def _build_rnngsn(self):
"""
Creates the updates and other return variables for the computation graph.
Returns
-------
List
the sample at the end of the computation graph, the train cost function, the train monitors,
the computation updates, the generated visible list, the generated computation updates, the ending
recurrent states
"""
# For training, the deterministic recurrence is used to compute all the
# {h_t, 1 <= t <= T} given Xs. Conditional GSNs can then be trained
# in batches using those parameters.
(u, h_ts), updates_train = theano.scan(fn=lambda x_t, u_tm1: self.recurrent_step(x_t, u_tm1),
sequences=self.input,
outputs_info=[self.u0, None],
name="rnngsn_computation_scan")
h_list = [T.zeros_like(self.input)]
for layer, w in enumerate(self.weights_list[:self.layers]):
if layer % 2 != 0:
h_list.append(T.zeros_like(T.dot(h_list[-1], w)))
else:
h_list.append((h_ts.T[(layer/2) * self.hidden_size:(layer/2 + 1) * self.hidden_size]).T)
gsn = GSN(
inputs_hook = (self.input_size, self.input),
input_size = self.input_size,
hiddens_hook = (self.hidden_size, GSN.pack_hiddens(h_list)),
params_hook = self.gsn_params,
outdir = os.path.join(self.outdir, 'gsn_noisy/'),
layers = self.layers,
walkbacks = self.walkbacks,
visible_activation = self.visible_activation_func,
hidden_activation = self.hidden_activation_func,
input_sampling = self.input_sampling,
mrg = self.mrg,
tied_weights = self.tied_weights,
cost_function = self.cost_function,
cost_args = self.cost_args,
add_noise = self.add_noise,
noiseless_h1 = self.noiseless_h1,
hidden_noise = self.hidden_noise,
hidden_noise_level = self.hidden_noise_level,
input_noise = self.input_noise,
input_noise_level = self.input_noise_level,
noise_decay = self.noise_decay,
noise_annealing = self.noise_annealing,
image_width = self.image_width,
image_height = self.image_height
)
cost = gsn.get_train_cost()
monitors = gsn.get_monitors() # frame-level error would be the 'mse' monitor from GSN
x_sample_recon = gsn.get_outputs()
# symbolic loop for sequence generation
(x_ts, u_ts), updates_generate = theano.scan(lambda u_tm1: self.recurrent_step(None, u_tm1),
outputs_info=[None, self.generate_u0],
n_steps=self.n_steps,
name="rnngsn_generate_scan")
return x_sample_recon, cost, monitors, updates_train, x_ts, updates_generate, u_ts[-1]
def _build_rnngsn(self):
"""
Creates the updates and other return variables for the computation graph.
Returns
-------
List
the sample at the end of the computation graph, the train cost function, the train monitors,
the computation updates, the generated visible list, the generated computation updates, the ending
recurrent states
"""
# For training, the deterministic recurrence is used to compute all the
# {h_t, 1 <= t <= T} given Xs. Conditional GSNs can then be trained
# in batches using those parameters.
(u, h_ts), updates_train = theano.scan(fn=lambda x_t, u_tm1: self.recurrent_step(x_t, u_tm1),
sequences=self.input,
outputs_info=[self.u0, None],
name="rnngsn_computation_scan")
h_list = [T.zeros_like(self.input)]
for layer, w in enumerate(self.weights_list[:self.layers]):
if layer % 2 != 0:
h_list.append(T.zeros_like(T.dot(h_list[-1], w)))
else:
h_list.append((h_ts.T[(layer/2) * self.hidden_size:(layer/2 + 1) * self.hidden_size]).T)
gsn = GSN(
inputs_hook = (self.input_size, self.input),
hiddens_hook = (self.hidden_size, GSN.pack_hiddens(h_list)),
params_hook = self.gsn_params,
outdir = os.path.join(self.outdir, 'gsn_noisy/'),
layers = self.layers,
walkbacks = self.walkbacks,
visible_activation = self.visible_activation_func,
hidden_activation = self.hidden_activation_func,
input_sampling = self.input_sampling,
mrg = self.mrg,
tied_weights = self.tied_weights,
cost_function = self.cost_function,
cost_args = self.cost_args,
add_noise = self.add_noise,
noiseless_h1 = self.noiseless_h1,
hidden_noise = self.hidden_noise,
hidden_noise_level = self.hidden_noise_level,
input_noise = self.input_noise,
input_noise_level = self.input_noise_level,
noise_decay = self.noise_decay,
noise_annealing = self.noise_annealing,
image_width = self.image_width,
image_height = self.image_height
)
cost = gsn.get_train_cost()
monitors = gsn.get_monitors() # frame-level error would be the 'mse' monitor from GSN
x_sample_recon = gsn.get_outputs()
# symbolic loop for sequence generation
(x_ts, u_ts), updates_generate = theano.scan(lambda u_tm1: self.recurrent_step(None, u_tm1),
outputs_info=[None, self.generate_u0],
n_steps=self.n_steps,
name="rnngsn_generate_scan")
return x_sample_recon, cost, monitors, updates_train, x_ts, updates_generate, u_ts[-1]
