def _reorg_lstm_parameters(ih, hh):
hidden_size = len(ih) // 4
return L.LSTMParameterSet(
ih[0 * hidden_size:1 * hidden_size], # W_input
hh[0 * hidden_size:1 * hidden_size], # U_input
ih[2 * hidden_size:3 * hidden_size], # W_cell
hh[2 * hidden_size:3 * hidden_size], # U_cell
ih[1 * hidden_size:2 * hidden_size], # W_forget
hh[1 * hidden_size:2 * hidden_size], # U_forget
ih[3 * hidden_size:4 * hidden_size], # W_output
hh[3 * hidden_size:4 * hidden_size]) # U_output
def _make_bias(self, params):
bias = []
for i in range(params.num_layers):
bias.append(
L.LSTMParameterSet(
np.random.rand(params.hidden_size).astype(np.float32),
np.random.rand(params.hidden_size).astype(np.float32),
np.random.rand(params.hidden_size).astype(np.float32),
np.random.rand(params.hidden_size).astype(np.float32),
np.random.rand(params.hidden_size).astype(np.float32),
np.random.rand(params.hidden_size).astype(np.float32),
np.random.rand(params.hidden_size).astype(np.float32),
np.random.rand(params.hidden_size).astype(np.float32)))
return bias
def _make_weights(self, params, skip=False):
weights = []
for i in range(params.num_layers):
w_out = params.hidden_size
w_in = params.hidden_size if i > 0 or skip else params.data_size
weights.append(
L.LSTMParameterSet(
np.random.rand(w_out, w_in).astype(np.float32),
np.random.rand(w_out, w_out).astype(np.float32),
np.random.rand(w_out, w_in).astype(np.float32),
np.random.rand(w_out, w_out).astype(np.float32),
np.random.rand(w_out, w_in).astype(np.float32),
np.random.rand(w_out, w_out).astype(np.float32),
np.random.rand(w_out, w_in).astype(np.float32),
np.random.rand(w_out, w_out).astype(np.float32)))
return weights