I got error

[PythonImageDeveloper]

I useed EfficientNet as backnone of ssd, When I run python object_detection/builders/model_builder_test.py for testing the API, I got this error:

Traceback (most recent call last):
  File "object_detection/builders/model_builder_test.py", line 203, in test_create_ssd_models_from_config
    model = model_builder.build(model_proto, is_training=True)
  File "/home/mm/API-TF/models/research/object_detection/builders/model_builder.py", line 125, in build
    return _build_ssd_model(model_config.ssd, is_training, add_summaries)
  File "/home/mm/API-TF/models/research/object_detection/builders/model_builder.py", line 248, in _build_ssd_model
    is_training=is_training)
  File "/home/mm/API-TF/models/research/object_detection/builders/model_builder.py", line 224, in _build_ssd_feature_extractor
    return feature_extractor_class(**kwargs)
TypeError: __init__() got an unexpected keyword argument 'conv_hyperparams_fn'

----------------------------------------------------------------------
Ran 16 tests in 0.021s

FAILED (errors=1, skipped=1)

[BowieHsu]

put this FeatureExtractor into keras CONFIG like mobilenet

[PythonImageDeveloper]

Hi @BowieHsu , Did you doing this work? is't works? If so, Please more explain about this, if possible, thanks I don't understand about this, FeatureExtractor into keras CONFIG like mobilenet

[BowieHsu]

@PythonImageDeveloper kind of like this.

[PythonImageDeveloper]

Hi @BowieHsu @CasiaFan I now got this error when I run python model_main.py in the models/research_object_detection :

    from object_detection.builders import model_builder
  File "/home/mm/API-TF/models/research/object_detection/builders/model_builder.py", line 56, in <module>
    from .efficientnet_feature_extractor import SSDEfficientNetFeatureExtractor, SSDEfficientNetFPNFeatureExtractor
ImportError: No module named 'object_detection.builders.efficientnet_feature_extractor'

[CasiaFan]

@PythonImageDeveloper Sorry for forgetting to add location for efficientnet files. I've updated it in README. These two file should better to put under same directory. In your case, it seems that you put efficient_feature_extractor under builders but efficientnet in other place.

[PythonImageDeveloper]

hi @BowieHsu @CasiaFan I solved this problem, but I have some question: How do I create config file for ssd_effcientnet? Can I modify the ssd_mobilnetv2 confilg file for ssd_efficientnet? If so,Do I must change the type of feature extractor from ssd_mobilenetv2 to ssd_efficientnet, right? But I don't know what put in the section of fine_tune_checkpoint path for this new network?

[CasiaFan]

@PythonImageDeveloper See this issue for sample config. If you'd like to use finetune_checkpoint, you could train EfficientNet on VOC dataset and use its checkpoint to finetune. But in case, since the pretrained checkpoint is not provided, just comment the line fine_tune_checkpoint

[PythonImageDeveloper]

Hi @CasiaFan I have my own dataset that is one class and have 10k images and I also have VOC and COCO datasets, In your opinion, what's options you select for training: 1- Train only 10k my own images 2- Frist train on COCO/VOC dataset and then transfer-learning over 10k my own dataset? Ok this option is good, but If I have one GPU 1080 ti, then can I train with one GPU for training over COCO dataset?

[CasiaFan]

@PythonImageDeveloper In any case, training for enough epochs is the best way. For option 1, add more augmentation methods like mixing up and random erasing, whereas overfitting is somehow inevitable; for the 2nd option, using small batch inputs and group normalization should improve, but a long time is required. So if you are in hurry, use the first method.

[CasiaFan]

@PythonImageDeveloper In any case, training for enough epochs is the best way. For option 1, add more augmentation methods like mixing up and random erasing, whereas overfitting is somehow inevitable; for the 2nd option, using small batch inputs and group normalization should improve, but a long time is required. So if you are in hurry, use the first method.