DeepAR model¶

In [1]:

#!pip install deepts_forecasting

#!pip install deepts_forecasting

Import libraries¶

In [2]:

import numpy as np
import torch
from torch.utils.data import DataLoader
import pytorch_lightning as pl
from pytorch_lightning.callbacks import EarlyStopping, LearningRateMonitor
from pytorch_lightning.loggers import TensorBoardLogger

from deepts_forecasting.utils.data import TimeSeriesDataSet
from deepts_forecasting.utils.data.encoders import TorchNormalizer
from deepts_forecasting.datasets import AirPassengersDataset
from deepts_forecasting.models.deepar import DeepAR

import numpy as np import torch from torch.utils.data import DataLoader import pytorch_lightning as pl from pytorch_lightning.callbacks import EarlyStopping, LearningRateMonitor from pytorch_lightning.loggers import TensorBoardLogger from deepts_forecasting.utils.data import TimeSeriesDataSet from deepts_forecasting.utils.data.encoders import TorchNormalizer from deepts_forecasting.datasets import AirPassengersDataset from deepts_forecasting.models.deepar import DeepAR

D:\Anaconda3\envs\DeepTS_Forecasting\lib\site-packages\tqdm\auto.py:22: TqdmWarning: IProgress not found. Please update jupyter and ipywidgets. See https://ipywidgets.readthedocs.io/en/stable/user_install.html
  from .autonotebook import tqdm as notebook_tqdm

Dataset¶

In [3]:

data = AirPassengersDataset().load()
data['year'] = data['Month'].dt.year
data['month'] = data['Month'].dt.month
data['group'] = '0'
data['time_idx'] = np.arange(len(data))
data['Passengers'] = data['Passengers'].astype(float)
data['month'] = data['month'].astype('str')
data.head()

data = AirPassengersDataset().load() data['year'] = data['Month'].dt.year data['month'] = data['Month'].dt.month data['group'] = '0' data['time_idx'] = np.arange(len(data)) data['Passengers'] = data['Passengers'].astype(float) data['month'] = data['month'].astype('str') data.head()

Out[3]:

	Month	Passengers	year	month	group	time_idx
0	1949-01-01	112.0	1949	1	0	0
1	1949-02-01	118.0	1949	2	0	1
2	1949-03-01	132.0	1949	3	0	2
3	1949-04-01	129.0	1949	4	0	3
4	1949-05-01	121.0	1949	5	0	4

Split train/test sets¶

In [4]:

max_encoder_length = 18
max_prediction_length = 12

training_cutoff = data["time_idx"].max() - max_encoder_length - max_prediction_length

training = TimeSeriesDataSet(
    data[lambda x: x.time_idx <= training_cutoff],
    max_encoder_length= max_encoder_length,
    min_encoder_length=max_encoder_length,
    max_prediction_length=max_prediction_length,
    min_prediction_length=max_prediction_length,
    time_idx="time_idx",
    target="Passengers",
    group_ids=["group"],
    static_categoricals=[],
    static_reals=[],
    time_varying_known_categoricals=['month'],
    time_varying_known_reals=[],
    time_varying_unknown_reals=["Passengers"],
    time_varying_unknown_categoricals=[],
    target_normalizer=TorchNormalizer(method="standard",
                                      transformation=None),
    )

training.get_parameters()
validation = TimeSeriesDataSet.from_dataset(training,
                                            data[lambda x: x.time_idx > training_cutoff])

batch_size = 16
train_dataloader = DataLoader(training, batch_size=batch_size, shuffle=False, drop_last=False)
val_dataloader = DataLoader(validation, batch_size=batch_size, shuffle=False, drop_last=False)

max_encoder_length = 18 max_prediction_length = 12 training_cutoff = data["time_idx"].max() - max_encoder_length - max_prediction_length training = TimeSeriesDataSet( data[lambda x: x.time_idx <= training_cutoff], max_encoder_length= max_encoder_length, min_encoder_length=max_encoder_length, max_prediction_length=max_prediction_length, min_prediction_length=max_prediction_length, time_idx="time_idx", target="Passengers", group_ids=["group"], static_categoricals=[], static_reals=[], time_varying_known_categoricals=['month'], time_varying_known_reals=[], time_varying_unknown_reals=["Passengers"], time_varying_unknown_categoricals=[], target_normalizer=TorchNormalizer(method="standard", transformation=None), ) training.get_parameters() validation = TimeSeriesDataSet.from_dataset(training, data[lambda x: x.time_idx > training_cutoff]) batch_size = 16 train_dataloader = DataLoader(training, batch_size=batch_size, shuffle=False, drop_last=False) val_dataloader = DataLoader(validation, batch_size=batch_size, shuffle=False, drop_last=False)

