Closed xsacha closed 3 years ago
I've added a test and converted a few models using this and confirmed it works.
Big warning to anyone planning on using it: Tensorflow Lite still does not support grouped convolutions during quantization. So this cannot run on the camera, unfortunately. It gives a strange error about the filter size being wrong but it's because it ignores the group number. The Tensorflow Lite bug is tracked here: https://github.com/tensorflow/tensorflow/issues/40044
This fails for tensorflow==2.3.1 as it does not support grouped convolutions on the CPU, but I suppose it's time to bump the minimal required tensorflow version...
Thanx! Please consider adding a test in test_onnx2keras.py aswell.