██████╗  █████╗ ██████╗ ██████╗ ██╗     ███████╗     ██████╗██╗███████╗ █████╗ ██████╗ 
██╔══██╗██╔══██╗██╔══██╗██╔══██╗██║     ██╔════╝    ██╔════╝██║██╔════╝██╔══██╗██╔══██╗
██████╔╝███████║██║  ██║██║  ██║██║     █████╗      ██║     ██║█████╗  ███████║██████╔╝
██╔═══╝ ██╔══██║██║  ██║██║  ██║██║     ██╔══╝      ██║     ██║██╔══╝  ██╔══██║██╔══██╗
██║     ██║  ██║██████╔╝██████╔╝███████╗███████╗    ╚██████╗██║██║     ██║  ██║██║  ██║
╚═╝     ╚═╝  ╚═╝╚═════╝ ╚═════╝ ╚══════╝╚══════╝     ╚═════╝╚═╝╚═╝     ╚═╝  ╚═╝╚═╝  ╚═╝

PaddlePaddle cifar100

To explore the limit performance of opensource classifier on CIFAR100 with PaddlePaddle.

Repo dynamic

Relate update

• 2021.05.27: Releases the training of the MnasNet model
• 2021.04.24: Releases the training of the ResNeXt model
• 2021.04.16: Releases the training of the shufflenet series model
• 2021.04.09: Releases the training of the vgg model
• 2021.04.08: Releases the training of the MobileNet v2 model
• 2021.04.07: Releases the training of the MobileNet v1 model
• 2021.04.04: Releases the training of the ViT(vision transformer) variants model
• 2021.03.30: Releases the training of the ResNet series model

Model results

Some nets might get the best result from other hyperparameters, You can set up other hyperparameters for training.

network	params	top1 Acc	top5 Acc	yaml
resnet18	11.2M	0.7681	0.9345	common.yml
resnet34	21.3M	0.7835	0.9439	common.yml
resnet50	23.8M	0.8020	0.9530	common.yml
resnet101	42.8M	0.8016	0.9540	common.yml
resnet152	58.5M	0.8083	0.9549	common.yml
vit_p4_d6_h12_e384	10.7M	0.5689	0.8370	vit.yml
vit_p4_d6_h12_e192	2.7M	0.5428	0.8207	vit.yml
vit_p4_d12_h12_e384	21.4M	0.5570	0.8296	vit.yml
vit_p4_d12_h12_e192	5.4M	0.5375	0.8201	vit.yml
mobilenetv1 x1.0	3.3M	0.6607	0.8792	mobilenetv1.yml
mobilenetv1 x0.75	1.9M	0.6481	0.8646	mobilenetv1.yml
mobilenetv1 x0.5	0.88M	0.6095	0.8493	mobilenetv1.yml
mobilenetv1 x0.25	0.24M	0.5645	0.8292	mobilenetv1.yml
mobilenetv2 x1.0	0.24M	0.677	0.8912	mobilenetv2.yml
mobilenetv2 x0.75	1.5M	0.6685	0.8874	mobilenetv2.yml
mobilenetv2 x0.5	0.8M	0.6265	0.8631	mobilenetv2.yml
mobilenetv2 x0.25	0.38M	0.5631	0.831	mobilenetv2.yml
vgg19 BN	39.3M	0.716	0.8997	vgg.yml
vgg16 BN	34.0M	0.7174	0.9053	vgg.yml
vgg13 BN	28.7M	0.7133	0.9029	vgg.yml
vgg11 BN	28.5M	0.6773	0.8808	vgg.yml
shufflenet v1 groups3	1.04M	0.6966	0.8975	shufflenetv1.yml
shufflenet v2 scale1.0 swish	1.37M	0.7116	0.9116	shufflenetv2.yml
shufflenet v2 scale1.0 relu	1.37M	0.6924	0.8952	shufflenetv2.yml
swin transformer tiny	6.93M	0.5303	0.7833	swin_transformer.yml
resnext50 32x4d	23.2M	0.7744	0.934	common.yml
resnext101 32x8d	87.1M	0.7938	0.9407	common.yml
mnasnet 1.0	3.3M	0.7039	0.8914	mnasnet.yml

Test curve

csv_dir download from VisualDL scalar

Requirements

python3 -m pip install -r requirements.txt

Usage

Training & Evaluation

Training and Evaluation are put together, using PaddlePaddle HighAPI(hapi). To train baseline PaddlePaddle-cifar100-resnet18 on a single gpu for 300 epochs run:

python3 main.py -c resnet18 -y yamls/common.yml

Visualization

VisualDL: VisualDL is the PaddlePaddle visual analysis tool.

visualdl --logdir logs --host 0.0.0.0

VisualDL display details:

Relevant papers

• ResNet: Deep Residual Learning for Image Recognition
• Transformer: Attention is not all you need: pure attention loses rank doubly exponentially with depth
• MobileNets: MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications
• MobileNetV2: MobileNetV2: Inverted Residuals and Linear Bottlenecks
• VGG: Very Deep Convolutional Networks for Large-Scale Image Recognition
• ShuffleNet: ShuffleNet: An Extremely Efficient Convolutional Neural Network for Mobile Devices
• ShuffleNet V2: ShuffleNet V2: Practical Guidelines for Efficient CNN Architecture Design
• Swin Transformer: Swin Transformer: Hierarchical Vision Transformer using Shifted Windows
• ResneXt: Aggregated Residual Transformations for Deep Neural Networks
• MnasNet: MnasNet: Platform-Aware Neural Architecture Search for Mobile