Other Parts Discussed in Thread: TDA4VM
The TIDL installation on the target TDA4VM populates the model_zoo. Each model is provided with both the TIDL import artifact files, and the tflite or onnx model that (I suppose) was the input of the import that produced the artifacts. However, when we import those tflite or onnx model, we don't get the same artifacts : while the artifacts on the target always have 100 % of the nodes delegated, the artifacts we obtained by importing the onnx or tflite files always have less than 100 % of the nodes delegated which probably explain the large differences in inference latency.
For example, the ssdlite_mobiledet_dsp_320x320_coco_20200519.tflite model artifacts on the target has 129 / 129 nodes delegated while the import we performed has only 128 / 129 nodes delegated. The "Cast" node is probably not delegated, however when importing the tflite model, we set the compile_options with an empty deny_list so all layers should delegated.
Any reason for not obtaining the same delegated layers ?
