[
https://issues.apache.org/jira/browse/MINIFICPP-301?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16251539#comment-16251539
]
ASF GitHub Bot commented on MINIFICPP-301:
------------------------------------------
Github user achristianson commented on a diff in the pull request:
https://github.com/apache/nifi-minifi-cpp/pull/183#discussion_r150864184
--- Diff: CMakeLists.txt ---
@@ -155,6 +155,13 @@ if (ENABLE_USB_CAMERA)
createExtension(USB-CAMERA-EXTENSIONS "USB CAMERA EXTENSIONS" "This
enables USB camera support" "extensions/usb-camera"
"${TEST_DIR}/usb-camera-tests" "TRUE" "thirdparty/libuvc-0.0.6")
endif()
+## TensorFlow extensions
+## Disabled by default because TF can be complex/environment-specific to
build
+option(ENABLE_TENSORFLOW "Disables the TensorFlow extensions." OFF)
+if (ENABLE_TENSORFLOW)
+ createExtension(TENSORFLOW-EXTENSIONS "TENSORFLOW EXTENSIONS" "This
enables TensorFlow support" "extensions/tensorflow"
"${TEST_DIR}/tensorflow-tests")
--- End diff --
This extension should be compatible with that platform. The key thing would
be building TF on (or for, with cross-compiling) the target.
We are dynamically linking TF so it just has to be available in the target
linking environment.
As for build time, our custom FindTensorFlow.cmake needs to find the
libtensorflow_cc.so and headers.
Ultimately, it would be nice to have the ability to build TF in thirdparty/
and link statically, but there's currently no way to static link TF without
giving up important features such as GPU acceleration.
My thought was to keep it as an externally-provided and dynamically-linked
library, and keep an eye on libtensorflow_cc as it matures and hopefully
supports static linking options and cleaner embedded cmake files we could put
in thirdparty/.
> Create processor to apply arbitrary Tensor Flow graphs to tensors
> -----------------------------------------------------------------
>
> Key: MINIFICPP-301
> URL: https://issues.apache.org/jira/browse/MINIFICPP-301
> Project: NiFi MiNiFi C++
> Issue Type: Improvement
> Reporter: Andrew Christianson
> Assignee: Andrew Christianson
>
> In many cases, it may be desirable to interpret/preprocess raw signal inputs
> on the edge, where MiNiFI runs, before sending semantic interpretations
> upstream.
> Tensor Flow is a data flow system for processing tensors, and many graphs
> exist or could be created which are well-suited to interpret signal inputs.
> It would therefore be useful to have a processor in MiNiFi - C++ which takes
> tensors (serialized as binary protocol buffers), feeds them into an input
> node on a supplied graph, reads tensors from an output node, and finally
> writes those output tensors as flow files containing binary protocol buffers.
> While there are many additional convenience utilities which would be helpful,
> such as converting various standard sensor types into tensors, the initial
> scope of this feature is a processor which processes arbitrary tensors
> through arbitrary graphs.
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)