-
Notifications
You must be signed in to change notification settings - Fork 221
Construction from Layer specification objects. #19
Conversation
…ral network, also used internally. The previous tuple types are currently converted automatically. All tests pass.
…onstruction parameters. A global construction parameter called dropout (optionally a float, or default 0.5) is used as the fallback. Updates #12.
…n layers, and expects a 2D tuple for the size of the pools. Default is no pooling, so (1,1) but can be (2,2) or (4,4). Updates #3.
…ts API is obvious to use.
…ll of the tests accordingly.
…xamples to follow...
I think this one is ready for review, @ssamot. Thanks :-) |
I think this is overall great, but it could be the case that having just on Layer type might prove too confusing (it might make new users confused about having to define a softmax layer with kernel). On the other hand if you have each layer type in each class it might be too confusing as well, too many objects flying around. I think the proper level of abstraction is Layer, PoolingLayer and ConvolutionLayer - what do you think? |
I'm not sure about having a separate Pooling layer because it's implemented as part of the convolution in For the Convolution, that sounds reasonable... If we can support more convolution layer types at the same time, why not! I'll make a separate ticket though. |
Agreed - rest looks great and can be pulled |
Created new ticket #20 for improving the convolution support. |
Construction from Layer specification objects.
New syntax for specifying layers that makes it easier to pass-in (and document) parameters. This provides error checking by default, too.
Updates #3, #6, #12, #18.