wassname/pytorch-ts
1
0
Fork 0
mirror of https://github.com/wassname/pytorch-ts.git synced 2026-07-26 13:37:40 +08:00
Code Issues Packages Projects Releases Wiki Activity

Deepvar cleanup (#15)

Browse Source 
* initial age feature

* added support for feat_static_cat in network

* added feature_static_real

also concat the log(scale) to the network

* fixed typo

* FEAT_STATIC_CAT are np.long
This commit is contained in:
Kashif Rasul
2020-06-18 12:30:55 +02:00
committed by GitHub Enterprise
parent e063a64ccc
commit 8e9f31eae1
3 changed files with 89 additions and 65 deletions
Download Patch File Download Diff File Expand all files Collapse all files
pts/model/deepvar/__init__.py
+1 -1
View File
@@ -1 +1 @@
from .deepvar_estimator import DeepVAREstimator
from .deepvar_estimator import DeepVAREstimator
pts/model/deepvar/deepvar_estimator.py
+39 -20
View File
@@ -1,5 +1,6 @@
from typing import List, Optional
from typing import List, Optional, Callable
import numpy as np
import torch
from pts import Trainer
@@ -24,8 +25,10 @@ from pts.transform import (
ExpandDimArray,
AddObservedValuesIndicator,
AddTimeFeatures,
AddAgeFeature,
VstackFeatures,
SetFieldIfNotPresent,
SetField,
TargetDimIndicator,
)
from .deepvar_network import DeepVARTrainingNetwork, DeepVARPredictionNetwork
@@ -46,9 +49,10 @@ class DeepVAREstimator(PTSEstimator):
num_parallel_samples: int = 100,
dropout_rate: float = 0.1,
use_feat_dynamic_real: bool = False,
use_feat_dynamic_cat: bool = False,
cardinality: List[int] = [1],
embedding_dimension: int = 5,
use_feat_static_cat: bool = False,
use_feat_static_real: bool = False,
cardinality: Optional[List[int]] = None,
embedding_dimension: Optional[List[int]] = None,
distr_output: Optional[DistributionOutput] = None,
rank: Optional[int] = 5,
scaling: bool = True,
@@ -82,9 +86,14 @@ class DeepVAREstimator(PTSEstimator):
self.num_parallel_samples = num_parallel_samples
self.dropout_rate = dropout_rate
self.use_feat_dynamic_real = use_feat_dynamic_real
self.use_feat_dynamic_cat = use_feat_dynamic_cat
self.cardinality = cardinality
self.embedding_dimension = embedding_dimension
self.use_feat_static_cat = use_feat_static_cat
self.use_feat_static_real = use_feat_static_real
self.cardinality = cardinality if cardinality and use_feat_static_cat else [1]
self.embedding_dimension = (
embedding_dimension
if embedding_dimension is not None
else [min(50, (cat + 1) // 2) for cat in self.cardinality]
)
self.conditioning_length = conditioning_length
self.use_marginal_transformation = use_marginal_transformation
@@ -130,14 +139,24 @@ class DeepVAREstimator(PTSEstimator):
}
)
remove_field_names = []
remove_field_names = [FieldName.FEAT_DYNAMIC_CAT]
if not self.use_feat_dynamic_real:
remove_field_names.append(FieldName.FEAT_DYNAMIC_REAL)
if not self.use_feat_dynamic_cat:
remove_field_names.append(FieldName.FEAT_DYNAMIC_CAT)
if not self.use_feat_static_real:
remove_field_names.append(FieldName.FEAT_STATIC_REAL)
return Chain(
[RemoveFields(field_names=remove_field_names)]
+ (
[SetField(output_field=FieldName.FEAT_STATIC_CAT, value=[0])]
if not self.use_feat_static_cat
else []
)
+ (
[SetField(output_field=FieldName.FEAT_STATIC_REAL, value=[0.0])]
if not self.use_feat_static_real
else []
)
+ [
AsNumpyArray(
field=FieldName.TARGET,
@@ -160,26 +179,28 @@ class DeepVAREstimator(PTSEstimator):
