I have quantized a detector model: Yolo X s of super-gradients repo using these commands:
snpe-dlc-quantize --input_dlc ../model640x480.dlc --output_dlc ../model640x480_quantized.dlc --input_list dataset/cropped640x480_snpe_raw/samples2.txt
Also, tried it via AIMET:
snpe-onnx-to-dlc --input_network ../aimet_model640x480.onnx --output_path ../aimet_model640x480.dlc --quantization_overrides aimet_model640x480.encodings
snpe-dlc-quantize --input_dlc ../aimet_model640x480.dlc --output_dlc ../aimet_model640x480_quant.dlc --input_list dataset/cropped640x480_snpe_raw/samples2.txt --override_params --enable_htp
snpe-dlc-graph-prepare --input_dlc ../aimet_model640x480_quant.dlc --output_dlc ../aimet_model640x480_quant_cache.dlc --htp_archs v68
The result in both ways of converting it, result in the same outcome:
- CPU predicts correctly
- DSP predicts all confidences randomly (e.g. objectness score is always ZERO)
Please consider that both CPU and DSP inferences use the same QUANTIZED DLC MODEL FILE.
Does DSP-infer read something else from the quantized-model-file than the CPU-infer ?
Dear developer,
What's the version of snpe you used, and you can share you model with us through git.
BR.
Yunxiang