[ConvNext-Det] Quantized ConvNext

hadhoryth commented 6 months ago

Issue Type

Documentation Feature Request

OS

Linux

onnx2tf version number

1.21.1

onnx version number

1.15.0

onnxruntime version number

1.17.1

onnxsim (onnx_simplifier) version number

0.4.33

tensorflow version number

2.16.1

Download URL for ONNX

convnext-det.onnx.zip

Parameter Replacement JSON

--

Description

I have a convnext model and want to generate quantized version of it.

Command what I'm using: onnx2tf -i convnext-det.onnx -cotof -rtpo Gelu -n -coion -osd

The conversion to the the tflite_32 and tflite_16 worked but the .pb files are not there so quantized model is not generated.

WARNING: All log messages before absl::InitializeLog() is called are written to STDERR
W0000 00:00:1715702204.803935 1208310 tf_tfl_flatbuffer_helpers.cc:390] Ignored output_format.
W0000 00:00:1715702204.803969 1208310 tf_tfl_flatbuffer_helpers.cc:393] Ignored drop_control_dependency.
W0000 00:00:1715702205.390732 1208310 tf_tfl_flatbuffer_helpers.cc:390] Ignored output_format.
W0000 00:00:1715702205.390751 1208310 tf_tfl_flatbuffer_helpers.cc:393] Ignored drop_control_dependency.

Maybe you can suggest what I'm doing wrong.

hadhoryth commented 6 months ago

My model has 6 channels as inputs and I also see you have the issue here https://github.com/PINTO0309/onnx2tf/issues/411 So I think it is related

PINTO0309 commented 6 months ago

Share the link to the onnx. It is very tedious to generate it myself.

hadhoryth commented 6 months ago

Sure, here is the link to gdrive

convnext-det.onnx.zip

PINTO0309 commented 6 months ago

Fixed a problem in the conversion process of GatherElements and upgraded to the new version.

Update: https://github.com/PINTO0309/onnx2tf/releases/tag/1.21.2

pip install onnx2tf -U

For now, please use the following tutorial for quantization for tensors other than 3-channel (RGB). This is because it is impossible for onnx2tf to determine whether the input data type is image, voice, or other sensor data from the ONNX input shape alone. The calibration data for INT8 quantization must be some data used for training, but the tool does not know what that data is.

Even if ConvNext-Det's input were 6 channels of RGBRGB, it would not be known from onnx2tf.

hadhoryth commented 6 months ago

@PINTO0309 Thank you for the incredibly fast update to the codebase. The conversion went through, but I still can not see the .pb files. Does it work for you?

Name: onnx2tf
Version: 1.21.2

I want to start with this: https://github.com/PINTO0309/onnx2tf?tab=readme-ov-file#9-int8-quantization-of-models-with-multiple-inputs-requiring-non-image-data

So I run onnx2tf to generate the pb file

onnx2tf -i convnext_det.onnx  -cotof -n -osd -rtpo Gelu

but my saved_model does not have it, only (convnext_det_float16.tflite, convnext_det_float32.tflite)

PINTO0309 commented 6 months ago

You have missed the warning message at the beginning of the model's conversion log.

WARNING: This model contains GroupConvolution and is automatically optimized for TFLite, but is not output because saved_model does not support GroupConvolution. If saved_model is needed, specify --disable_group_convolution to retransform the model.

onnx2tf -i convnext-det.onnx  -cotof -osd -dgc

ls -l saved_model

drwxr-xr-x 2 xxxxx xxxxx  1048576  5月 15 18:10 assets
-rwxr-xr-x 1 xxxxx xxxxx  8094560  5月 15 18:10 convnext-det_float16.tflite
-rwxr-xr-x 1 xxxxx xxxxx 16072900  5月 15 18:10 convnext-det_float32.tflite
-rwxr-xr-x 1 xxxxx xxxxx       54  5月 15 18:10 fingerprint.pb
-rwxr-xr-x 1 xxxxx xxxxx 16861549  5月 15 18:10 saved_model.pb
drwxr-xr-x 2 xxxxx xxxxx  1048576  5月 15 18:10 variables

hadhoryth commented 6 months ago

You are right! It worked for me! Thank you!

PINTO0309 commented 4 months ago

Fix: onnx2tf>=1.24.0

# Keras API 3, Convertible without `-dgc`.
pip install -U tf-keras~=2.16
pip install -U tensorflow>=2.17.0
pip install -U onnx2tf

onnx2tf -i convnext-det.onnx -cotof -osd