Deep learning examples

This notebooks contains examples with neural network models.

Table of Contents

• Creating TSDataset

• Architecture

• Testing models

  • DeepAR

  • TFT

  • Simple Model

import torch
import random

import pandas as pd
import numpy as np

from etna.datasets.tsdataset import TSDataset
from etna.pipeline import Pipeline
from etna.transforms import DateFlagsTransform
from etna.transforms import LagTransform
from etna.transforms import LinearTrendTransform
from etna.metrics import SMAPE, MAPE, MAE
from etna.analysis import plot_backtest
from etna.models import SeasonalMovingAverageModel

import warnings
warnings.filterwarnings("ignore")

1. Creating TSDataset

We are going to take transformed [Household Electric Power Consumption] dataset. Let’s load and look at it.

original_df = pd.read_csv("data/example_dataset.csv")
original_df.head()

	timestamp	segment	target
0	2019-01-01	segment_a	170
1	2019-01-02	segment_a	243
2	2019-01-03	segment_a	267
3	2019-01-04	segment_a	287
4	2019-01-05	segment_a	279

Our library works with the spacial data structure TSDataset. Let’s create it as it was done in “Get started” notebook.

df = TSDataset.to_dataset(original_df)
ts = TSDataset(df, freq="D")
ts.head(5)

segment	segment_a	segment_b	segment_c	segment_d
feature	target	target	target	target
timestamp
2019-01-01	170	102	92	238
2019-01-02	243	123	107	358
2019-01-03	267	130	103	366
2019-01-04	287	138	103	385
2019-01-05	279	137	104	384

2. Architecture

Our library uses PyTorch Forecasting to work with time series neural networks. To include it in our current architecture we use PytorchForecastingTransform class.

Let’s look at it closer.

from etna.transforms import PytorchForecastingTransform

?PytorchForecastingTransform

"""
Init signature:
PytorchForecastingTransform(
    max_encoder_length: int = 30,
    min_encoder_length: int = None,
    min_prediction_idx: int = None,
    min_prediction_length: int = None,
    max_prediction_length: int = 1,
    static_categoricals: List[str] = [],
    static_reals: List[str] = [],
    time_varying_known_categoricals: List[str] = [],
    time_varying_known_reals: List[str] = [],
    time_varying_unknown_categoricals: List[str] = [],
    time_varying_unknown_reals: List[str] = [],
    variable_groups: Dict[str, List[int]] = {},
    constant_fill_strategy: Dict[str, Union[str, float, int, bool]] = {},
    allow_missing_timesteps: bool = True,
    lags: Dict[str, List[int]] = {},
    add_relative_time_idx: bool = True,
    add_target_scales: bool = True,
    add_encoder_length: Union[bool, str] = True,
    target_normalizer: Union[pytorch_forecasting.data.encoders.TorchNormalizer, pytorch_forecasting.data.encoders.NaNLabelEncoder, pytorch_forecasting.data.encoders.EncoderNormalizer, str, List[Union[pytorch_forecasting.data.encoders.TorchNormalizer, pytorch_forecasting.data.encoders.NaNLabelEncoder, pytorch_forecasting.data.encoders.EncoderNormalizer]], Tuple[Union[pytorch_forecasting.data.encoders.TorchNormalizer, pytorch_forecasting.data.encoders.NaNLabelEncoder, pytorch_forecasting.data.encoders.EncoderNormalizer]]] = 'auto',
    categorical_encoders: Dict[str, pytorch_forecasting.data.encoders.NaNLabelEncoder] = None,
    scalers: Dict[str, Union[sklearn.preprocessing._data.StandardScaler, sklearn.preprocessing._data.RobustScaler, pytorch_forecasting.data.encoders.TorchNormalizer, pytorch_forecasting.data.encoders.EncoderNormalizer]] = {},
)
Docstring:      Transform for models from PytorchForecasting library.
Init docstring:
Parameters for TimeSeriesDataSet object.

Reference
---------
https://github.com/jdb78/pytorch-forecasting/blob/v0.8.5/pytorch_forecasting/data/timeseries.py#L117
"""

We can see a pretty scary signature, but don’t panic, we will look at the most important parameters.

