def test_hierarchical_cnn_encoders(use_residual, hybridize) -> None:
num_ts = 2
ts_len = 10
num_static_feat = 2
num_dynamic_feat = 5
test_data = nd.arange(num_ts * ts_len).reshape(shape=(num_ts, ts_len, 1))
test_static_feat = nd.random.randn(num_ts, num_static_feat)
test_dynamic_feat = nd.random.randn(num_ts, ts_len, num_dynamic_feat)
chl_dim = [30, 30, 30]
ks_seq = [3] * len(chl_dim)
dial_seq = [1, 3, 9]
cnn = HierarchicalCausalConv1DEncoder(
dial_seq,
ks_seq,
chl_dim,
use_residual,
use_dynamic_feat=True,
use_static_feat=True,
)
cnn.collect_params().initialize()
if hybridize:
cnn.hybridize()
true_shape = (num_ts, ts_len, 31) if use_residual else (num_ts, ts_len, 30)
assert (cnn(test_data, test_static_feat,
test_dynamic_feat)[1].shape == true_shape)
def __init__(
self,
freq: str,
prediction_length: int,
cardinality: List[int],
embedding_dimension: int,
decoder_mlp_layer: List[int],
decoder_mlp_static_dim: int,
scaler: Scaler = NOPScaler(),
context_length: Optional[int] = None,
quantiles: Optional[List[float]] = None,
trainer: Trainer = Trainer(),
num_parallel_samples: int = 100,
) -> None:
encoder = HierarchicalCausalConv1DEncoder(
dilation_seq=[1, 3, 9],
kernel_size_seq=([3] * len([30, 30, 30])),
channels_seq=[30, 30, 30],
use_residual=True,
use_dynamic_feat=True,
use_static_feat=True,
)
super(CNN2QRForecaster, self).__init__(
freq=freq,
prediction_length=prediction_length,
encoder=encoder,
cardinality=cardinality,
embedding_dimension=embedding_dimension,
decoder_mlp_layer=decoder_mlp_layer,
decoder_mlp_static_dim=decoder_mlp_static_dim,
context_length=context_length,
scaler=scaler,
quantiles=quantiles,
trainer=trainer,
num_parallel_samples=num_parallel_samples,
)
def __init__(
self,
freq: str,
prediction_length: int,
context_length: Optional[int] = None,
use_past_feat_dynamic_real: bool = False,
use_feat_dynamic_real: bool = False,
use_feat_static_cat: bool = False,
cardinality: List[int] = None,
embedding_dimension: List[int] = None,
add_time_feature: bool = True,
add_age_feature: bool = False,
enable_encoder_dynamic_feature: bool = True,
enable_decoder_dynamic_feature: bool = True,
seed: Optional[int] = None,
decoder_mlp_dim_seq: Optional[List[int]] = None,
channels_seq: Optional[List[int]] = None,
dilation_seq: Optional[List[int]] = None,
kernel_size_seq: Optional[List[int]] = None,
use_residual: bool = True,
quantiles: Optional[List[float]] = None,
distr_output: Optional[DistributionOutput] = None,
trainer: Trainer = Trainer(),
scaling: Optional[bool] = None,
scaling_decoder_dynamic_feature: bool = False,
num_forking: Optional[int] = None,
max_ts_len: Optional[int] = None,
is_iqf: bool = True,
batch_size: int = 32,
) -> None:
assert (distr_output is None) or (quantiles is None)
assert (
prediction_length > 0
), f"Invalid prediction length: {prediction_length}."
assert decoder_mlp_dim_seq is None or all(
d > 0 for d in decoder_mlp_dim_seq
), "Elements of `mlp_hidden_dimension_seq` should be > 0"
assert channels_seq is None or all(
d > 0 for d in channels_seq
), "Elements of `channels_seq` should be > 0"
assert dilation_seq is None or all(
d > 0 for d in dilation_seq
), "Elements of `dilation_seq` should be > 0"
# TODO: add support for kernel size=1
assert kernel_size_seq is None or all(
d > 1 for d in kernel_size_seq
), "Elements of `kernel_size_seq` should be > 0"
assert quantiles is None or all(
0 <= d <= 1 for d in quantiles
), "Elements of `quantiles` should be >= 0 and <= 1"
self.decoder_mlp_dim_seq = (
decoder_mlp_dim_seq if decoder_mlp_dim_seq is not None else [30]
)
self.channels_seq = (
channels_seq if channels_seq is not None else [30, 30, 30]
)
self.dilation_seq = (
dilation_seq if dilation_seq is not None else [1, 3, 9]
)
self.kernel_size_seq = (
kernel_size_seq if kernel_size_seq is not None else [7, 3, 3]
)
self.quantiles = (
quantiles
if (quantiles is not None) or (distr_output is not None)
else [0.025, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 0.975]
)
self.is_iqf = is_iqf
assert (
len(self.channels_seq)
== len(self.dilation_seq)
== len(self.kernel_size_seq)
), (
f"mismatch CNN configurations: {len(self.channels_seq)} vs. "
f"{len(self.dilation_seq)} vs. {len(self.kernel_size_seq)}"
)
if seed:
np.random.seed(seed)
mx.random.seed(seed, trainer.ctx)
# `use_static_feat` and `use_dynamic_feat` always True because network
# always receives input; either from the input data or constants
encoder = HierarchicalCausalConv1DEncoder(
dilation_seq=self.dilation_seq,
kernel_size_seq=self.kernel_size_seq,
channels_seq=self.channels_seq,
use_residual=use_residual,
use_static_feat=True,
use_dynamic_feat=True,
prefix="encoder_",
)
decoder = ForkingMLPDecoder(
dec_len=prediction_length,
final_dim=self.decoder_mlp_dim_seq[-1],
hidden_dimension_sequence=self.decoder_mlp_dim_seq[:-1],
prefix="decoder_",
)
quantile_output = (
QuantileOutput(self.quantiles, is_iqf=self.is_iqf)
if self.quantiles
else None
)
super().__init__(
encoder=encoder,
decoder=decoder,
quantile_output=quantile_output,
distr_output=distr_output,
freq=freq,
prediction_length=prediction_length,
context_length=context_length,
use_past_feat_dynamic_real=use_past_feat_dynamic_real,
use_feat_dynamic_real=use_feat_dynamic_real,
use_feat_static_cat=use_feat_static_cat,
enable_encoder_dynamic_feature=enable_encoder_dynamic_feature,
enable_decoder_dynamic_feature=enable_decoder_dynamic_feature,
cardinality=cardinality,
embedding_dimension=embedding_dimension,
add_time_feature=add_time_feature,
add_age_feature=add_age_feature,
trainer=trainer,
scaling=scaling,
scaling_decoder_dynamic_feature=scaling_decoder_dynamic_feature,
num_forking=num_forking,
max_ts_len=max_ts_len,
batch_size=batch_size,
)