Define model¶

In [5]:

pl.seed_everything(123)
# create PyTorch Lighning Trainer with early stopping
early_stop_callback = EarlyStopping(monitor="val_loss", min_delta=1e-4,
                                    patience=60, verbose=False, mode="min")
lr_logger = LearningRateMonitor()

trainer = pl.Trainer(
    max_epochs=300,
    gpus=0,  # run on CPU, if on multiple GPUs, use accelerator="ddp"
    gradient_clip_val=0.1,
    limit_train_batches=30,  # 30 batches per epoch
    callbacks=[lr_logger, early_stop_callback],
    logger=TensorBoardLogger("lightning_logs")
)

model = DeepAR.from_dataset(training,
                            hidden_size=32,
                            rnn_layers=3,
                                 )
model.summarize

pl.seed_everything(123) # create PyTorch Lighning Trainer with early stopping early_stop_callback = EarlyStopping(monitor="val_loss", min_delta=1e-4, patience=60, verbose=False, mode="min") lr_logger = LearningRateMonitor() trainer = pl.Trainer( max_epochs=300, gpus=0, # run on CPU, if on multiple GPUs, use accelerator="ddp" gradient_clip_val=0.1, limit_train_batches=30, # 30 batches per epoch callbacks=[lr_logger, early_stop_callback], logger=TensorBoardLogger("lightning_logs") ) model = DeepAR.from_dataset(training, hidden_size=32, rnn_layers=3, ) model.summarize

Global seed set to 123
GPU available: False, used: False
TPU available: False, using: 0 TPU cores
IPU available: False, using: 0 IPUs

Out[5]:

<bound method LightningModule.summarize of DeepAR(
  (loss): NormalDistributionLoss()
  (logging_metrics): ModuleList()
  (mu_layer): Linear(in_features=32, out_features=1, bias=True)
  (sigma_layer): Linear(in_features=32, out_features=1, bias=True)
  (activation): Softplus(beta=1, threshold=20)
  (encode_rnn): LSTM(7, 32, num_layers=3, batch_first=True)
  (decode_rnn): LSTM(6, 32, num_layers=3, batch_first=True)
  (embeddings): ModuleDict(
    (month): Embedding(12, 6)
  )
)>

In [6]:

model.hparams

model.hparams

Out[6]:

"categorical_groups":                {}
"embedding_labels":                  {'month': array(['1', '10', '11', '12', '2', '3', '4', '5', '6', '7', '8', '9'],
      dtype=object)}
"embedding_paddings":                []
"embedding_sizes":                   {'month': [12, 6]}
"hidden_size":                       32
"learning_rate":                     0.001
"log_interval":                      -1
"log_val_interval":                  None
"logging_metrics":                   ModuleList()
"loss":                              NormalDistributionLoss()
"max_encoder_length":                18
"max_prediction_length":             12
"monotone_constaints":               {}
"output_transformer":                TorchNormalizer()
"rnn_layers":                        3
"static_categoricals":               []
"static_reals":                      []
"time_varying_categoricals_decoder": ['month']
"time_varying_categoricals_encoder": ['month']
"time_varying_reals_decoder":        []
"time_varying_reals_encoder":        ['Passengers']
"x_categoricals":                    ['month']
"x_reals":                           ['Passengers']

Train model with early stopping¶

In [7]:

trainer.fit(
    model, train_dataloader=train_dataloader, val_dataloaders=val_dataloader,
)

# (given that we use early stopping, this is not necessarily the last epoch)
best_model_path = trainer.checkpoint_callback.best_model_path
best_model = DeepAR.load_from_checkpoint(best_model_path)