• time_varying_known_reals — known real values that change across the time (real regressors), now it it necessary to add “time_idx” variable to the list;

• time_varying_unknown_reals — our real value target, set it to ["target"];

• max_prediction_length — our horizon for forecasting;

• max_encoder_length — length of past context to use;

• static_categoricals — static categorical values, for example, if we use multiple segments it can be some its characteristics including identifier: “segment”;

• time_varying_known_categoricals — known categorical values that change across the time (categorical regressors);

• target_normalizer — class for normalization targets across different segments.

Our library currently supports these models: * DeepAR, * TFT.

3. Testing models

In this section we will test our models on example.

3.1 DeepAR

Before training let’s fix seeds for reproducibility.

torch.manual_seed(42)
random.seed(42)
np.random.seed(42)

Creating transforms for DeepAR.

from pytorch_forecasting.data import GroupNormalizer

HORIZON = 7

transform_date = DateFlagsTransform(day_number_in_week=True, day_number_in_month=False, out_column="dateflag")
num_lags = 10
transform_lag = LagTransform(in_column="target", lags=[HORIZON+i for i in range(num_lags)], out_column="target_lag")
lag_columns = [f"target_lag_{HORIZON+i}" for i in range(num_lags)]

transform_deepar = PytorchForecastingTransform(
    max_encoder_length=HORIZON,
    max_prediction_length=HORIZON,
    time_varying_known_reals=["time_idx"]+lag_columns,
    time_varying_unknown_reals=["target"],
    time_varying_known_categoricals=["dateflag_day_number_in_week"],
    target_normalizer=GroupNormalizer(groups=["segment"]),
)

Now we are going to start backtest.

from etna.models.nn import DeepARModel


model_deepar = DeepARModel(max_epochs=150, learning_rate=[0.01], gpus=0, batch_size=64)
metrics = [SMAPE(), MAPE(), MAE()]

pipeline_deepar = Pipeline(model=model_deepar,
                           horizon=HORIZON,
                           transforms=[transform_lag, transform_date, transform_deepar])

metrics_deepar, forecast_deepar, fold_info_deepar = pipeline_deepar.backtest(ts, metrics=metrics, n_folds=3, n_jobs=1)

[Parallel(n_jobs=1)]: Using backend SequentialBackend with 1 concurrent workers.
GPU available: False, used: False
TPU available: False, using: 0 TPU cores
IPU available: False, using: 0 IPUs

  | Name                   | Type                   | Params
------------------------------------------------------------------
0 | loss                   | NormalDistributionLoss | 0
1 | logging_metrics        | ModuleList             | 0
2 | embeddings             | MultiEmbedding         | 35
3 | rnn                    | LSTM                   | 2.2 K
4 | distribution_projector | Linear                 | 22
------------------------------------------------------------------
2.3 K     Trainable params
0         Non-trainable params
2.3 K     Total params
0.009     Total estimated model params size (MB)

[Parallel(n_jobs=1)]: Done   1 out of   1 | elapsed:  1.9min remaining:    0.0s
GPU available: False, used: False
TPU available: False, using: 0 TPU cores
IPU available: False, using: 0 IPUs

  | Name                   | Type                   | Params
------------------------------------------------------------------
0 | loss                   | NormalDistributionLoss | 0
1 | logging_metrics        | ModuleList             | 0
2 | embeddings             | MultiEmbedding         | 35
3 | rnn                    | LSTM                   | 2.2 K
4 | distribution_projector | Linear                 | 22
------------------------------------------------------------------
2.3 K     Trainable params
0         Non-trainable params
2.3 K     Total params
0.009     Total estimated model params size (MB)

[Parallel(n_jobs=1)]: Done   2 out of   2 | elapsed:  3.9min remaining:    0.0s
GPU available: False, used: False
TPU available: False, using: 0 TPU cores
IPU available: False, using: 0 IPUs

  | Name                   | Type                   | Params
------------------------------------------------------------------
0 | loss                   | NormalDistributionLoss | 0
1 | logging_metrics        | ModuleList             | 0
2 | embeddings             | MultiEmbedding         | 35
3 | rnn                    | LSTM                   | 2.2 K
4 | distribution_projector | Linear                 | 22
------------------------------------------------------------------
2.3 K     Trainable params
0         Non-trainable params
2.3 K     Total params
0.009     Total estimated model params size (MB)

