DepthAi: Access to NN FP16 output layer causes JVM dump

[ecmnet]

Running a neural network via Java on DepthAI works fine as long as I do not try to access FP16 data.

Example

NNData seg = nn.getNNData(); FloatBuffer output_0 = seg.getLayerFp16("output_0"); float value = output_0.get();

Any access to the data of this Buffer fails and causes a dump of the JVM. The limit and capacity of this buffer seems to be correct
with 10752 bytes (shown as DirectByteBuffer) as I have 2688 FP16 values (2688*4) in the output layer.

Could it be that FP16 -> Java Float is not supported properly?

[saudet]

Could you try to use FloatPointer instead, something like seg.getLayerFp16(new BytePointer("output_0")).get(0)?

[ecmnet]

@saudet Perfect, this seems to work. What's the background here?

[saudet]

JavaCPP doesn't allocate memory for new Buffer objects returned, and with the @StdVector annotation, the memory allocated gets released on return of the function. We could allocate memory with ByteBuffer.allocateDirect(), but then there isn't any way to deallocate that memory. We could also use Java arrays, but... I don't think there's a good way to go about this. Pointer is there for cases like that anyway and it works well enough.

[ecmnet]

Thanks for your explanation. I‘ll close the issue.

[saudet]

I've improved somewhat the mappings for this in commit 0dca099 such that we now get a float[] from getLayerFp16(String). This way we don't have any allocation issues since everything is on the Java heap. Please give it a try with the snapshots: http://bytedeco.org/builds/