Directly provide pre-trained models

[futurely]

MXNet's recent development was several times more active than Caffe's. Pull requests in this project were reviewed and merged in days instead of in months.

https://github.com/dmlc/mxnet/pulse/monthly
https://github.com/BVLC/caffe/pulse/monthly
September 22, 2015 – October 22, 2015

Metrics	MXNet	Caffe
Active Pull Requests	154	52
Merged Pull Requests	153	38
Proposed Pull Requests	1	14
Authors Pushed Commits	25	37
Commits Pushed to master	422	48
Commits Pushed to All Branches	1000	103
Files Changed on Master	324	237
Additions on Master	25,233	4,463
Deletions on Master	2,904	1,896

On the other hand, Caffe had 6480 stars and 3714 forks with a fork vs. star ratio of 0.57 while MXNet had 669 stars and 186 forks with a ratio of 0.28 as of the time of writing. MXNet's active user conversion rate was only half of Caffe's.

Being able to easily download pre-trained models from a centralized place was one of the key factors why Caffe became so popular because it greatly saved the user's time to get near the state of the art baseline feature extractors for further fine-tuning on their own datasets. Unfortunately, there's not yet a similar model zoo for MXNet. All that can be found now is just a simple tool to convert Caffe models.

MXNet should setup a model zoo as soon as possible to quickly increase the active users.

[antinucleon]

Thanks for advice! It's very valuable and informative.

We will move the pretrained model from notebook to some kind of model zoo this week. What's more, we will release a model trained on 21k classes.

[antinucleon]

Update: http://mxnet.readthedocs.org/en/latest/pretrained.html

[futurely]

Christian Szegedy, Julian Ibarz, and Vincent Vanhoucke. Scene Classification with Inception-7. CVPR Large-scale Scene Understanding Challenge Workshop 2015.