def get_dims(self,
y=None,
u_state=None,
u_obs=None,
n_timesteps=None,
n_batch=None):
if y is not None:
n_timesteps = y.shape[0]
n_batch = y.shape[1]
elif u_state is not None:
n_timesteps = u_state.shape[0]
n_batch = u_state.shape[1]
elif u_obs is not None:
n_timesteps = u_obs.shape[0]
n_batch = u_obs.shape[1]
else:
if n_timesteps is None and n_batch is None:
raise Exception(
"either provide n_timesteps and n_batch directly, "
"or provide any of (y, u_state, u_obs, u_switch). "
f"Got following types: "
f"y: {type(y)}, "
f"u_state: {type(u_state)}, "
f"u_obs: {type(u_obs)}, "
f"n_timesteps: {type(n_timesteps)}, "
f"n_batch: {type(n_batch)}")
return TensorDims(
timesteps=n_timesteps,
particle=self.n_particle,
batch=n_batch,
state=self.n_state,
target=self.n_target,
ctrl_target=self.n_ctrl_target,
ctrl_state=self.n_ctrl_state,
)
n_hidden_rnn: int
reconstruction_weight: float
rao_blackwellized: bool
dims_img = (1, 32, 32)
config = PymunkKVAEConfig(
dataset_name=consts.Datasets.box,
experiment_name="kvae",
dims_img=dims_img,
dims=TensorDims(
timesteps=20,
particle=1, # KVAE does not use SMC
batch=34, # pytorch-magma bug with 32.
state=10,
target=int(np.prod(dims_img)),
switch=None, # --> n_hidden_rnn
auxiliary=2,
ctrl_target=None,
ctrl_state=None,
),
#
batch_size_eval=4,
num_samples_eval=256,
prediction_length=40,
n_epochs=400,
lr=2e-3,
lr_decay_rate=0.85,
lr_decay_steps=20,
grad_clip_norm=1500.0,
weight_decay=0,
def make_default_config(dataset_name):
timefeat = TimeFeatType.timefeat
(
n_timefeat,
n_staticfeat,
n_latent,
freq,
cardinalities,
prediction_length_rolling,
prediction_length_full,
) = get_n_feat_and_freq(dataset_name=dataset_name, timefeat=timefeat)
assert len(cardinalities["cardinalities_feat_static_cat"]) == 1
n_static_embedding = min(
50, (cardinalities["cardinalities_feat_static_cat"][0] + 1) // 2
)
n_ctrl_all = n_ctrl_static = n_ctrl_dynamic = 64
# n_ctrl_static = n_static_embedding
# n_ctrl_dynamic = 32
# n_ctrl_all = n_ctrl_static + n_ctrl_dynamic # we cat
dims = TensorDims(
timesteps=past_lengths[dataset_name],
particle=10,
batch=50,
state=n_latent,
target=1,
switch=5,
# ctrl_state=None,
# ctrl_switch=n_staticfeat + n_timefeat,
# ctrl_obs=n_staticfeat + n_timefeat,
ctrl_state=n_ctrl_dynamic,
ctrl_target=n_ctrl_static,
ctrl_switch=n_ctrl_all, # switch takes cat feats
ctrl_encoder=n_ctrl_all, # encoder takes cat feats
timefeat=n_timefeat,
staticfeat=n_staticfeat,
cat_embedding=n_static_embedding,
auxiliary=None,
)
