add CapsNet example #8787
add CapsNet example #8787
Conversation
|
Do you actually update your bias in capsule layer in each training iteration? |
|
As stated in the routing algorithm of CapsNet (line 2), |
| * * * | ||
| ## **Prerequisities** | ||
|
|
||
| MXNet version above (0.11.0) |
There was a problem hiding this comment.
For the version here, have you tested on some more recent versions like the current master or 1.0.0rc?
There was a problem hiding this comment.
I tested it on MXNet 0.12.1 and it works well
example/capsnet/capsulenet.py
Outdated
| data_flatten = mx.sym.flatten(data=data) | ||
| squared_error = mx.sym.square(x_recon-data_flatten) | ||
| recon_error = mx.sym.mean(squared_error) | ||
| loss = mx.symbol.MakeLoss((1-0.392)*margin_loss(y_onehot, out_caps)+0.392*recon_error) |
There was a problem hiding this comment.
I think it's better to make 0.392 as part of the function arguments.
There was a problem hiding this comment.
Thank you for review and I added 0.392 as option named recon_loss_weight
|
@piiswrong Do you think this example can be merged in? It looks good. |
|
@sxjscience Thank you for review, |
|
@piiswrong Can this example be merged? |
|
@Soonhwan-Kwon Yes. Could you rebase to master? needs CI to pass |
|
@piiswrong Thank you for review. I rebased to the master. |
* add capsnet example's layer * add capsnet example * add recon_loss_weight option and tensorboard for plot * update readme to install tensorboard * fix print of loss scaled to 1/batchsize
* add capsnet example's layer * add capsnet example * add recon_loss_weight option and tensorboard for plot * update readme to install tensorboard * fix print of loss scaled to 1/batchsize
Description
This example is MXNet implementation of CapsNet:
Sara Sabour, Nicholas Frosst, Geoffrey E Hinton. Dynamic Routing Between Capsules. NIPS 2017
We achieved
the best test error rate=0.29%andaverage test error=0.303%. It is the best accuracy and fastest training time result among other implementations(Keras, Tensorflow at 2017-11-23).The result on paper is
0.25% (average test error rate).※ tensorflow implementation's batch size is 128 but MXNet and Keras implementation 's batchsize are 100.
Checklist
Essentials
make lint)Comments