def weight_relu_initialization(link, mean=0.0, relu_a=0.0, way='forward'):
dim = len(link.weight.data.shape)
if dim == 2:
# fc layer
channel_out, channel_in = link.weight.data.shape
y_k, x_k = 1, 1
elif dim == 4:
# conv layer
channel_out, channel_in, y_k, x_k = link.weight.data.shape
n_i, n_i_next = NN.select_way(way, channel_in * y_k * x_k, channel_out * y_k * x_k)
# calculate variance
variance = initializer.variance_relu(n_i, n_i_next, a=relu_a)
# orthogonal matrix
w = []
for i in six.moves.range(channel_out):
w.append(initializer.orthonorm(mean, variance, (channel_in, y_k * x_k), initializer.gauss, np.float32))
return np.reshape(w, link.weight.data.shape)
def bias_initialization(conv, constant=0):
return initializer.const(conv.b.data.shape,
constant=0,
dtype=np.float32)