This means to support most caffe models. Need a lot of work.
-
Modify the python script in convertor/src/net.py and convertor/src/layer.py, to generate a JSON file that contains multibranch information.
-
Create some kind of
mergelayer
orconcatLyaer
in kernels/net/net.cl, that accepts multipleinputFeatureMap
that current implementation doesn't support. You will need to have
- different access pattern for
readFmCache
than other layers
-
add the newly created
LayerType
in cpp/network.h -
add corresponding code for
forward
andLayer
method in cpp/network.cpp.-
Layer
method should process the multibranch information in JSON, thenext*, prev*
pointer may be changed tonext**, prev**
. -
The
forward
method is more complex than usual, since it needs multipleinputFeatureMap
. You may need to- Transfer rest of those
inputFeatureMap
even in FPGA memory traffic reduction mode.
- Transfer rest of those
-
-
Test/Debug on cpu/fpga