Features : tensorflow, ensemble of 3 models (VGG-like with batch size 64/128, resnet 32layers), best accuracy with a single model is 99.74%, data augmentation (rotation, shift, zoom)
Features : tensorflow, ensemble of 5 models obtained with different hyper-params and same architecture (4 conv-layers, 1 fc-layer), best accuracy with a single model is 0.9968
Features : keras (theano-base), ensemble of 5 models obtained with different hyper-params and same architecture (6 conv-layers), data augmentation (elastic distortion)
Features : tflearn, ensemble of 11 models (5 conv-nets, 3 highway-nets, 3 rnn), weights for ensemble are also trained, data augmentation (shift, rotation, blur)
Features : keras, ensemble of 3 models obtained with different filter size and same architecture (VGG-like), best accuracy with a single model is 0.9959, data augmentation (shift, rotation)
Features : tensorflow, ensemble of 5 models obtained with same hyper-params and same architecture (VGG-like), best accuracy with a single model is 0.9935, data augmentation (scale, rotation)
Features : keras, ensemble of 50 models obtained with same hyper-params and same architecture (3 conv-layers, 1 fc-layer), data augmentation (infmnist)