retinanet.py

####################### retinanet.py
def default_null_model(
    pyramid_feature_size=256,
    name='null_submodel'
):
    options = {
        'kernel_size' : 3,
        'strides'     : 1,
        'padding'     : 'same',
    }

    inputs  = keras.layers.Input(shape=(None, None, pyramid_feature_size))
    outputs = inputs
    
    return keras.models.Model(inputs=inputs, outputs=outputs, name=name)
  
def default_null_submodel():
    return [('null', default_null_model())]

def retinanet_null(
    inputs,
    backbone,
    create_pyramid_features = __create_pyramid_features,
    submodels               = None,
    name                    = 'retinanet'
):
    if submodels is None:
        submodels = default_null_submodel()

    image = inputs

    _, C3, C4, C5 = backbone.outputs  # we ignore C2

    # compute pyramid features as per https://arxiv.org/abs/1708.02002
    features = create_pyramid_features(C3, C4, C5)

    # for all pyramid levels, run available submodels
    pyramid = __build_pyramid(submodels, features)

    return keras.models.Model(inputs=inputs, outputs=[pyramid], name=name)

def retinanet_null_bbox(inputs, name='retinanet-null-bbox', *args, **kwargs):
    model = retinanet_null(inputs=inputs, *args, **kwargs)

    # we expect the anchors, regression and classification values as first output
    regression     = model.outputs[0]

    # construct the model
    return keras.models.Model(inputs=inputs, outputs=[regression], name=name)
####################### resnet.py
def ResNet50RetinaNetNull(inputs, weights='imagenet', *args, **kwargs):
    image = inputs

    # load pretrained imagenet weights?
    if weights == 'imagenet':
        weights_path = keras.applications.imagenet_utils.get_file(
            'ResNet-50-model.keras.h5',
            WEIGHTS_PATH_NO_TOP, cache_subdir='models', md5_hash='1e511c75e9ab5c16900652ad1f6044ce'
        )
    else:
        weights_path = weights

    resnet = keras_resnet.models.ResNet50(image, include_top=False, freeze_bn=True)

    model = keras_retinanet.models.retinanet.retinanet_null_bbox(inputs=inputs, backbone=resnet, *args, **kwargs)
    model.load_weights(weights_path, by_name=True)
    return model

def create_model(weights='imagenet'):
    image = keras.layers.Input((640, 640, 3))
    return ResNet50RetinaNetNull(image, weights=weights)
  
# ValueError: `Concatenate` layer requires inputs with matching shapes except for the concat axis. 
# Got inputs shapes: [(None, 80, 80, 256), (None, 40, 40, 256), (None, 20, 20, 256), (None, 10, 10, 256), (None, 5, 5, 256)]