TFLu: Port TFL detection postprocess operator

[mansnils]

This is fixing issue: #43565

[gbaned]

@mansnils Can you please resolve conflicts? Thanks!

[gbaned]

@mansnils Any update on this PR? Please. Thanks!

[mansnils]

Removing support for unknown output dimensions. Because it does not work with TFLiteEvalTensor API and it would require major changes in TFLu.
Most models seen with this operator have unknown output dimensions. So to run those they have to be modified first so that the out dimensions are known beforehand.

[advaitjain]

Still having issues unfortunately :-/

The CI is failing with:

tensorflow/lite/micro/kernels/detection_postprocess.cc: In function ‘TfLiteStatus tflite::{anonymous}::Eval(TfLiteContext*, TfLiteNode*)’:
tensorflow/lite/micro/kernels/detection_postprocess.cc:591:58: error: potential null pointer dereference [-Werror=null-dereference]
   tflite::micro::GetTensorData<float>(num_detections)[0] =
   ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~^
       size_of_sorted_indices;
       ~~~~~~~~~~~~~~~~~~~~~~                              
tensorflow/lite/micro/kernels/detection_postprocess.cc:687:58: error: potential null pointer dereference [-Werror=null-dereference]
   tflite::micro::GetTensorData<float>(num_detections)[0] = output_box_index;

but it is not reproducible locally.

My current guess is that the gcc version used as part of CI is old (gcc-6.3.0). I have a change under review that updates

tensorflow/tensorflow/tools/ci_build/Dockerfile.micro

Line 4 in 3d3c4aa

FROM python:3.5-stretch

to

python:3.9.0-buster

which should mean that we use gcc 8.3.0 as part of CI. Let's see what happens when that change get merged.

[mansnils]

@advaitjain It seems like the updated docker didn't help, or do we need to rebase?

[advaitjain]

Still having issues unfortunately :-/

The CI is failing with:

tensorflow/lite/micro/kernels/detection_postprocess.cc: In function ‘TfLiteStatus tflite::{anonymous}::Eval(TfLiteContext*, TfLiteNode*)’:
tensorflow/lite/micro/kernels/detection_postprocess.cc:591:58: error: potential null pointer dereference [-Werror=null-dereference]
   tflite::micro::GetTensorData<float>(num_detections)[0] =
   ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~^
       size_of_sorted_indices;
       ~~~~~~~~~~~~~~~~~~~~~~                              
tensorflow/lite/micro/kernels/detection_postprocess.cc:687:58: error: potential null pointer dereference [-Werror=null-dereference]
   tflite::micro::GetTensorData<float>(num_detections)[0] = output_box_index;

but it is not reproducible locally.

My current guess is that the gcc version used as part of CI is old (gcc-6.3.0). I have a change under review that updates

tensorflow/tensorflow/tools/ci_build/Dockerfile.micro

Line 4 in 3d3c4aa

FROM python:3.5-stretch

to

python:3.9.0-buster

which should mean that we use gcc 8.3.0 as part of CI. Let's see what happens when that change get merged.

I wasn't able to reproduce locally because I hadn't done a make clean_downloads.

Created #45286 to address this issue. This comment has some additional context.