unable to replicate yolov4 onnx conversion example

[chad-green]

Bug Report

Which model does this pertain to?

yolov4

Describe the bug

Attempted to replicate results using the steps described in Conversion.ipynb The result is a much different graph that gives errors when running with the inference.ipynb It only generates one output of shape: Output shape: [(1, 33, 84)]

Whereas your onnx model has output shape: Output shape: [(1, 52, 52, 3, 85), (1, 26, 26, 3, 85), (1, 13, 13, 3, 85)] How did you change the output shape?

Reproduction instructions

System Information

OS Platform and Distribution (Linux Ubuntu 18.04):
ONNX version (1.7.0):
Backend/Runtime version (ONNX Runtime 1.5.2):

Provide a code snippet to reproduce your errors. run the suggested tutorials as written: https://github.com/onnx/models/blob/master/vision/object_detection_segmentation/yolov4/dependencies/Conversion.ipynb https://github.com/onnx/models/blob/master/vision/object_detection_segmentation/yolov4/dependencies/inference.ipynb

[AnakTeka]

I'm having a same problem. It seems that the google drive that was linked had different file that the one that used for the conversion?

[hoangkhoiLE]

I have the same question !!!

[hoangkhoiLE]

@ravimashru Please help us !

[ravimashru]

@hoangkhoiLE I worked off the model weights already available here: https://github.com/onnx/models/tree/master/vision/object_detection_segmentation/yolov4/model

You could maybe take a look at what changed in the repository https://github.com/hunglc007/tensorflow-yolov4-tflite.

[hoangkhoiLE]

@ravimashru Thank you for your reponse !

At first, when i used the model in https://github.com/onnx/models/tree/master/vision/object_detection_segmentation/yolov4/model, it gives me an error:

RuntimeError: /onnxruntime_src/onnxruntime/core/session/inference_session.cc:215 onnxruntime::InferenceSession::InferenceSession(const onnxruntime::SessionOptions&, const string&, onnxruntime::logging::LoggingManager*) status.IsOK() was false. Given model could not be parsed while creating inference session. Error message: Protobuf parsing failed.

when i just tried to load the model: ''' sess = rt.InferenceSession("yolov4.onnx")

outputs = sess.get_outputs() output_names = list(map(lambda output: output.name, outputs)) input_name = sess.get_inputs()[0].name '''

I have tried some different versions onnx but the error is still exists.

So i continued to try with jupyter notebook Conversion: https://github.com/onnx/models/blob/master/vision/object_detection_segmentation/yolov4/dependencies/Conversion.ipynb

I have verified well the link of yolov4 weights, they keeps the same link with the same id. So the change in the link https://github.com/hunglc007/tensorflow-yolov4-tflite may not cause impact.

Just by curiosity, can you show us how you load model in : https://github.com/onnx/models/tree/master/vision/object_detection_segmentation/yolov4/model and which env python and its dependencies that work well !

Thank you,

Best regards

[ravimashru]

I use it exactly like you: sess = rt.InferenceSession("yolov4.onnx")

The versions that are currently working for me:

numpy==1.19.4
onnx==1.7.0
onnxruntime==1.5.2
protobuf==3.13.0
six==1.15.0
typing-extensions==3.7.4.3

Also, I see that someone had a similar error in https://github.com/onnx/models/issues/306 which turned out to be due to a corrupted yolov4.onnx file.

You could check the MD5 checksum of the file you have and see if it matches mine (2e0eeb4de8da2a0663ae3eb4a0dabbce) to ensure we have the same weights file. :)

[hoangkhoiLE]

I use it exactly like you: sess = rt.InferenceSession("yolov4.onnx")

The versions that are currently working for me:

numpy==1.19.4
onnx==1.7.0
onnxruntime==1.5.2
protobuf==3.13.0
six==1.15.0
typing-extensions==3.7.4.3

Also, I see that someone had a similar error in #306 which turned out to be due to a corrupted yolov4.onnx file.

You could check the MD5 checksum of the file you have and see if it matches mine (2e0eeb4de8da2a0663ae3eb4a0dabbce) to ensure we have the same weights file. :)

@ravimashru Thank you for your kind reponse. I redownload the file and it work right now. Thank you a lot and have a good day.

[vitowen9580]

I have same issue. Anyone solved it?

[s-tory]

This issue was solved by the following operation.

!git clone https://github.com/hunglc007/tensorflow-yolov4-tflite.git
%pushd tensorflow-yolov4-tflite/
!git reset --hard 3f8ed00b73a577e9a3cba8bc559f6064b565dbaa

Maybe, the implementation of TF model's outputs was changed on Jul 2-6 2020. https://github.com/hunglc007/tensorflow-yolov4-tflite/commits/master/save_model.py

---

YOLOv4 -> .onnx conversion commands:

!python save_model.py --weights yolov4.weights --output yolov4.tf --input_size 416 --model yolov4
!python -m tf2onnx.convert --saved-model yolov4.tf --output yolov4_1_416_416_3_converted_by_myself.onnx --opset 11 --verbose

Log w/o !git reset:

2021-02-16 05:52:32,083 - INFO - tf2onnx: inputs: ['input_1:0']
2021-02-16 05:52:32,084 - INFO - tf2onnx: outputs: ['Identity:0']

Log with !git reset:

2021-02-16 05:26:46,820 - INFO - tf2onnx: inputs: ['input_1:0']
2021-02-16 05:26:46,821 - INFO - tf2onnx: outputs: ['Identity:0', 'Identity_1:0', 'Identity_2:0']

---

My package versions:

easydict                  1.9
numpy                     1.19.0
onnx                      1.7.0
onnxruntime               1.6.0
protobuf                  3.12.2
tensorflow                2.2.0
tf2onnx                   1.9.0

[janesser]

Hi,

i could help myself with:

def reshape_output(trt_output):
    if len(trt_output) % (52*52) == 0:
        return trt_output.reshape(1, 52, 52, 3, 85)
    elif len(trt_output) % (26*26) == 0:
        return trt_output.reshape(1, 26, 26, 3, 85)
    elif len(trt_output) % (13*13) == 0:
        return trt_output.reshape(1, 13, 13, 3, 85)
    else:
        print('unknown trt_output size {}'.format(len(trt_output)))
        return []

Still the resulting coordinates are far off from where they should be. Might be rounding issue (tensorrt) or model misuse, i can't tell yet.

Hope above helps, keep me posted if you find a way to make it work out.

Best, Jan