Cannot convert tf to onnx in object detection

arthurkafer commented 3 years ago

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 commented 3 years ago

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 Screenshot from 2021-04-20 22-29-34

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 commented 3 years ago

Thanks for the fast update!

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

AIWintermuteAI / aXeleRate

Cannot convert tf to onnx in object detection #46