Implement new initializations for conv neuralnet layer #2797
Conversation
| @@ -280,6 +292,10 @@ void CNeuralConvolutionalLayer::init() | |||
| SG_ADD(&m_pooling_height, "pooling_height", "Pooling Height", MS_NOT_AVAILABLE); | |||
| SG_ADD(&m_stride_x, "stride_x", "X Stride", MS_NOT_AVAILABLE); | |||
| SG_ADD(&m_stride_y, "stride_y", "Y Stride", MS_NOT_AVAILABLE); | |||
|
|
|||
| // SG_ADD(&m_initialization_mode, "initialization_mode", "Initialization Mode", | |||
There was a problem hiding this comment.
Shall I typecast it to machine_int_t* like done below or make a new Parameter::add and a new entry in lib/DataType.h?
|
@lisitsyn Made the changes you suggested. |
| * of initialization of the network */ | ||
| INIT_MANUAL=0, | ||
| /** Sets standard deviation to | ||
| * \sqrt{\frac{2}{input\_width\times input\_height\times input\_channels}} |
There was a problem hiding this comment.
Please use doxygen syntax for equations like here: https://github.com/shogun-toolbox/shogun/blob/develop/src/shogun/kernel/GaussianKernel.h
|
@lisitsyn sorry for the delay. I had my end-semester exams. made the changes you suggested. Please review. |
|
@sanuj this time excuse my delay :) can we name MANUAL thing as NORMAL and AUTO as HE_NORMAL (as per author of the paper, like in Keras https://github.com/fchollet/keras/blob/master/keras/initializations.py) ? |
|
@lisitsyn I have updated the code although for some reason i'm unable to run the unit-tests locally. Whenever I do |
|
@lisitsyn ping :) |
|
Ok this looks ok, thanks! |
Implement new initializations for conv neuralnet layer
PR for issue #2700
@lisitsyn Please have a look. I'm not sure if this is the correct way to implement this.