- 深度学习 14 1
this is sth related to matrix decomposition, when there's multiple matrix multiplication, the eigen value define the scale...(sorr...
the 3 r-g-b channel activation feature map will sum up to get a new feature. So the number is only determined by the filter number...
You just let the net structure have 2x9+4x9 units in the later layer, the training will automatically regress to these numbers wit...
Since the initial number are random initialized, the numbers are different from the 1st seconds, so don't need to worry about this...
- 2017-07-10 23:39
深度学习 14 1