def __init__(self, name, n_inputs, n_outputs, activation=None, use_bias=True, weights=None, biases=None):
super().__init__(name)
self.n_inputs = n_inputs
self.n_outputs = n_outputs
self.use_bias = use_bias
if activation is None:
activation = Activation.getInitialized("tanh")
else:
if not Activation.isObjectRegistered(activation):
if isinstance(activation, dict):
activation = Activation(**activation)
elif isinstance(activation, str):
activation = Activation(class_name=activation)
else:
raise Exception("{} is not a "\
"registered activation. Use {}".format(activation, Activation.registeredClasses()))
self.activation = activation
if weights is None:
# Between -1 and 1
self.weights = (np.random.random((n_outputs, n_inputs)) * 2 - 1)
else:
assert isinstance(weights, np.ndarray)
assert weights.shape == (n_outputs, n_inputs)
self.weights = weights
if biases is None:
# Between -1 and 1
self.biases = (np.random.random((n_outputs, 1)) * 2 - 1) * 0.001
else:
assert isinstance(biases, np.ndarray)
assert biases.shape == (n_outputs, 1)
self.biases = biases
# Mutation mask ... create only once.
self.mutation_mask = np.zeros_like(self.weights)
self.setBiases(arr)
else:
raise Exception()
if __name__ == "__main__":
layer = Layer("dense", "input_lay2er", 14, 16, use_bias=False, activation=Activation.getInitialized("relu"))
layer2 = Layer("dense", "input_lay22er", 16, 16, activation=Activation.getInitialized("relu"))
layer3 = Layer("dense", "input_lay222er", 16, 16, activation=Activation.getInitialized("relu"))
layer4 = Layer("dense", "output_layer", 16, 4, activation=Activation.getInitialized("relu"))
model = SequentialModel([layer, layer2, layer3, layer4])
inputs = model.generateRandomInputs(-1, 1)
np.savez_compressed(r"C:\tmp\sequence_test.npz", state=model.__getstate__(), **model.getArrs())
#model2 = model.duplicate(True)
#model3 = model.duplicate(True)
#print (model.compute(inputs), model2.compute(inputs), model3.compute(inputs))
