This is a modified version of braingineer's ikelos cwrnn, update it to suit both keras 2.1.4 version and 2.0.4 version
The main change is that adding a keras version detector for selecting suitable function. E.g.
if ClockworkRNN.k_v > ClockworkRNN.target_v:
self.cell.recurrent_kernel = self.cell.recurrent_kernel * self.mask
else:
self.recurrent_kernel = self.recurrent_kernel * self.mask
And add a simple help how to use it:
model = Sequential()
model.add(ClockworkRNN(units=90,
period_spec=[1, 2, 4, 8, 16],
input_shape=train_x.shape[1:], # ---(samples, timesteps, dimension)
dropout_W=0.4,
return_sequences=True,
debug=cwrnn_debug)) # debug is for developing mode, you can remove
model.add(Dropout(.2))
model.add(TimeDistributed(Dense(units=1, activation='linear')))
model.compile(loss='mse', optimizer='sgd', metrics=['accuracy'])
model.fit(train_x, train_y, epochs=epochs, batch_size=1, verbose=1)
The test code has been run at Windows 7 with Anaconda 2, Keras 2.1.4, Theano. And it also has been run at Cent OS 7 with Anaconda 2, Keras 2.0.4, Theano.
Predict Result:
Training and Test curve for epochs from 1 to 100:
