retraining-tappas-models.rst

Retraining TAPPAS Models

Instructions

If you wish to use a TAPPAS pipeline/demo with your own model, the easiest solution is:

1. Read the relevant TAPPAS pipeline/demo README page:

   • It lists the ModelZoo models that are used on this pipeline
   • If Model Zoo retraining dockers are available for those models, links are given; In addition it describes how to reconfigure TAPPAS for the retrained models

2. Follow the retraining docker's instructions to retrain the model with your own data
3. Follow the retraining docker's instructions to compile the network
4. Reconfigure TAPPAS scripts to run your new .hef file(s)

Notes

• Easy post-processing JSON reconfiguration is currently only available for YOLO architecures. For other architectures, recreating post-processing .so file is required. The relevant code and header files are mentioned on the README pages.

• Models that use on-chip RGBX->RGB layers does not appear yet on the ModelZoo. Therefore, many models that are used for iMX demos (and others; see on the README of each app) should be added those layers manually to create models that will fit TAPPAS:

  • Use the non-rgbx network from the above table for retraining and compilation, with one simple modification - On the model alls file (on hailo_model_zoo/cfg/alls), add right after normalization command:

    reshape_rgb = input_conversion(input_layer1, tf_rgbx_to_hailo_rgb)