cassdlcm/mediapipe
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

enable NNAPI in CPU path

Browse Source 
With NNAPI delegate added, Android devices with floating point
NNAPI device(s) could run facedectioncpu and objectdectioncpu
accelerated. On device without accelerators, tflite interpreter
will simple fall back to CPU.
This commit is contained in:
Koan-Sin Tan 2019-11-08 15:39:23 +08:00
parent fce372d153
commit aea83912f9
1 changed files with 4 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
mediapipe/calculators/tflite/tflite_inference_calculator.cc
View File

@ -305,6 +305,10 @@ REGISTER_CALCULATOR(TfLiteInferenceCalculator);
#endif #endif
} else { } else {
// Read CPU input into tensors. // Read CPU input into tensors.
interpreter_->SetAllowFp16PrecisionForFp32(1);
delegate_ = tflite::NnApiDelegate();
interpreter_->ModifyGraphWithDelegate(delegate_);
const auto& input_tensors = const auto& input_tensors =
cc->Inputs().Tag("TENSORS").Get<std::vector<TfLiteTensor>>(); cc->Inputs().Tag("TENSORS").Get<std::vector<TfLiteTensor>>();
RET_CHECK_GT(input_tensors.size(), 0); RET_CHECK_GT(input_tensors.size(), 0);