search

snuspl / dolphin

14 stars 2 forks source link

Remove required numInput and numOutput options for layer configuration #150

Closed beomyeol closed 8 years ago

beomyeol commented 8 years ago

This pull request removes num_input and num_output options for layer configurations. The num_input and num_output not only are meaningful only with 1-dimensional inputs, but also input and output shapes of layers can be inferred by other options for layers, such as kernel size for convolutional layers and the number of outputs for fully connected layers.

Moreover, the codes for instantiating layers and generating layer parameters are repeatedly used at several points. This PR introduces the utility class for neural network, NeuralNetworkUtils, for neural network related operations including generating layer instances and layer parameters.

In addition, this pull request brings back the support for arbitrary dimensional inputs that was lost in #148. While input data had its dimension information with ND4J library, each layer has its input shape with which layers can process spatial operations on input data.

This closes #146.

jsjason commented 8 years ago

I'll do a pass today.

jsjason commented 8 years ago

I finished my pass, @beomyeol.

bgchun commented 8 years ago

@beomyeol @jsjason Is this PR ready to be merged?

jsjason commented 8 years ago

@bgchun @beomyeol Sorry, I didn't have the chance to do another pass up till now. I'll do a review right away.

jsjason commented 8 years ago

@beomyeol I did another pass. I'll merge this right away after you check my comments. Thanks.

beomyeol commented 8 years ago

@jsjason Thank you for your review. I've addressed your comments and pushed one more commit for updating README. Please take another look :)

jsjason commented 8 years ago

@beomyeol I did another pass for the new README commit you added. Thanks.

beomyeol commented 8 years ago

@jsjason Thank you for your comments. I've address them.

jsjason commented 8 years ago

@beomyeol Great. I'll merge.