All models are hosted at http://data.dmlc.ml/mxnet/models/ and licensed udner CC0.
This model is a pretrained model on ILSVRC2012 dataset. The model is trained by only random crop and mirror augmentation. This model is able to achieve 89.9% Top-5 accuracy on ILSVRC2012-Validation Set.
This model is converted from TensorFlow released pretrained model. By single crop on 299 x 299 image from 384 x 384 image, this model is able to achieve 76.88% Top-1 Accuracy and 93.344% Top-5 Accuracy on ILSVRC2012-Validation Set.
This model is a pretrained model on full imagenet dataset with 14,197,087 images in 21,841 classes. The model is trained by only random crop and mirror augmentation. This model is able to achieve 37.19% Top-1 accuracy on training data. This model is about 50% more complex than standard Inception-BN Network