Find file History
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Type Name Latest commit message Commit time
Failed to load latest commit information.


This is an implementation of a deep convolutional neural network model inspired by the paper Krizhevsky,Sutskever, Hinton 2012 used to classify images from the ImageNet 2012 competition. The model used here is adapted from the Caffenet architecture described here.

The neon implementation does not use the conv layer weight grouping as used in the Caffe model.

Model script

The model run script is included below

Trained weights

The trained weights file can be downloaded from AWS using the following link: [trained Alexnet model weights][S3_WEIGHTS_FILE] [S3_WEIGHTS_FILE]:


This model is acheiving 58.6% top-1 and 81.1% top-5 accuracy on the validation data set after 90 epochs of training. The training here is using a single, random crop on every epoch and flipping the images across the vertical axis. These results improve further with additional data augmentation added to the training as decsribed in Krizhevsky,Sutskever, Hinton 2012.


To run the model, first the ImageNet data set needs to be uploaded and converted to the format compatible with neon (see instructions). Note there has been some changes to the format of the mean data subtraction; users with the old format may be prompted to run an update script before proceeding.

This script was tested with the neon release v1.4.0 (commit SHA bc196cb). Make sure that your local repo is synced to this commit and run the installation procedure before proceeding.

If neon is installed into a virtualenv, make sure that it is activated before running the commands below. Also, the commands below use the GPU backend by default so add -b cpu if you are running on a system without a compatible GPU.

To test the model performance on the validation data set use the following command:

python -w path/to/dataset/batches --model_file alexnet.p --test_only

To train the model from scratch for 90 epochs, use the command:

python -w path/to/dataset/batches -s alexnet_weights.p -e 90

Additional options are available to add features like saving checkpoints and displaying logging information, use the --help option for details.


Machine and GPU specs:

Intel(R) Core(TM) i5-4690 CPU @ 3.50GHz
Ubunutu 14.04
CUDA Driver Version 7.0

The run times for the fprop and bprop pass and the parameter update are given in the table below. The iteration row is the combined runtime for all functions in a training iteration. These results are for each minibatch consisting of 128 images of shape 224x224x3. The model was run 12 times, the first two passes were ignored and the last 10 were used to get the benchmark results.

|    Func     |      Mean    |
| fprop       |   40.1 msec  |
| bprop       |   83.1 msec  |
| iteration   |  123.2 msec  |


ImageNet Classification with Deep Convolutional Neural Networks
Alex Krizhevsky and Sutskever, Ilya and Geoffrey E. Hinton
Advances in Neural Information Processing Systems 25
eds.F. Pereira and C.J.C. Burges and L. Bottou and K.Q. Weinberger
pp. 1097-1105, 2012