time_features=self.time_features,
pred_length=self.prediction_length,
),
AddAgeFeature(
target_field=FieldName.TARGET,
output_field=FieldName.FEAT_AGE,
pred_length=self.prediction_length,
log_scale=True,
dtype=self.dtype,
),
VstackFeatures(
output_field=FieldName.FEAT_TIME,
input_fields=[FieldName.FEAT_TIME]
input_fields=[FieldName.FEAT_TIME, FieldName.FEAT_AGE]
+ (
[FieldName.FEAT_DYNAMIC_REAL]
if self.use_feat_dynamic_real
else []
)
+ (
[FieldName.FEAT_DYNAMIC_CAT]
if self.use_feat_dynamic_cat
else []
),
),
SetFieldIfNotPresent(field=FieldName.FEAT_STATIC_CAT, value=[0]),
TargetDimIndicator(
field_name="target_dimension_indicator",
target_field=FieldName.TARGET,
),
AsNumpyArray(field=FieldName.FEAT_STATIC_CAT, expected_ndim=1),
AsNumpyArray(field=FieldName.FEAT_STATIC_CAT, expected_ndim=1, dtype=np.long),
AsNumpyArray(field=FieldName.FEAT_STATIC_REAL, expected_ndim=1),
InstanceSplitter(
target_field=FieldName.TARGET,
is_pad_field=FieldName.IS_PAD,
@@ -214,7 +235,6 @@ class DeepVAREstimator(PTSEstimator):
embedding_dimension=self.embedding_dimension,
lags_seq=self.lags_seq,
scaling=self.scaling,
conditioning_length=self.conditioning_length,
).to(device)
def create_predictor(
@@ -239,7 +259,6 @@ class DeepVAREstimator(PTSEstimator):
embedding_dimension=self.embedding_dimension,
lags_seq=self.lags_seq,
scaling=self.scaling,
conditioning_length=self.conditioning_length,
).to(device)
copy_parameters(trained_network, prediction_network)
pts/model/deepvar/deepvar_network.py
+49 -44
View File
@@ -5,7 +5,7 @@ import torch.nn as nn
from pts.core.component import validated
from pts.model import weighted_average
from pts.modules import DistributionOutput, MeanScaler, NOPScaler
from pts.modules import DistributionOutput, MeanScaler, NOPScaler, FeatureEmbedder
class DeepVARTrainingNetwork(nn.Module):
@@ -23,9 +23,8 @@ class DeepVARTrainingNetwork(nn.Module):
dropout_rate: float,
lags_seq: List[int],
target_dim: int,
conditioning_length: int,
cardinality: List[int] = [1],
embedding_dimension: int = 1,
cardinality: List[int],
embedding_dimension: List[int],
scaling: bool = True,
**kwargs,
) -> None:
@@ -43,7 +42,6 @@ class DeepVARTrainingNetwork(nn.Module):
self.target_dim = target_dim
self.scaling = scaling
self.target_dim_sample = target_dim
self.conditioning_length = conditioning_length
assert len(set(lags_seq)) == len(lags_seq), "no duplicated lags allowed!"
lags_seq.sort()
@@ -65,9 +63,8 @@ class DeepVARTrainingNetwork(nn.Module):
self.proj_dist_args = distr_output.get_args_proj(num_cells)
self.embed_dim = 1
self.embed = nn.Embedding(
num_embeddings=self.target_dim, embedding_dim=self.embed_dim
self.embed = FeatureEmbedder(
cardinalities=cardinality, embedding_dims=embedding_dimension
)
if scaling:
@@ -124,7 +121,8 @@ class DeepVARTrainingNetwork(nn.Module):
lags: torch.Tensor,
scale: torch.Tensor,
time_feat: torch.Tensor,
target_dimension_indicator: torch.Tensor,
feat_static_cat: torch.Tensor,
feat_static_real: torch.Tensor,
unroll_length: int,
begin_state: Optional[Union[List[torch.Tensor], torch.Tensor]] = None,
) -> Tuple[
@@ -146,18 +144,16 @@ class DeepVARTrainingNetwork(nn.Module):
)
# (batch_size, target_dim, embed_dim)
index_embeddings = self.embed(target_dimension_indicator)
embedded_cat = self.embed(feat_static_cat)
