[YOLOv5] Yolov5 shows zero accuracy after quantization

[Corallo]

Issue Type

Others

onnx2tf version number

1.9.6

onnx version number

1.13.0

tensorflow version number

2.11.0

Download URL for ONNX

https://mega.nz/folder/PBshwDga#KwqmDZScWLIhdqOb9sqgzg

Parameter Replacement JSON

I don't know where to get this JSON

Description

1. I am trying to develop a guide about deploying a model on an edge device that needs model quantized by tensor. The quantization is the main blocking part.
2. No errors, I am using a custom yolov5 (the custom part is simply the relu activation), exported to onnx, and then converted to tflite with onnx2tf, with the command onnx2tf -i ./fully_trained_normal.onnx -oiqt -qt per-tensor -ioqd uint8 then I am running the yolov5 eval script python3 val.py --weights fully_trained_normal_integer_quant.tflite . While running the eval script on the onnx model produces a map score of ~0.44, I get 0 with the quantized model. I was wondering if you know I am doing something obviously wrong that I am not aware of.
3. Checked the quantized model on netron, it looks all right, I tried with some of the other quantized version of the model that have both filter and bias of the convolution quantized as int8, same problem apply, I tried also to export to both int8 and uint8. Checked on this repo closed issues, didn't find similar problems.
4. To obtain a working model for the edge device
5. I linked a mega folder with the onnx model and the tflite converted version.

Any suggestion would be appreciated :)

[PINTO0309]

Duplicate of #244 , #263, #269

[Corallo]

I see, looking at those other issues, I understand it is an architecture problem that doesn't support well by design the quantization by tensor.
Do you have a suggestion about object detection model that don't have this issue with quantization by tensor?
I've seen your model zoo, but there I don't see any indication of which model might have this issue