def __init__(self, H = 3, d = 1, ny = 1, M = 3, debug_output = False):
"""
H: number of hidden units
d: number of inputs
ny: number of outputs
M: number of mixture components
"""
self.c = ny
ny = 2*M + M * ny # M mixing coefficients + M variances + M*ny means
self.M = M
self.count_fwd = 0
TLP.__init__(self, H, d, ny, linear_output = True, error_function = 'mdn',
debug_output = debug_output)
def __init__(self, H = 3, d = 1, ny = 1, T = 100):
"""
Create a fully-connected neural network with one hidden recurrent layer .
@param H: number of hidden units
@param ny: number of output units
@param T: number of time-steps
"""
TLP.__init__(self, H, d, ny)
self.T = T
z = np.zeros([T, H+1]) # hidden unit activations + bias
self.z = z[None, :]
self.z[:,:,0] = 1 # set extra bias unit to one
dj = np.zeros([T, H+1])
self.dj = dj[None, :]
# init recurrent weights
self.wh = np.random.normal(loc=0.0, scale = 1,size=[H, H]) / np.sqrt(H) # TODO: check?
self.Nwh = H**2
def _forward(self, x, w1 = None, w2 = None): # perform forward propagation #t0=datetime.now() y = TLP._forward(self, x, w1, w2) #self.count_fwd = self.count_fwd + 1 # the outputs are ordered as follows: # y[0:M]: M mixing coefficients alpha # y[M:2*M]: M kernel widths sigma # y[2*M:M*c]: M x c kernel centre components M = int(self.M) # calculate mixing coefficients, variances and means from network output y[0:M] = self.softmax(y[0:M]) # alpha # avoid overflow y[M:2*M] = np.minimum(y[M:2*M], np.log(np.finfo(float).max)) y[M:2*M] = np.exp(y[M:2*M]) # sigma # avoid underflow y[M:2*M] = np.maximum(y[M:2*M], np.finfo(float).eps) #print (datetime.now()-t0).microseconds return y
def __init__(self, H = 3, d = 1, ny = 1): TLP.__init__(self, H, d, ny, linear_output = True, error_function = 'bayes')