# calcualte mean absolute error on validation set
actuals = torch.cat([model.transform_output(prediction=y, target_scale=x['target_scale'])
                     for x, y in iter(val_dataloader)])
predictions = best_model.predict(val_dataloader)
mae = (actuals - predictions).abs().mean()
# print('predictions shape is:', predictions.shape)
# print('actuals shape is:', actuals.shape)
print(torch.cat([actuals, predictions]))
print('MAE is:', mae)

trainer.fit( model, train_dataloader=train_dataloader, val_dataloaders=val_dataloader, ) # (given that we use early stopping, this is not necessarily the last epoch) best_model_path = trainer.checkpoint_callback.best_model_path best_model = DeepAR.load_from_checkpoint(best_model_path) # calcualte mean absolute error on validation set actuals = torch.cat([model.transform_output(prediction=y, target_scale=x['target_scale']) for x, y in iter(val_dataloader)]) predictions = best_model.predict(val_dataloader) mae = (actuals - predictions).abs().mean() # print('predictions shape is:', predictions.shape) # print('actuals shape is:', actuals.shape) print(torch.cat([actuals, predictions])) print('MAE is:', mae)

D:\Anaconda3\envs\DeepTS_Forecasting\lib\site-packages\pytorch_lightning\trainer\trainer.py:735: LightningDeprecationWarning: `trainer.fit(train_dataloader)` is deprecated in v1.4 and will be removed in v1.6. Use `trainer.fit(train_dataloaders)` instead. HINT: added 's'
  rank_zero_deprecation(

  | Name            | Type                   | Params
-----------------------------------------------------------
0 | loss            | NormalDistributionLoss | 0     
1 | logging_metrics | ModuleList             | 0     
2 | mu_layer        | Linear                 | 33    
3 | sigma_layer     | Linear                 | 33    
4 | activation      | Softplus               | 0     
5 | encode_rnn      | LSTM                   | 22.1 K
6 | decode_rnn      | LSTM                   | 22.0 K
7 | embeddings      | ModuleDict             | 72    
-----------------------------------------------------------
44.3 K    Trainable params
0         Non-trainable params
44.3 K    Total params
0.177     Total estimated model params size (MB)

D:\Anaconda3\envs\DeepTS_Forecasting\lib\site-packages\pytorch_lightning\trainer\data_loading.py:132: UserWarning: The dataloader, val_dataloader 0, does not have many workers which may be a bottleneck. Consider increasing the value of the `num_workers` argument` (try 16 which is the number of cpus on this machine) in the `DataLoader` init to improve performance.
  rank_zero_warn(
Global seed set to 123
D:\Anaconda3\envs\DeepTS_Forecasting\lib\site-packages\pytorch_lightning\trainer\data_loading.py:132: UserWarning: The dataloader, train_dataloader, does not have many workers which may be a bottleneck. Consider increasing the value of the `num_workers` argument` (try 16 which is the number of cpus on this machine) in the `DataLoader` init to improve performance.
  rank_zero_warn(
D:\Anaconda3\envs\DeepTS_Forecasting\lib\site-packages\pytorch_lightning\trainer\data_loading.py:432: UserWarning: The number of training samples (6) is smaller than the logging interval Trainer(log_every_n_steps=50). Set a lower value for log_every_n_steps if you want to see logs for the training epoch.
  rank_zero_warn(

Epoch 75: 100%|██████████| 7/7 [00:00<00:00, 29.41it/s, loss=0.0683, v_num=8, val_loss=19.80, train_loss=0.586]   
tensor([[[417.0000],
         [391.0000],
         [419.0000],
         [461.0000],
         [472.0000],
         [535.0000],
         [622.0000],
         [606.0000],
         [508.0000],
         [461.0000],
         [390.0000],
         [432.0000]],

        [[392.7820],
         [405.8823],
         [433.1111],
         [421.9375],
         [418.5957],
         [468.5001],
         [492.4781],
         [481.0977],
         [422.1546],
         [357.6677],
         [326.4588],
         [337.1531]]], dtype=torch.float64)
MAE is: tensor(67.8474, dtype=torch.float64)