| Data: | The ILSVRC2012 dataset (http://www.image-net.org/challenges/LSVRC/2012/) for image classification. |
|---|---|
| Purpose | This folder contains examples that demonstrate how to use CNTK to define AlexNet (http://papers.nips.cc/paper/4824-imagenet-classification-with-deep-convolutional-neural-networks) for image classification. |
| Network | AlexNet. |
| Training | Stochastic gradient descent with momentum. |
| Comments | See below. |
We use the ILSVRC2012 datasets to demonstrate how to train the AlexNet which won the ILSVRC 2012 challenge. AlexNet is an important milestone, as for the first time it was shown that deep convolutional neural networks can outperform traditional manual feature design for vision tasks by a significant margin.
ILSVRC2012 datasets are not included in the CNTK distribution. You may obtain it through http://image-net.org.
We give examples for both Python and BrainScript. Compared to the original AlexNet, and the Caffe implementation of AlexNet (https://github.com/BVLC/caffe/tree/master/models/bvlc_alexnet), our model differs slightly in that we no longer split the convolution layers into two groups (model parallelism). As a result our model has very slightly more parameters, but achieves better accuracy.
Pre-trained AlexNet models can be found here.