Author: Petra Vidnerová, The Czech Academy of Sciences, Institute of Computer Science
RBF layer for Tensorflow 2.0 (as custom layer derived from tf.keras.layers.Layer)
You need rbflayer.py to use RBF layers in your code. See test.py for very simple examples.
Feel free to use or modify the code.
Tensorflow, Scikit-learn, optionally Matplotlib (only for the toy example in test.py)
# creating RBF network
rbflayer = RBFLayer(10,
initializer=InitCentersRandom(X),
betas=2.0,
input_shape=(num_inputs,))
model = Sequential()
model.add(rbflayer)
model.add(Dense(n_outputs, use_bias=False))
or using KMeans clustering for RBF centers
# creating RBFLayer with centers found by KMeans clustering
rbflayer = RBFLayer(10,
initializer=InitCentersKMeans(X),
betas=2.0,
input_shape=(num_inputs,))
If you need any other setup of centers or widhts, you can very easily define your own initializer,
just write your subclass of tensorflow.keras.initializers.Initializer.
Because you have created Keras model with a custom layer, you need to take it into account if you need to save it to file and load it. Saving is no problem:
model.save("some_fency_file_name.h5")
but while loading you have to specify your custom object RBFLayer:
rbfnet = load_model("some_fency_file_name.h5", custom_objects={'RBFLayer': RBFLayer})
You can also load weights (centers or widhts) from file (.npy file with an numpy array of the right shape), see IntFromFile in initializer.py and example in test.py.
Old repo that was written in 2017 for Keras.
If you need help, do not hesitate to contact me via petra@cs.cas.cz or write an Issue.
In case you use this RBF layer for any experiments that result in publication, please consider citing it. Thanks ❤️
Vidnerová, Petra. RBF-Keras: an RBF Layer for Keras Library. 2019. Available at https://github.com/PetraVidnerova/rbf_keras
Thanks to the author of the very first citation: Lukas Brausch, et al. Towards a wearable low-cost ultrasound device for classification of muscle activity and muscle fatigue. 2019 doi:10.1145/3341163.3347749
This work was partially supported by the Czech Grant Agency grant 18-23827S and institutional support of the Institute of Computer Science RVO 67985807.