Tensorflow Lite BBAI-64 Custom TFLite Image Classification Model compilation for the libtidl_tfl_delegate

General Discussion
1

So this is my progress so far:
After poking around with tidl-tools for a while I’ve managed to strip off the major husk from the tflrt_delegate.py and made my own compile.py script.
With it I’ve managed to build the artifacts required by libtidl_tfl_delegate.so on Ubuntu PC.
Right now the Custom Image Classification model based on Keras MobileNetV2 can be trained with train.py script.
Proven to work on bbai64-emmc-flasher-debian-11.8-xfce-edgeai-arm64-2023-10-07-10gb.img.xz image.

1 Like
2

Have managed to train, compile and run the ssd_mobilenet_v2_320x320_coco17_tpu-8 object detection model using libtidl_tfl_delegate.so on bbai-64.
The inference time is about 7ms for 300x300 input.

1 Like