config = RsglsIssmGtsExpConfig(
experiment_name="rsgls",
dataset_name=dataset_name,
#
n_epochs=50,
n_epochs_no_resampling=5,
n_epochs_freeze_gls_params=1,
n_epochs_until_validate_loss=1,
lr=5e-3,
weight_decay=1e-5,
grad_clip_norm=10.0,
num_samples_eval=100,
batch_size_val=100, # 10
# gpus=tuple(range(3, 4)),
# dtype=torch.float64,
# architecture, prior, etc.
state_prior_scale=1.0,
state_prior_loc=0.0,
make_cov_from_cholesky_avg=True,
extract_tail_chunks_for_train=False,
switch_link_type=SwitchLinkType.individual,
switch_link_dims_hidden=(64,),
switch_link_activations=nn.LeakyReLU(0.1, inplace=True),
recurrent_link_type=SwitchLinkType.individual,
is_recurrent=True,
n_base_A=20,
n_base_B=20,
n_base_C=20,
n_base_D=20,
n_base_Q=20,
n_base_R=20,
n_base_F=20,
n_base_S=20,
requires_grad_R=True,
requires_grad_Q=True,
requires_grad_S=True,
# obs_to_switch_encoder=True,
# state_to_switch_encoder=False,
switch_prior_model_dims=tuple(),
# TODO: made assumption that this is used for ctrl_state...
input_transform_dims=(64,) + (dims.ctrl_state,),
switch_transition_model_dims=(64,),
# state_to_switch_encoder_dims=(64,),
obs_to_switch_encoder_dims=(64,),
b_fn_dims=tuple(),
d_fn_dims=tuple(), # (64,),
switch_prior_model_activations=LeakyReLU(0.1, inplace=True),
input_transform_activations=LeakyReLU(0.1, inplace=True),
switch_transition_model_activations=LeakyReLU(0.1, inplace=True),
# state_to_switch_encoder_activations=LeakyReLU(0.1, inplace=True),
obs_to_switch_encoder_activations=LeakyReLU(0.1, inplace=True),
b_fn_activations=LeakyReLU(0.1, inplace=True),
d_fn_activations=LeakyReLU(0.1, inplace=True),
# initialisation
init_scale_A=0.95,
init_scale_B=0.0,
init_scale_C=None,
init_scale_D=0.0,
init_scale_R_diag=[1e-5, 1e-1],
init_scale_Q_diag=[1e-4, 1e0],
init_scale_S_diag=[1e-5, 1e-1],
# set from outside due to dependencies.
dims=dims,
freq=freq,
time_feat=timefeat,
add_trend=add_trend_map[dataset_name],
prediction_length_rolling=prediction_length_rolling,
prediction_length_full=prediction_length_full,
normalisation_params=normalisation_params[dataset_name],
LRinv_logdiag_scaling=1.0,
LQinv_logdiag_scaling=1.0,
A_scaling=1.0,
B_scaling=1.0,
C_scaling=1.0,
D_scaling=1.0,
LSinv_logdiag_scaling=1.0,
F_scaling=1.0,
eye_init_A=True,
)
return config
def make_default_config(dataset_name):
timefeat = TimeFeatType.timefeat
(
n_timefeat,
n_staticfeat,
n_latent,
freq,
cardinalities,
prediction_length_rolling,
prediction_length_full,
) = get_n_feat_and_freq(dataset_name=dataset_name, timefeat=timefeat)
assert len(cardinalities["cardinalities_feat_static_cat"]) == 1
n_static_embedding = min(
50, (cardinalities["cardinalities_feat_static_cat"][0] + 1) // 2)
n_ctrl_all = n_ctrl_static = n_ctrl_dynamic = 64
# n_ctrl_static = n_static_embedding
# n_ctrl_dynamic = 32
# n_ctrl_all = n_ctrl_static + n_ctrl_dynamic # we cat
dims = TensorDims(
timesteps=past_lengths[dataset_name],
particle=1,
batch=50,
state=16, # n_latent,
target=1,
switch=None,
ctrl_state=n_ctrl_dynamic,
ctrl_target=n_ctrl_static,
ctrl_switch=n_ctrl_all,
ctrl_encoder=None, # KVAE uses pseudo-obs only, no controls for enc.
