dlcheng / cifar10-keras

91.5% accuracy of CIFAR10 on keras

cifar10-keras

To train the model, run

python train.py

a training log will be saved at the same time.

The CNN structure is:

[CONV 64 + BN + DROP 0.4] [CONV 64 + BN] [CONV 64 + BN] [Pooling 2X2]

[CONV 128 + BN + DROP 0.4] [CONV 128 + BN] [CONV 128 + BN] [Pooling 2X2]

[CONV 256 + BN + DROP 0.4] [CONV 256 + BN] [CONV 512 + BN] [CONV 512 + BN] [Pooling 2X2]

[Flatten] [FC 512 + BN + DROP 0.5] [FC 10]

BN is the batchnormalization that regulates the input of each slice. It can speed up the training significantly.

Results: After 300 epochs, this model can reach 91.5% accuracy on the validation data with image augmentation.

Each epoch spends 97s on GTX 1060.

The accuracy should be in the top 20 of Kaggle CIFAR10 competition. See the score board https://www.kaggle.com/c/cifar-10/leaderboard