contains various common training tricks and their pytorch implementations.

I follow the hyperparameter settings in paper Improved Regularization of Convolutional Neural Networks with Cutout, which is init lr = 0.1 divide by 5 at 60th, 120th, 160th epochs, train for 200 epochs with batchsize 128 and weight decay 5e-4, Nesterov momentum of 0.9. You could also use the hyperparameters from paper Regularizing Neural Networks by Penalizing Confident Output Distributions and Random Erasing Data Augmentation, which is initial lr = 0.1, lr divied by 10 at 150th and 225th epochs, and training for 300 epochs with batchsize 128, this is more commonly used. You could decrese the batchsize to 64 or whatever suits you, if you dont have enough gpu memory.

You can choose whether to use TensorBoard to visualize your training procedure

Results

The result I can get from a certain model, since I use the same hyperparameters to train all the networks, some networks might not get the best result from these hyperparameters, you could try yourself by finetuning the hyperparameters to get better result.