timefeat=n_timefeat,
staticfeat=n_staticfeat,
cat_embedding=n_static_embedding,
auxiliary=5,
)
config = KvaeGtsExpConfig(
experiment_name="kvae",
dataset_name=dataset_name,
#
n_epochs=50,
n_epochs_until_validate_loss=1,
lr=5e-3,
weight_decay=1e-5,
grad_clip_norm=10.0,
num_samples_eval=100,
# Note: These batch sizes barely fit on the GPU.
batch_size_val=10 if dataset_name
in ["exchange_rate_nips", "wiki2000_nips", "wiki2000_nips"] else 2,
# architecture, prior, etc.
state_prior_scale=1.0,
state_prior_loc=0.0,
make_cov_from_cholesky_avg=True,
extract_tail_chunks_for_train=False,
switch_link_type=SwitchLinkType.shared,
switch_link_dims_hidden=tuple(), # linear used in KVAE LSTM -> alpha
switch_link_activations=tuple(),
# they have 1 Dense layer after LSTM.
recurrent_link_type=SwitchLinkType.shared,
n_hidden_rnn=50,
rao_blackwellized=True,
reconstruction_weight=1.0, # They use 0.3 w/o rao-BW.
dims_encoder=(64, 64),
dims_decoder=(64, 64),
activations_encoder=LeakyReLU(0.1, inplace=True),
activations_decoder=LeakyReLU(0.1, inplace=True),
n_base_A=20,
n_base_B=20,
n_base_C=20,
n_base_D=None, # KVAE does not have D
n_base_Q=20,
n_base_R=20,
n_base_F=None,
n_base_S=None,
requires_grad_R=True,
requires_grad_Q=True,
requires_grad_S=None,
input_transform_dims=tuple() + (dims.ctrl_state, ),
input_transform_activations=LeakyReLU(0.1, inplace=True),
# initialisation
init_scale_A=0.95,
init_scale_B=0.0,
init_scale_C=None,
init_scale_D=None,
init_scale_R_diag=[1e-4, 1e-1],
init_scale_Q_diag=[1e-4, 1e-1],
init_scale_S_diag=None,
# init_scale_S_diag=[1e-5, 1e0],
# set from outside due to dependencies.
dims=dims,
freq=freq,
time_feat=timefeat,
add_trend=add_trend_map[dataset_name],
prediction_length_rolling=prediction_length_rolling,
prediction_length_full=prediction_length_full,
normalisation_params=normalisation_params[dataset_name],
LRinv_logdiag_scaling=1.0,
LQinv_logdiag_scaling=1.0,
A_scaling=1.0,
B_scaling=1.0,
C_scaling=1.0,
D_scaling=1.0,
LSinv_logdiag_scaling=1.0,
F_scaling=1.0,
eye_init_A=True,
)
return config
switch_prior_scale: float
switch_prior_loc: float
requires_grad_switch_prior: bool
dims_img = (1, 32, 32)
config = PymunkASGLSConfig(
dataset_name=consts.Datasets.box,
experiment_name="arsgls",
dims_img=dims_img,
dims=TensorDims(
timesteps=20,
particle=32,
batch=34, # pytorch-magma bug with 32.
state=10,
target=int(np.prod(dims_img)),
switch=8,
auxiliary=2,
ctrl_target=None,
ctrl_state=None,
),
#
batch_size_eval=8,
num_samples_eval=256,
prediction_length=40,
n_epochs=400,
lr=2e-3,
lr_decay_rate=0.85,
lr_decay_steps=20,
grad_clip_norm=1500.0,
weight_decay=0,
b_fn_dims: tuple
b_fn_activations: (nn.Module, tuple)
d_fn_dims: tuple
d_fn_activations: (nn.Module, tuple)
n_epochs_no_resampling: int
weight_decay: float
grad_clip_norm: float
dims = TensorDims(
timesteps=50,
particle=64,
batch=100,
state=3,
target=2,
switch=5,
auxiliary=None,
ctrl_target=None,
ctrl_state=None,
)
config = PendulumSGLSConfig(
dataset_name=consts.Datasets.pendulum_3D_coord,
experiment_name="default",
batch_size_eval=1000,
num_samples_eval=100,
lr=1e-2,
weight_decay=1e-5,
grad_clip_norm=500.0,
n_epochs=100,
n_epochs_no_resampling=10,