09:24 tomeu:wishes Rockchip's compiler crashed less often...
09:25 phh: like you're giving rknn2 new models to see how it compiles it and fails?
09:26 tomeu: I'm passing a slice of mobiledet, and it segfaults
09:27 phh: wouldn't it be easier to just give it custom models with just the one layer/activation you're currently studying?
09:28 tomeu: yes, that is what I do, but now I need to debug 3 layers together
09:28 tomeu: as the independent layers work fine
09:40 phh: ah okay. well what I said is still relevant, you can still make custom 3 layer model, potentially simpler that wouldn't break rknn?
09:42 phh: but anyway yeah my experience with rknn2 was very bad (I don't quite remember segfaults), but it was blunt failing on most models I could give it
09:42 phh: anyway having a full custom model compiler supporting all NN framework sounds overkilly complicated compared to providing framework' backends
10:02 tomeu: hmm, what do you mean by custom 3 layer model?
10:07 phh: you're using mobiledet, but you could hand craft in tensorflow a simple 3 layers {cnn+relu} that should workk?
10:07 tomeu: ah, I'm not using the whole of it, just a slice
10:08 tomeu: I extract the ops I want from the json of the tflite, with python
10:08 tomeu: has worked fine for VSI's compiler
10:09 tomeu: will check how it is different from similar hand-crafted models when I get back to it
10:14 phh: I wouldn't be surprised that rockchip's tflite parser does broken assumptions that breaks editing json manually
11:14 tomeu: could be, though they bundle tflite2onnx
11:15 tomeu: ah, finally valgrind finished:
11:15 tomeu: ==41405== at 0x93DD64CF: tflite::FlatBufferModel::ReadAllMetadata() const (in /home/tomeu/.local/lib/python3.9/site-packages/tensorflow/lite/python/interpreter_wrapper/_pywrap_tensorflow_interpreter_wrapper.so)
11:16 tomeu: so it is flatbuffers that crashes
11:16 tomeu: amid all this talk of supply chain attacks...
11:17 tomeu: wonder if rockchip supplies their sources to some customers, or if all are running these random forks downloaded from internet forums in their products
11:35 phh: i'm pretty sure they don't
11:45 tomeu: they could least deliver the blobs for their own code, and let people build the rest of the FOSS software that they reuse...
12:02 phh: I presume they added non-trivial changes to tensorflow that are too ugly too publish