Train CIFAR10 with PyTorch
PyTorch on the CIFAR10 dataset with various up-to-date neural networks
Additional Updates
Training
- Can be run using IDE such as visual code.
- Can use terminar by
python main.py --network_name VGG16
- Currently
New features
-
Save file with all support network name
-
Added save check point for every epoch
- Best accuracy epoch is save at checkpoint/network/network_best.pth
- Option to save network at each epoch
- Network will automatic resum training
- Train from begining (i.e. epoch 0) by
python main.py --network_name VGG16 --resume False
- E.g. VGG11 training data is saved at checkpoint/VGG11/VGG11_epoch#no_epoch.pth
-
Learning rate is now controlled
-
Bring tensorboardx Support
- Using command
tensorboard --logdir=log_dir --host localhost --port 8088
- Open browser at http://localhost:8088/
- Only save scalar values of training testing accuracy, loss, epoch, and learning rate
-
Avaiabled trained data, learning rate are at range [0 25 50 100 ] are 0.1 ,0.01, and 0.0001 (total 100 epochs). In the original repo, author use learning rate of 0.1, 0.01, and 0.001 for each 150 epochs (total 450 epochs).
Todo
- Pretrain data of all network
Accuracy
At best test accuracy epoch
Model | Train Acc. | Test Acc. | Ref. Acc | Year |
---|---|---|---|---|
VGG11 | 99.17% | 90.92% | 92.64% | 2014 |
VGG13 | 99.49% | 90.73% | - | 2014 |
VGG16 | 99.42% | 90.76% | - | 2014 |
VGG19 | 99.15% | 90.15% | - | 2014 |
GoogLeNet/Inception v1 | 99.93% | 93.70% | - | 2015 |
ResNet18 | 99.76% | 94.14% | 93.02% | 2015 |
ResNet34 | 99.89% | 94.60% | - | 2015 |
ResNet50 | -% | -% | - | 2015 |
ResNet101 | -% | -% | - | 2015 |
ResNet152 | -% | -% | - | 2015 |
PreActResNet18 | -% | -% | 95.11% | 2016 |
PreActResNet34 | -% | -% | - | 2016 |
PreActResNet50 | -% | -% | - | 2016 |
PreActResNet101 | -% | -% | - | 2016 |
PreActResNet101 | -% | -% | - | 2016 |
DenseNet121 | 99.79% | 94.26% | 95.04% | 2016 |
DenseNet161 | -% | -% | - | 2016 |
DenseNet169 | -% | -% | - | 2016 |
DenseNet201 | -% | -% | - | 2016 |
ResNeXt29(2x64d) | 99.81% | 94.25% | 94.82% | 2016 |
ResNeXt29(4x64d) | 99.92% | 94.04% | - | 2016 |
ResNeXt29(8x64d) | -% | -% | - | 2016 |
ResNeXt29(16x64d) | -% | -% | - | 2016 |
ResNeXt29(32x4d) | 99.85% | 94.55% | 94.73% | 2016 |
SENet18 | 99.55% | 93.56% | - | 2017 |
ShuffleNetG2 | 96.18% | 90.32% | - | 2017 |
ShuffleNetG3 | 96.29% | 90.71% | - | 2017 |
ShuffleNetG8 | -% | -% | - | 2017 |
ShuffleNetV2 | -% | -% | - | 2018 |
PNASNetA | -% | -% | - | 2017 |
PNASNetB | -% | -% | - | 2017 |
DPN26 | 99.61% | 93.86% | 95.16% | 2017 |
DPN92 | -% | -% | - | 2017 |
MobileNet | 95.02% | 89.83% | - | 2017 |
MobileNetV2 | 97.12% | 92.17% | 94.43% | 2018 |
EfficientNetB0 | 91.26% | 87.85% | - | 2019 |
Ref. accuracy is reported in the original repo at KuangLiu Repo