[Parallel(n_jobs=1)]: Done   3 out of   3 | elapsed:  6.0min remaining:    0.0s
[Parallel(n_jobs=1)]: Done   3 out of   3 | elapsed:  6.0min finished

Let’s compare results across different segments.

metrics_deepar

	segment	SMAPE	MAPE	MAE	fold_number
0	segment_a	6.496923	6.262331	33.577118	0
0	segment_a	3.110433	3.115227	16.091409	1
0	segment_a	3.247720	3.195171	17.676841	2
3	segment_b	6.794060	6.527645	17.035167	0
3	segment_b	4.234161	4.282767	10.278190	1
3	segment_b	4.020539	4.129931	9.492033	2
1	segment_c	3.464639	3.436801	5.854686	0
1	segment_c	5.573871	5.410155	9.875741	1
1	segment_c	3.951570	3.798992	6.880014	2
2	segment_d	6.403644	6.308158	54.982370	0
2	segment_d	6.815848	6.975395	55.210536	1
2	segment_d	3.138582	3.075240	26.546858	2

To summarize it we will take mean value of SMAPE metric because it is scale tolerant.

score = metrics_deepar["SMAPE"].mean()
print(f"Average SMAPE for DeepAR: {score:.3f}")

Average SMAPE for DeepAR: 4.771

Visualize results.

plot_backtest(forecast_deepar, ts, history_len=20)

3.2 TFT

Let’s move to the next model.

torch.manual_seed(42)
random.seed(42)
np.random.seed(42)

transform_date = DateFlagsTransform(day_number_in_week=True, day_number_in_month=False, out_column="dateflag")
num_lags = 10
transform_lag = LagTransform(in_column="target", lags=[HORIZON+i for i in range(num_lags)], out_column="target_lag")
lag_columns = [f"target_lag_{HORIZON+i}" for i in range(num_lags)]

transform_tft = PytorchForecastingTransform(
    max_encoder_length=HORIZON,
    max_prediction_length=HORIZON,
    time_varying_known_reals=["time_idx"],
    time_varying_unknown_reals=["target"],
    time_varying_known_categoricals=["dateflag_day_number_in_week"],
    static_categoricals=["segment"],
    target_normalizer=GroupNormalizer(groups=["segment"]),
)

from etna.models.nn import TFTModel

model_tft = TFTModel(max_epochs=200, learning_rate=[0.01], gpus=0, batch_size=64)

pipeline_tft = Pipeline(model=model_tft,
                        horizon=HORIZON,
                        transforms=[transform_lag, transform_date, transform_tft])

metrics_tft, forecast_tft, fold_info_tft = pipeline_tft.backtest(ts, metrics=metrics, n_folds=3, n_jobs=1)

[Parallel(n_jobs=1)]: Using backend SequentialBackend with 1 concurrent workers.
GPU available: False, used: False
TPU available: False, using: 0 TPU cores
IPU available: False, using: 0 IPUs

   | Name                               | Type                            | Params
----------------------------------------------------------------------------------------
0  | loss                               | QuantileLoss                    | 0
1  | logging_metrics                    | ModuleList                      | 0
2  | input_embeddings                   | MultiEmbedding                  | 47
3  | prescalers                         | ModuleDict                      | 96
4  | static_variable_selection          | VariableSelectionNetwork        | 1.8 K
5  | encoder_variable_selection         | VariableSelectionNetwork        | 1.9 K
6  | decoder_variable_selection         | VariableSelectionNetwork        | 1.3 K
7  | static_context_variable_selection  | GatedResidualNetwork            | 1.1 K
8  | static_context_initial_hidden_lstm | GatedResidualNetwork            | 1.1 K
9  | static_context_initial_cell_lstm   | GatedResidualNetwork            | 1.1 K
10 | static_context_enrichment          | GatedResidualNetwork            | 1.1 K
11 | lstm_encoder                       | LSTM                            | 2.2 K
12 | lstm_decoder                       | LSTM                            | 2.2 K
13 | post_lstm_gate_encoder             | GatedLinearUnit                 | 544
14 | post_lstm_add_norm_encoder         | AddNorm                         | 32
15 | static_enrichment                  | GatedResidualNetwork            | 1.4 K
16 | multihead_attn                     | InterpretableMultiHeadAttention | 676
17 | post_attn_gate_norm                | GateAddNorm                     | 576
18 | pos_wise_ff                        | GatedResidualNetwork            | 1.1 K
19 | pre_output_gate_norm               | GateAddNorm                     | 576
20 | output_layer                       | Linear                          | 119
----------------------------------------------------------------------------------------
18.9 K    Trainable params
0         Non-trainable params
18.9 K    Total params
0.075     Total estimated model params size (MB)

