The CIFAR-10 classifiers of various neural networks. Datasets from cifar-10 datasets.

tqdm

chainer v1.5.1

Other packages around chainer.

$ wget https://www.cs.toronto.edu/~kriz/cifar-10-python.tar.gz
$ gzip -d cifar-10-python.tar.gz
$ tar -xf cifar-10-python.tar
$ python data.py

Training can be done in various algorithms (optimizer = {Adam, AdaGrad, SGD}, with/without batch-normalization, dropout rate, different data augumentation methods) and neural architectures {Fully-connected neural network, several Convolutional neural networks}.

In either case, if you use GPU, add option -g 0 or --gpu 0. If you want to export result figures, -p on or --plot on, write log, -l on or --log on, and save models -s on or --save on.

$ python train_nn.py -d on

"Simple" means fully-connected. To know the model, see model_nn.py. -d on option makes input data normalized and augumented. To know detail, see datahandler.py.

The test accuracy should be around 60%.

$ python train_cnn.py -d on

You can change models by options -m alexor -m alexbn, which represents with/without batch-normalization layers. To know the model, see model_cnn.py.

The test accuracy should be around 85%.

$ python train_cnn_crop.py -d on

By cropping the images, input data is augumented further and the prediction accuracy will be even better! The test accuracy should be...

shiba24, Jan, 2016