# assert_shape(index_embeddings, (-1, self.target_dim, self.embed_dim))
# (batch_size, seq_len, target_dim * embed_dim)
repeated_index_embeddings = (
index_embeddings.unsqueeze(1)
.expand(-1, unroll_length, -1, -1)
.reshape((-1, unroll_length, self.target_dim * self.embed_dim))
)
static_feat = torch.cat((embedded_cat, feat_static_real, scale.log()), dim=1)
# (batch_size, seq_len, embed_dim)
repeated_static_feat = static_feat.unsqueeze(1).expand(-1, unroll_length, -1)
# (batch_size, sub_seq_len, input_dim)
inputs = torch.cat((input_lags, repeated_index_embeddings, time_feat), dim=-1)
inputs = torch.cat((input_lags, repeated_static_feat, time_feat), dim=-1)
# unroll encoder
outputs, state = self.rnn(inputs, begin_state)
@@ -174,13 +170,14 @@ class DeepVARTrainingNetwork(nn.Module):
def unroll_encoder(
self,
feat_static_real: torch.Tensor,
past_time_feat: torch.Tensor,
past_target_cdf: torch.Tensor,
past_observed_values: torch.Tensor,
past_is_pad: torch.Tensor,
future_time_feat: Optional[torch.Tensor],
future_target_cdf: Optional[torch.Tensor],
target_dimension_indicator: torch.Tensor,
feat_static_cat: torch.Tensor,
) -> Tuple[
torch.Tensor,
Union[List[torch.Tensor], torch.Tensor],
@@ -197,6 +194,8 @@ class DeepVARTrainingNetwork(nn.Module):
Parameters
----------
feat_static_real
past_time_feat
Past time features (batch_size, history_length, num_features)
past_target_cdf
@@ -213,8 +212,8 @@ class DeepVARTrainingNetwork(nn.Module):
future_target_cdf
Future marginal CDF transformed target values (batch_size,
prediction_length, target_dim)
target_dimension_indicator
Dimensionality of the time series (batch_size, target_dim)
feat_static_cat
Returns
-------
@@ -237,18 +236,13 @@ class DeepVARTrainingNetwork(nn.Module):
)
if future_time_feat is None or future_target_cdf is None:
time_feat = past_time_feat[
:,- self.context_length :, ...
]
time_feat = past_time_feat[:, -self.context_length :, ...]
sequence = past_target_cdf
sequence_length = self.history_length
subsequences_length = self.context_length
else:
time_feat = torch.cat(
(
past_time_feat[:,- self.context_length :, ...],
future_time_feat,
),
(past_time_feat[:, -self.context_length :, ...], future_time_feat,),
dim=1,
)
sequence = torch.cat((past_target_cdf, future_target_cdf), dim=1)
@@ -274,7 +268,8 @@ class DeepVARTrainingNetwork(nn.Module):
lags=lags,
scale=scale,
time_feat=time_feat,
target_dimension_indicator=target_dimension_indicator,
feat_static_cat=feat_static_cat,
feat_static_real=feat_static_real,
unroll_length=subsequences_length,
begin_state=None,
)
@@ -310,7 +305,8 @@ class DeepVARTrainingNetwork(nn.Module):
def forward(
self,
target_dimension_indicator: torch.Tensor,
feat_static_cat: torch.Tensor,
feat_static_real: torch.Tensor,
past_time_feat: torch.Tensor,
past_target_cdf: torch.Tensor,
past_observed_values: torch.Tensor,
@@ -325,8 +321,8 @@ class DeepVARTrainingNetwork(nn.Module):
Parameters
----------
target_dimension_indicator
Indices of the target dimension (batch_size, target_dim)
feat_static_cat
past_time_feat
Dynamic features of past time series (batch_size, history_length,
num_features)
@@ -365,13 +361,14 @@ class DeepVARTrainingNetwork(nn.Module):
# unroll the decoder in "training mode", i.e. by providing future data
# as well
rnn_outputs, _, scale, _, inputs = self.unroll_encoder(
feat_static_real=feat_static_real,
past_time_feat=past_time_feat,
past_target_cdf=past_target_cdf,