[Parallel(n_jobs=1)]: Done   1 out of   1 | elapsed:  4.7min remaining:    0.0s
GPU available: False, used: False
TPU available: False, using: 0 TPU cores
IPU available: False, using: 0 IPUs

   | Name                               | Type                            | Params
----------------------------------------------------------------------------------------
0  | loss                               | QuantileLoss                    | 0
1  | logging_metrics                    | ModuleList                      | 0
2  | input_embeddings                   | MultiEmbedding                  | 47
3  | prescalers                         | ModuleDict                      | 96
4  | static_variable_selection          | VariableSelectionNetwork        | 1.8 K
5  | encoder_variable_selection         | VariableSelectionNetwork        | 1.9 K
6  | decoder_variable_selection         | VariableSelectionNetwork        | 1.3 K
7  | static_context_variable_selection  | GatedResidualNetwork            | 1.1 K
8  | static_context_initial_hidden_lstm | GatedResidualNetwork            | 1.1 K
9  | static_context_initial_cell_lstm   | GatedResidualNetwork            | 1.1 K
10 | static_context_enrichment          | GatedResidualNetwork            | 1.1 K
11 | lstm_encoder                       | LSTM                            | 2.2 K
12 | lstm_decoder                       | LSTM                            | 2.2 K
13 | post_lstm_gate_encoder             | GatedLinearUnit                 | 544
14 | post_lstm_add_norm_encoder         | AddNorm                         | 32
15 | static_enrichment                  | GatedResidualNetwork            | 1.4 K
16 | multihead_attn                     | InterpretableMultiHeadAttention | 676
17 | post_attn_gate_norm                | GateAddNorm                     | 576
18 | pos_wise_ff                        | GatedResidualNetwork            | 1.1 K
19 | pre_output_gate_norm               | GateAddNorm                     | 576
20 | output_layer                       | Linear                          | 119
----------------------------------------------------------------------------------------
18.9 K    Trainable params
0         Non-trainable params
18.9 K    Total params
0.075     Total estimated model params size (MB)

[Parallel(n_jobs=1)]: Done   2 out of   2 | elapsed: 10.0min remaining:    0.0s
GPU available: False, used: False
TPU available: False, using: 0 TPU cores
IPU available: False, using: 0 IPUs

   | Name                               | Type                            | Params
----------------------------------------------------------------------------------------
0  | loss                               | QuantileLoss                    | 0
1  | logging_metrics                    | ModuleList                      | 0
2  | input_embeddings                   | MultiEmbedding                  | 47
3  | prescalers                         | ModuleDict                      | 96
4  | static_variable_selection          | VariableSelectionNetwork        | 1.8 K
5  | encoder_variable_selection         | VariableSelectionNetwork        | 1.9 K
6  | decoder_variable_selection         | VariableSelectionNetwork        | 1.3 K
7  | static_context_variable_selection  | GatedResidualNetwork            | 1.1 K
8  | static_context_initial_hidden_lstm | GatedResidualNetwork            | 1.1 K
9  | static_context_initial_cell_lstm   | GatedResidualNetwork            | 1.1 K
10 | static_context_enrichment          | GatedResidualNetwork            | 1.1 K
11 | lstm_encoder                       | LSTM                            | 2.2 K
12 | lstm_decoder                       | LSTM                            | 2.2 K
13 | post_lstm_gate_encoder             | GatedLinearUnit                 | 544
14 | post_lstm_add_norm_encoder         | AddNorm                         | 32
15 | static_enrichment                  | GatedResidualNetwork            | 1.4 K
16 | multihead_attn                     | InterpretableMultiHeadAttention | 676
17 | post_attn_gate_norm                | GateAddNorm                     | 576
18 | pos_wise_ff                        | GatedResidualNetwork            | 1.1 K
19 | pre_output_gate_norm               | GateAddNorm                     | 576
20 | output_layer                       | Linear                          | 119
----------------------------------------------------------------------------------------
18.9 K    Trainable params
0         Non-trainable params
18.9 K    Total params
0.075     Total estimated model params size (MB)

