question : what is the previous step to quantitize nanodet model ?

[wwdok]

Hi, @PINTO0309 , thank you for providing these great samples, now i am interested in using nanodet tflite model, you have provided the convertion python scripts, but i have a question, what is the file used to export tflite in saved_model_320x320 ? is it model_float32.pb ? if so, how to generate it? Could you please give me some instructions ? Many thanks !

[PINTO0309]

If I try to explain it properly, it will take a long time. Please wait a moment while I prepare it.

[wwdok]

Oh, sorry for this will take you a long time, let me narrow down the scope of the the question, the model_float32.pb comes from format convertion or transfer retraining ? You can just explain the key points, i will spend time to dive into it, thanks !

[PINTO0309]

Environment

1. PyTorch
2. ONNX Runtime
3. OpenVINO 2021.1+
4. tf-nightly

Conversion Workflow

PyTorch -> ONNX -> OpenVINO IR -> openvino2tensorflow -> TFLite

ONNX -> OpenVINO IR

$ python3 ${INTEL_OPENVINO_DIR}/deployment_tools/model_optimizer/mo.py \
 --input_model nanodet_320x320.onnx \
 --input i \
 --input_shape [1,3,320,320] \
 --output_dir openvino/nanodet_320x320/FP32 \
 --data_type FP32

Three types of files will be generated: .xml, .bin, and .mapping.

Edit and optimize the .xml to remove layers of meaningless garbage

1. Before

2. After

OpenVINO IR -> TFLite, saved_model, pb, quantized tflite, h5

https://github.com/PINTO0309/openvino2tensorflow.git

$ openvino2tensorflow \
  --model_path onnx/openvino/nanodet_320x320/FP32/nanodet_320x320.xml \
  --output_saved_model True \
  --output_h5 True \
  --output_pb True \
  --output_no_quant_float32_tflite True \
  --output_weight_quant_tflite True \
  --output_float16_quant_tflite True

[PINTO0309]

@wwdok

the model_float32.pb comes from format convertion or transfer retraining ?

format convertion

--output_pb True

[wwdok]

Wow, the convertion process is longer than i thought, i am really thankful for this helpful guide !