def _init_pars(self):
spec = rnn.parameters(
self.n_inpt, self.n_hiddens, self.n_output, self.skip_to_out,
self.hidden_transfers)
self.parameters = ParameterSet(**spec)
self.parameters.data[:] = np.random.standard_normal(
self.parameters.data.shape).astype(theano.config.floatX)
def parameters(n_inpt,
n_hiddens,
n_output,
skip_to_out=False,
hidden_transfers=None,
out_transfer=None,
prefix=''):
spec = rnn.parameters(n_inpt, n_hiddens, n_output, skip_to_out,
hidden_transfers, out_transfer, prefix)
if hidden_transfers is not None:
hiddens_inoutsizes = [rnn.inout_size(i) for i in hidden_transfers]
else:
hiddens_inoutsizes = [(1, 1) for _ in n_hiddens]
hiddens_insizes, hiddens_outsizes = zip(*hiddens_inoutsizes)
total_hidden_outsizes = [
i * j for i, j in zip(n_hiddens, hiddens_outsizes)
]
for i, j in enumerate(total_hidden_outsizes):
spec['initial_hidden_means_%i' % i] = j
spec['initial_hidden_vars_%i' % i] = j
del spec['initial_hiddens_%i' % i]
return spec
def parameters(n_inpt, n_hiddens, n_output, skip_to_out=False,
hidden_transfers=None, out_transfer=None, prefix=''):
spec = rnn.parameters(n_inpt, n_hiddens, n_output, skip_to_out,
hidden_transfers, out_transfer, prefix)
if hidden_transfers is not None:
hiddens_inoutsizes = [rnn.inout_size(i) for i in hidden_transfers]
else:
hiddens_inoutsizes = [(1, 1) for _ in n_hiddens]
hiddens_insizes, hiddens_outsizes = zip(*hiddens_inoutsizes)
total_hidden_outsizes = [i * j for i, j in zip(n_hiddens, hiddens_outsizes)]
for i, j in enumerate(total_hidden_outsizes):
spec['initial_hidden_means_%i' % i] = j
spec['initial_hidden_vars_%i' % i] = j
del spec['initial_hiddens_%i' % i]
return spec
def _gen_par_spec(self):
"""Return the parameter specification of the generating model."""
n_output = self.assumptions.visible_layer_size(self.n_inpt)
return rnn.parameters(
self.n_latent + self.n_inpt, self.n_hiddens_gen,
n_output)
def _init_pars(self):
spec = rnn.parameters(self.n_inpt, self.n_hiddens, self.n_output,
self.skip_to_out, self.hidden_transfers)
self.parameters = ParameterSet(**spec)
self.parameters.data[:] = np.random.standard_normal(
self.parameters.data.shape).astype(theano.config.floatX)