[Parallel(n_jobs=1)]: Done   3 out of   3 | elapsed: 14.7min remaining:    0.0s
[Parallel(n_jobs=1)]: Done   3 out of   3 | elapsed: 14.7min finished

metrics_tft

	segment	SMAPE	MAPE	MAE	fold_number
0	segment_a	4.003736	3.939512	20.710637	0
0	segment_a	6.366790	6.099132	31.872040	1
0	segment_a	2.728972	2.656444	15.460903	2
3	segment_b	5.884539	5.649675	14.540667	0
3	segment_b	5.377515	5.408940	12.965875	1
3	segment_b	4.069312	4.167871	9.468142	2
1	segment_c	3.423244	3.487631	5.776851	0
1	segment_c	3.581810	3.489848	6.513208	1
1	segment_c	6.113726	6.112950	10.985452	2
2	segment_d	9.495453	9.456682	79.086417	0
2	segment_d	3.744241	3.850278	32.086225	1
2	segment_d	2.347024	2.286240	20.012992	2

score = metrics_tft["SMAPE"].mean()
print(f"Average SMAPE for TFT: {score:.3f}")

Average SMAPE for TFT: 4.761

plot_backtest(forecast_tft, ts, history_len=20)

3.3 Simple model

For comparison let’s train a much more simpler model.

model_sma = SeasonalMovingAverageModel(window=5, seasonality=7)
linear_trend_transform = LinearTrendTransform(in_column='target')

pipeline_sma = Pipeline(model=model_sma,
                        horizon=HORIZON,
                        transforms=[linear_trend_transform])

metrics_sma, forecast_sma, fold_info_sma = pipeline_sma.backtest(ts, metrics=metrics, n_folds=3, n_jobs=1)

[Parallel(n_jobs=1)]: Using backend SequentialBackend with 1 concurrent workers.
[Parallel(n_jobs=1)]: Done   1 out of   1 | elapsed:    0.1s remaining:    0.0s
[Parallel(n_jobs=1)]: Done   2 out of   2 | elapsed:    0.2s remaining:    0.0s
[Parallel(n_jobs=1)]: Done   3 out of   3 | elapsed:    0.3s remaining:    0.0s
[Parallel(n_jobs=1)]: Done   3 out of   3 | elapsed:    0.3s finished

metrics_sma

	segment	SMAPE	MAPE	MAE	fold_number
0	segment_a	6.343943	6.124296	33.196532	0
0	segment_a	5.346946	5.192455	27.938101	1
0	segment_a	7.510347	7.189999	40.028565	2
3	segment_b	7.178822	6.920176	17.818102	0
3	segment_b	5.672504	5.554555	13.719200	1
3	segment_b	3.327846	3.359712	7.680919	2
1	segment_c	6.430429	6.200580	10.877718	0
1	segment_c	5.947090	5.727531	10.701336	1
1	segment_c	6.186545	5.943679	11.359563	2
2	segment_d	4.707899	4.644170	39.918646	0
2	segment_d	5.403426	5.600978	43.047332	1
2	segment_d	2.505279	2.543719	19.347565	2

score = metrics_sma["SMAPE"].mean()
print(f"Average SMAPE for Seasonal MA: {score:.3f}")

Average SMAPE for Seasonal MA: 5.547

plot_backtest(forecast_sma, ts, history_len=20)

As we can see, neural networks are a bit better in this particular case.