past_observed_values=past_observed_values,
past_is_pad=past_is_pad,
future_time_feat=future_time_feat,
future_target_cdf=future_target_cdf,
target_dimension_indicator=target_dimension_indicator,
feat_static_cat=feat_static_cat,
)
# put together target sequence
@@ -413,8 +410,6 @@ class DeepVARTrainingNetwork(nn.Module):
# assert_shape(loss, (-1, -1, 1))
self.distribution = distr
return (loss.mean(), likelihoods) + distr_args
@@ -431,7 +426,8 @@ class DeepVARPredictionNetwork(DeepVARTrainingNetwork):
def sampling_decoder(
self,
past_target_cdf: torch.Tensor,
target_dimension_indicator: torch.Tensor,
feat_static_cat: torch.Tensor,
feat_static_real: torch.Tensor,
time_feat: torch.Tensor,
scale: torch.Tensor,
begin_states: Union[List[torch.Tensor], torch.Tensor],
@@ -445,8 +441,10 @@ class DeepVARPredictionNetwork(DeepVARTrainingNetwork):
past_target_cdf
Past marginal CDF transformed target values (batch_size,
history_length, target_dim)
target_dimension_indicator
Indices of the target dimension (batch_size, target_dim)
feat_static_cat
feat_static_real
time_feat
Dynamic features of future time series (batch_size, history_length,
num_features)
@@ -469,7 +467,8 @@ class DeepVARPredictionNetwork(DeepVARTrainingNetwork):
repeated_past_target_cdf = repeat(past_target_cdf)
repeated_time_feat = repeat(time_feat)
repeated_scale = repeat(scale)
repeated_target_dimension_indicator = repeat(target_dimension_indicator)
repeated_feat_static_cat = repeat(feat_static_cat)
repeated_feat_static_real = repeat(feat_static_real)
# slight difference for GPVAR and DeepVAR, in GPVAR, its a list
if self.cell_type == "LSTM":
@@ -494,7 +493,8 @@ class DeepVARPredictionNetwork(DeepVARTrainingNetwork):
lags=lags,
scale=repeated_scale,
time_feat=repeated_time_feat[:, k : k + 1, ...],
target_dimension_indicator=repeated_target_dimension_indicator,
feat_static_cat=repeated_feat_static_cat,
feat_static_real=repeated_feat_static_real,
unroll_length=1,
)
@@ -521,7 +521,8 @@ class DeepVARPredictionNetwork(DeepVARTrainingNetwork):
def forward(
self,
target_dimension_indicator: torch.Tensor,
feat_static_cat: torch.Tensor,
feat_static_real: torch.Tensor,
past_time_feat: torch.Tensor,
past_target_cdf: torch.Tensor,
past_observed_values: torch.Tensor,
@@ -533,8 +534,10 @@ class DeepVARPredictionNetwork(DeepVARTrainingNetwork):
All tensors should have NTC layout.
Parameters
----------
target_dimension_indicator
Indices of the target dimension (batch_size, target_dim)
feat_static_cat
feat_static_real
past_time_feat
Dynamic features of past time series (batch_size, history_length,
num_features)
@@ -566,18 +569,20 @@ class DeepVARPredictionNetwork(DeepVARTrainingNetwork):
# unroll the decoder in "prediction mode", i.e. with past data only
_, state, scale, _, _ = self.unroll_encoder(
feat_static_real=feat_static_real,
past_time_feat=past_time_feat,
past_target_cdf=past_target_cdf,
past_observed_values=past_observed_values,
past_is_pad=past_is_pad,
future_time_feat=None,
future_target_cdf=None,
target_dimension_indicator=target_dimension_indicator,
feat_static_cat=feat_static_cat,
)
return self.sampling_decoder(
past_target_cdf=past_target_cdf,
target_dimension_indicator=target_dimension_indicator,
feat_static_cat=feat_static_cat,
feat_static_real=feat_static_real,
time_feat=future_time_feat,
scale=scale,
begin_states=state,