I want to train the model and then I need a int8 onnx model, what should I do step by step?

[xiaoxiongli]

Dear @NJU-Jet :

I feel confused that if I use the command "python train.py --opt options/train/base7.yaml --name base7_D4C28_bs16ps64_lr1e-3 --scale 3 --bs 16 --ps 64 --lr 1e-3 --gpu_ids 0" to train a model, this model is a float32 model or a int8 model?

I want to train the model and then I need a int8 onnx model, what should I do step by step?
do I need to run or modify generate_tflite.py to get a int8 model and then convert to onnx int8 model?
or just the pb model we got from the command "python train.py --opt options/train/base7.yaml --name base7_D4C28_bs16ps64_lr1e-3 --scale 3 --bs 16 --ps 64 --lr 1e-3 --gpu_ids 0" is enough to convert to a int8 onnx model?

Thank you very much!

[NJU-Jet]

Hi, @xiaoxiongli ：
You will get a float32 model (.pb file) by running this command and then you can convert it to onnx model. Please refer:
https://github.com/onnx/tensorflow-onnx#getting-started.

As far as I know, there isn't a direct way to convert .tflite to onnx model. :)

[xiaoxiongli]

Dear @NJU-Jet :

my goal is that to run the int8 model in openvino or mnn or tnn, of cource I can use "https://github.com/onnx/tensorflow-onnx#getting-started" you mentioned before to convert the pb model to onnx model, but this convertion can only got a float32 onnx model, How can I use int8 model in openvino or mnn or tnn?

Can I change the float32 onnx model to int8 model using openvino/mnn/tnn, does this way affect the PSNR?

please help... sorry for disturbs you again, and thank you very much~

[NJU-Jet]

Maybe this works: https://github.com/jackwish/tflite2onnx.
Don't worry, it's my pleasure to help you.

[NJU-Jet]

In my understanding, you can change the float32 onnx model to int8 model using openvino/mnn/tnn and PSNR will drop within 0.1dB, because the quantization algorithm is the same as tflite.

[xiaoxiongli]

it's clear, I got it! it's very kind of you ^_^

[June1124]

Hi@xiaoxiongli, I meet the same point with u, convert INT8 or FP16 tflite model to onnx format, Do you have good advice? Thank u~