Training DNN models with CIFAR-10/MNIST/FashinMNIST dataset

• Python 3.6+
• PyTorch 1.0+

Use cifar10.py for the training.

$ ./cifar10.py

Default configurations are defined in Params class of cifar-10.py

Some configurations can be overridden with command line options.

  --model   : Specify the model for a training. (Case sentitive)
              Currently, vgg{9,11,13,16,19}, resnet{9,18,34,50,101,152},
              wrn_16_10, wrn_28_10
              (Default : resnet18)
  --epochs  : Specify how meny epochs the model is trained. (Default 16)
  --workdir : Specify working directory to store weights and checkpoint.
              (Default: Current directory)
  --resume  : Resume from checkpoint
  --no_cuda : Force to run on CPU

cifar10.py script can be also running on Jupyter Netbook. (And Google Colab)
Please copy models directory at the same location of the ipynb and use Params class to configure.
On the Jupyter Notebook, you can see loss/accuracy results with matplotlib.

• Training ResNet18 model.

$ ./cifar10.py --model resnet18

You can get weights cifar10-resnet18_weights_xxx.pth for each xxx epochs under weights directory. symlink cifar10-resnet18_weights.pth points to the best accuracy weights.

• Resume from vgg11 training and continue to train until 50th epoch.

$ ./cifar10.py --model vgg11 --resume --epochs 50

This project optionally supports MNIST/FashinMNIST dataset.
You can use mnist.py and fashion_mnist.py for training the models with mnist dataset.
AlexNet and VGG is tailored for 32x32 input, so that these models can't use with MNIST/FashionMNIST.

• Input image data is not normalized (To use weight with deploying model to AI SoCs)
• Optimizer : Adam
• Learning Rate : 0.01
• Epochs : 16

• AlexNet / CIFAR-10
• ResNet9 / CIFAR-10
• ResNet18 / CIFAR-10
• ResNet34 / CIFAR-10
• ResNet50 / CIFAR-10
• WideResNet 16-8 / CIFAR-10
• WideResNet 16-10 / CIFAR-10
• WideResNet 28-8 / CIFAR-10
• WideResNet 28-10 / CIFAR-10
• VGG 11 / CIFAR-10
• VGG 13 / CIFAR-10
• VGG 16 / CIFAR-10
• VGG 19 / CIFAR-10
• ResNet9 / MNIST
• ResNet18 / MNIST
• ResNet9 / FashionMNIST
• ResNet18 / FashionMNIST

This project is licensed under the MIT License - see the LICENSE file for details

AlexNet, ResNet, WideResNet, and VGG model is ported from bearpaw/pytorch-classification