How to Check Output Object Names Regarding Classes

murkoc · February 17, 2023, 1:09pm

Hi,

I have a tflite model with 14 classes that was created via TFLite modelmaker for object detection. It was created with the default parameter values. TFLite model output shows the 4 outputs but I can not decide which output is which class related. The netron output is below for the end tree:

I wonder what this outputs do and how could I decide my android source scripts parameters regarding these.

Any reply would be appreciated. Thank you in advance.

George_Soloupis · February 17, 2023, 2:29pm

It is kinda confusing since it has not info about the arrays and the dimensions of each class.
It is stating the explnation for each output and only “float32” value.
Are you sure everything has worked properly during the training?
Have you verified that the tflite model is working with the Python API before you go into android?

PINTO0309 · February 18, 2023, 7:25am

When I analyzed tflite’s TFLite_Detection_PostProcess flatbuffers binary two years ago, what I found was that the following four things were the output.

github.com/PINTO0309/onnx2tf

Yolov7-tiny to TensorflowLite conversion results in a dynamic output model incompatible with TfLite Java API

opened 12:56AM - 31 Jan 23 UTC

closed 06:09AM - 03 Feb 23 UTC

ingura

feature request discussion OP:NonMaxSuppression GPU Delegate

### Issue Type Others ### onnx2tf version number 1.5.36 ### onnx version num…ber 1.12.0 ### tensorflow version number 2.10.1 ### Download URL for ONNX pip install onnx==1.12.0 ### Parameter Replacement JSON ```json none ``` ### Description Hi, your library is awesome! I converted the Yolov7-tiny from PyTorch to TfLite using: `onnx2tf -i yolov7-tiny.onnx -o models-NHWC-final/ -osd -oh5 -cotof` I am trying to use it on an android device. The model works when tested on a PC however the Tensorflow Java API for Android does not support dynamic output models according to their documentation: https://www.tensorflow.org/lite/guide/inference. While the resulting yolo tflite model has a dynamic number of outputs( the number of outputs change with the number of indications/detentions) On the other hand, if I follow the conversion path PyTorch -> ONNX-> Tensorflow I do get yolov7 with a fixed output size so I suspect it is possible to achieve this with onnx2tf as well while also doing the NcHW to NHWc conversion in the process. Is there a way to have onnx2tf output a fixed/static output .tflite model for yolov7-tiny? Thank you

bounding_boxes
class_labels
class_confidences
num_of_boxes

However, it is very difficult to know which output OP corresponds to which output name because the output order of tflite output OPs generated from TensorFlow Model Maker and TF OD API cannot be controlled by user operation.

murkoc · February 18, 2023, 11:05pm

Yes it works on Python API and there is no trick for the issue on that side. Actually, it seems ok on training but I am not sure if it worked proper during the training. I have also had training with int32 value but it was still confusing that nothing was exact because model maker has no controllable parameters. Trying to check it on the output of the script, it gives outputs but they are confusing, as well. I was trying each object output and it is not regular for each class. An example one class detected and it gives an output but any other object in the classes provides the same output. That’s what I am trying to solve. I am going to try array size approach which you indicated at your reply. Thank you.

murkoc · February 19, 2023, 9:43am

What I need here is to have coordinates and sizes but another problem here is that my aim is to detect 14 classes objects but a lot of them have the same size. If anyone could recommend some other approaches it’d be great

murkoc · February 21, 2023, 9:12am

Finally I found the problem. I have 14 classes and trying to use it with image classification. But it is saying that image classification on tflite provides one output. That’s what I learnt thank you for all your replies.