Cannot convert tf to onnx in object detection

[arthurkafer]

Describe the bug
In a simple object detection training, I need to convert the trained model to onnx, but an error saying AttributeError: module 'tensorflow.keras.backend' has no attribute 'get_session' happens;

To Reproduce
To reproduce, I ran a simple person detector training and the converter set to 'onnx'

Expected behavior
Just to convert it so I can use it on my Jetson Nano.

Screenshots

Environment (please complete the following information):

• Using Google Colab right now

Additional context
I saw that there aren't any examples of object detection, but I assume that it would work as well.

This is my config dict:

{
	"model":{
		"type":                 "Detector",
		"architecture":         "MobileNet1_0", # MobileNet7_5
		"input_size":           224,
		"anchors":              [0.57273, 0.677385, 1.87446, 2.06253, 3.33843, 5.47434, 7.88282, 3.52778, 9.77052, 9.16828],
		"labels":               ["bobina"],
		"coord_scale" : 		1.0,
		"class_scale" : 		1.0,
		"object_scale" : 		5.0,
		"no_object_scale" : 	3.0 # 1.0e
	},
	"weights" : {
		"full":   				"mobilenet_1_0_224_tf_no_top.h5",
		"backend":   		  "" # mobilenet_1_0_224_tf_no_top.h5
	},
	"train" : {
		"actual_epoch":         15,
		"train_image_folder":   "bobinas_vert/imgs",
		"train_annot_folder":   "bobinas_vert/anns",
		"train_times":          10,
		"valid_image_folder":   "bobinas_vert/imgs_validation",
		"valid_annot_folder":   "bobinas_vert/anns_validation",
		"valid_times":          5,
		"valid_metric":         "mAP",
		"batch_size":           8,
		"learning_rate":        1e-4,
		"saved_folder":   		"TESTE_ZERO_MEU",
		"first_trainable_layer": "", #conv_pw_13_bn
		"augumentation":		True,
		"is_only_detect" : 		False
	},
	"converter" : {
		"type":   				["onnx"]
	}
}

[AIWintermuteAI]

Hi there!
Yes, the onnx converter option was broken by moving to tf 2.0, too many API changes.
I fixed it now, can you check using dev branch?

If running in colab, you will need to replace the first cell with

#we need imgaug 0.4 for image augmentations to work properly, see https://stackoverflow.com/questions/62580797/in-colab-doing-image-data-augmentation-with-imgaug-is-not-working-as-intended
!pip uninstall -y imgaug && pip uninstall -y albumentations && pip install imgaug==0.4 && pip install tf2onnx
!git clone https://github.com/AIWintermuteAI/aXeleRate.git
!cd aXeleRate && git checkout dev
import sys
sys.path.append('/content/aXeleRate')
from axelerate import setup_training, setup_inference

And then run training as usual. I tested all three types of networks (classifier, detector and segnet) on my local computer - they all worked as expected, outputting .onnx file to project folder. I also tested just detector in Colab

Btw, in your config,

"weights" : {
"full": "mobilenet_1_0_224_tf_no_top.h5",

is not supposed to be used like that. For full weights, you're supposed to pass the pass to full weights, normally it is done for resuming the training.

You're passing "no_top_model, which is needs to go to backend.

[arthurkafer]

Thanks for the fast update!

I forgot to change the weights for this test, and now it worked fine.