From: MyungJoo Ham Date: Mon, 28 May 2018 06:54:07 +0000 (+0900) Subject: [Filter] Add Readme entry for developers. X-Git-Tag: v0.0.1~212 X-Git-Url: http://review.tizen.org/git/?a=commitdiff_plain;h=4bbf874272e471556d76afb9fc8a43cf7f60caac;p=platform%2Fupstream%2Fnnstreamer.git [Filter] Add Readme entry for developers. Basic requirements for tensor_filter component are described. Signed-off-by: MyungJoo Ham --- diff --git a/tensor_filter/README.md b/tensor_filter/README.md index e69de29..f206ee6 100644 --- a/tensor_filter/README.md +++ b/tensor_filter/README.md @@ -0,0 +1,28 @@ +# NNStreamer::tensor_filter + +## Motivation + +This is the main element of the whole NNStreamer suite. This connects gstreamer data stream with neural network frameworks (NNFW) such as Tensorflow or Caffe. ```tensor_filter``` is supposed to attach an instance of neural network model with the given NNFW as a filter to a gstreamer stream. The input/output stream data type is ```other/tensor```. + +## Components + +### Main ```tensor_filter.c``` + +This is the main placeholder for all different subcomponents. With the property, ```FRAMEWORK```, this main component loads the proper subcomponent (e.g., tensorflow-lite support, custom support, or other addtional NNFW supports). + +The main component is supposed process the standard properties for subcomponents as well as processing the input/output dimensions. + +The subcomponents as supposed to fill in ```GstTensor_Filter_Framework``` struct and register it with ```supported``` array in ```tensor_filter.h```. + +Note that the registering sturcture may be updated later. (We may follow what ```Linux.kernel/drivers/devfreq/devfreq.c``` does) + +### Tensorflow-lite support, ```tensor_filter_tensorflow_lite.c``` + +This should fill in ```GstTensor_Filter_Framework``` supporting tensorflow_lite. + +### Custom function support, ```tensor_filter_custom.c``` + +This should fill in ```GstTensor_Filter_Framework``` supporting dlopen'ed custom shared objects, requiring such shared objects to provide its own defined functions. + +### We may add other NNFW as well (tensorflow, caffe, ...) +