ZFTurbo/Keras-inference-time-optimizer

Error with retinanet model

Opened this issue · 4 comments

Maxfashko commented

Hi, @ZFTurbo!
I tryed convert retinanet model with backbone resnet50 and it faled.
I used https://github.com/fizyr/keras-retinanet.

  File "/usr/local/lib/python3.6/dist-packages/kito/__init__.py", line 330, in reduce_keras_model
    new_layer = clone_model(layer)
  File "/usr/local/lib/python3.6/dist-packages/keras/models.py", line 251, in clone_model
    return _clone_functional_model(model, input_tensors=input_tensors)
  File "/usr/local/lib/python3.6/dist-packages/keras/models.py", line 106, in _clone_functional_model
    new_layer = layer.__class__.from_config(layer.get_config())
  File "/usr/local/lib/python3.6/dist-packages/keras/engine/base_layer.py", line 1109, in from_config
    return cls(**config)
  File "/usr/local/lib/python3.6/dist-packages/keras/legacy/interfaces.py", line 91, in wrapper
    return func(*args, **kwargs)
  File "/usr/local/lib/python3.6/dist-packages/keras/layers/convolutional.py", line 490, in __init__
    **kwargs)
  File "/usr/local/lib/python3.6/dist-packages/keras/layers/convolutional.py", line 118, in __init__
    self.bias_initializer = initializers.get(bias_initializer)
  File "/usr/local/lib/python3.6/dist-packages/keras/initializers.py", line 508, in get
    return deserialize(identifier)
  File "/usr/local/lib/python3.6/dist-packages/keras/initializers.py", line 503, in deserialize
    printable_module_name='initializer')
  File "/usr/local/lib/python3.6/dist-packages/keras/utils/generic_utils.py", line 138, in deserialize_keras_object
    ': ' + class_name)
ValueError: Unknown initializer: PriorProbability

I would be grateful for any help.

tonmoyborah commented

Hi @ZFTurbo , great work here. I can see you have contributed in the keras-retinanet repo of fizyr too and also tried to support that here. Still facing the same issue as above. Any workaround for this?

ZFTurbo commented

@tonmoyborah

Current code supports retinanet. Here is the example. I will add it in test_bench later:

def get_RetinaNet_model():
    from keras.models import load_model
    from keras.utils import custom_object_scope
    from keras_resnet.layers import BatchNormalization
    from keras_retinanet.layers import UpsampleLike, Anchors, RegressBoxes, ClipBoxes, FilterDetections
    from keras_retinanet.initializers import PriorProbability

    custom_objects = {
        'BatchNormalization': BatchNormalization,
        'UpsampleLike': UpsampleLike,
        'Anchors': Anchors,
        'RegressBoxes': RegressBoxes,
        'PriorProbability': PriorProbability,
        'ClipBoxes': ClipBoxes,
        'FilterDetections': FilterDetections,
    }

    with custom_object_scope(custom_objects):
        model = load_model("../retinanet_resnet50_500_classes_0.4594_converted.h5")
    return model, custom_objects

from keras.utils import custom_object_scope
model, custom_objects = get_RetinaNet_model()
with custom_object_scope(custom_objects):
    model_reduced = reduce_keras_model(model)
  • 👍1
tonmoyborah commented

I'll try this. I made it work by changing deserealize function in keras code but the resulting model didn't provide any speedup. Models other than retinanet are showing huge improvements

ZFTurbo commented

I observed the same behaviour for RetinaNet. While many BN layers were removed speed of inference stays the same.

P.S. Added RetinaNet in test_bench.

  • 🎉2