Support for GoogLeNet about digits HOT 4 CLOSED

This is not a bug, it's expected behavior. But I'll admit it's pretty annoying. If you click on the helptip next to "Custom Network" on the "Custom Network" tab, you'll see some more information about the current limitations on creating custom networks DIGITS.

GoogLeNet does some really cool tricks in their network (See this paper, page 7 for a visualization of the network structure). The problem is that the training network (train_val.prototxt) and the deploy network (deploy.prototxt) are drastically different.

Having two textareas for submitting a train_val network and a deploy network is cumbersome. In anticipation of this change: BVLC/caffe#1245, I just put one textarea in DIGITS and do some ugly, tedious stuff in the backend to split that one network into the two networks required by caffe.

Once that "All-in-one nets" feature comes out, you'll be able to use GoogLeNet and other interesting architectures, but for now you have to stick with the format used by LeNet, AlexNet and VGG (and many others).

from digits.

Closing this bug as it is an intentional limitation for now.

from digits.

@lukeyeager Is there a simple way to import GoogLeNet net definition to DIGITS for model training now? Since we don't have idea when the "All-in-one nets" will be ready.

from digits.

I'm reopening [and renaming] this issue since it's coming up a lot.

from digits.