def _create_one_shooting_block(self,
name,
nr_of_channels,
nr_of_particles,
particle_size,
particle_dimension,
only_pass_through=False):
if only_pass_through:
shooting_model = self.shooting_model(
in_features=nr_of_channels,
nonlinearity=self.nonlinearity,
nr_of_particles=nr_of_particles,
is_pass_through=True,
**self._shooting_model_kwargs)
else:
shooting_model = self.shooting_model(
in_features=nr_of_channels,
nonlinearity=self.nonlinearity,
state_initializer=self._state_initializer,
costate_initializer=self._costate_initializer,
nr_of_particles=nr_of_particles,
particle_size=particle_size,
particle_dimension=particle_dimension,
**self._shooting_model_kwargs)
shooting_block = shooting_blocks.ShootingBlockBase(
name=name,
shooting_integrand=shooting_model,
integrator=self._integrator)
return shooting_block
def setup_shooting_block(integrator=None,
in_features=20,
shooting_model='updown',
parameter_weight=1.0,
nr_of_particles=10,
inflation_factor=2,
nonlinearity='relu',
use_particle_rnn_mode=False,
use_particle_free_rnn_mode=False,
optimize_over_data_initial_conditions=False,
optimize_over_data_initial_conditions_type='linear'):
if shooting_model == 'updown':
smodel = shooting_models.AutoShootingIntegrandModelUpDown(
in_features=in_features,
nonlinearity=nonlinearity,
parameter_weight=parameter_weight,
inflation_factor=inflation_factor,
nr_of_particles=nr_of_particles,
particle_dimension=1,
particle_size=in_features,
use_analytic_solution=True,
use_rnn_mode=use_particle_rnn_mode,
optimize_over_data_initial_conditions=
optimize_over_data_initial_conditions,
optimize_over_data_initial_conditions_type=
optimize_over_data_initial_conditions_type)
elif shooting_model == 'updown_universal':
smodel = shooting_models.AutoShootingIntegrandModelUpDownUniversal(
in_features=in_features,
nonlinearity=nonlinearity,
parameter_weight=parameter_weight,
inflation_factor=inflation_factor,
nr_of_particles=nr_of_particles,
particle_dimension=1,
particle_size=in_features,
use_analytic_solution=True,
optional_weight=0.1,
use_rnn_mode=use_particle_rnn_mode,
optimize_over_data_initial_conditions=
optimize_over_data_initial_conditions,
optimize_over_data_initial_conditions_type=
optimize_over_data_initial_conditions_type)
elif shooting_model == 'periodic':
smodel = shooting_models.AutoShootingIntegrandModelUpdownPeriodic(
in_features=in_features,
nonlinearity=nonlinearity,
parameter_weight=parameter_weight,
inflation_factor=inflation_factor,
nr_of_particles=nr_of_particles,
particle_dimension=1,
particle_size=in_features,
use_analytic_solution=True,
use_rnn_mode=use_particle_rnn_mode,
optimize_over_data_initial_conditions=
optimize_over_data_initial_conditions,
optimize_over_data_initial_conditions_type=
optimize_over_data_initial_conditions_type)
elif shooting_model == 'simple':
smodel = shooting_models.AutoShootingIntegrandModelSimple(
in_features=in_features,
nonlinearity=nonlinearity,
parameter_weight=parameter_weight,
nr_of_particles=nr_of_particles,
particle_dimension=1,
particle_size=in_features,
use_analytic_solution=True,
use_rnn_mode=use_particle_rnn_mode)
print('Using shooting model {}'.format(shooting_model))
par_initializer = pi.VectorEvolutionParameterInitializer(
only_random_initialization=True, random_initialization_magnitude=0.5)
smodel.set_state_initializer(state_initializer=par_initializer)
shooting_block = shooting_blocks.ShootingBlockBase(
name='simple',
shooting_integrand=smodel,
use_particle_free_rnn_mode=use_particle_free_rnn_mode,
integrator=integrator)
return shooting_block
inflation_factor=inflation_factor,
use_rnn_mode=use_particle_rnn_mode)
elif args.shooting_model == 'universal':
smodel = smodels.AutoShootingIntegrandModelUniversal(
**shootingintegrand_kwargs,
use_analytic_solution=use_analytic_solution,
inflation_factor=inflation_factor,
use_rnn_mode=use_particle_rnn_mode)
block_name = 'sblock'
sblock = sblocks.ShootingBlockBase(
name=block_name,
shooting_integrand=smodel,
integrator_name=args.method,
use_particle_free_rnn_mode=use_particle_free_rnn_mode,
use_particle_free_time_dependent_mode=args.
use_particle_free_time_dependent_mode,
nr_of_particle_free_time_dependent_steps=args.
nr_of_particle_free_time_dependent_steps,
integrator_options={'step_size': args.stepsize})
use_shooting = False
integrator = None
func = None
if args.use_rnn:
weight_decay = 0.0001
print('Using ResNetRNN: weight = {}'.format(weight_decay))
simple_resnet = sdnn.ResNetRNN(nr_of_layers=args.nr_of_layers,
inflation_factor=inflation_factor)
elif args.use_double_resnet_rnn:
elif current_model == 'DEBUG':
shooting_model = shooting_models.DEBUGAutoShootingIntegrandModelSimple(
in_features=in_features_size,
nonlinearity=nonlinearity,
nr_of_particles=nr_of_particles,
particle_dimension=1,
particle_size=in_features_size,
parameter_weight=parameter_weight)
else:
raise ValueError('Unknown model to check: {}'.format(current_model))
use_analytic_solution = True
shooting_block = shooting_blocks.ShootingBlockBase(
name='test',
shooting_integrand=shooting_model,
integrator_options=integrator_options)
shooting_model.use_analytic_solution = use_analytic_solution
print('\n\nChecking model: {}'.format(current_model))
print('-------------------------------------\n')
# create some sample data
sample_data = torch.randn([1, 1, in_features_size])
# run through the shooting block once to get the necessary parameters
shooting_block(x=sample_data)
# create overall time-vector, for simplicity make sure it corresponds to the step-size
t_np = np.array(range(0, 16)) * stepsize
t = torch.from_numpy(t_np)
parameter_weight=parameter_weight)
elif current_model == "general":
shooting_model = shooting_models.AutoShootingIntegrandModelGeneralUpDown(
in_features=in_features_size,
nonlinearity=nonlinearity,
nr_of_particles=nr_of_particles,
particle_dimension=1,
particle_size=in_features_size,
parameter_weight=parameter_weight,
inflation_factor=5)
else:
raise ValueError('Unknown model to check: {}'.format(current_model))
shooting_block = shooting_blocks.ShootingBlockBase(
name='test', shooting_integrand=shooting_model)
print('\n\nChecking model: {}'.format(current_model))
print('-------------------------------------\n')
# create some sample data
sample_data = torch.randn([100, 1, in_features_size])
sample_data_init = torch.randn([100, 1, in_features_size])
autodiff_gradient_results = dict()
analytic_gradient_results = dict()
for use_analytic_solution in (True, False):
shooting_model.use_analytic_solution = use_analytic_solution
# run through the shooting block once to get the necessary parameters
