• Chainer
  • $ git clone https://github.com/pfnet/chainer.git
  • $ cd chainer; python setup.py install
  • $ pip install chainer-cuda-deps
• progressbar2
  • $ pip install progressbar2

$ bash download.sh

$ python dataset.py

$ python train.py

See the help messages with --help option for details.

You can choose model file to be trained from models dir. Cifar10, Network In Network (NIN), and VGG-net (with some variants) are already prepared. The architecture of VGG_mini is derived from here. The original paper of VGG-net is found in here.

• There are four types of base architectures

  • Cifar10
  • NIN (Network In Network)
  • VGG (similar to original VGG-net)
  • VGG_mini (same as nagadomi's model)

• Variants are different at Batch Normalization and Parameterized ReLU

  • BN: Batch Normalization for first two blocks of convolutional layer
  • ABN: Batch Normalization for all convolutional layers
  • PReLU: Parameterized ReLU for all activation functions

• VGG_mini_PReLU didn't perform well

• Default setting ($ python train.py) is VGG_mini_ABN, and it performed the below result (92.28% accuracy at epoch 150):

I don't know why but chainer.functions.accuracy function returns a little better value. So I tested experimental results with test.py.

$ python test.py --eval normal \
--model results/VGG_mini_ABN/VGG_mini_ABN.py \
--param results/VGG_mini_ABN/VGG_mini_ABN_Adam_epoch_100.chainermodel \
--norm 0 --batchsize 128 --gpu 0

$ python draw_loss.py --logfile log.txt --outfile log.jpg

• The label vector will be passed to softmax_cross_entropy should be in shape (N,). (N, 1) causes divergence of weights.
• If model.to_cpu() and model.to_gpu() are called during training, test scores are fixed and never updated.