Transformer 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.transformer import TransformerModel

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.transformer import TransformerModel

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 [7]:

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 = TransformerModel.from_dataset(training,
                                      dim_feedforward=32,
                                      n_head=1,
                                      n_layers=2,
                                      d_model=16,
                                      )
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 = TransformerModel.from_dataset(training, dim_feedforward=32, n_head=1, n_layers=2, d_model=16, ) 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[7]:

<bound method LightningModule.summarize of TransformerModel(
  (loss): L1Loss()
  (logging_metrics): ModuleList()
  (encoder_input_linear): Linear(in_features=7, out_features=16, bias=True)
  (decoder_input_linear): Linear(in_features=6, out_features=16, bias=True)
  (encoder_positional_encoding): PositionalEncoding(
    (dropout): Dropout(p=0.1, inplace=False)
  )
  (decoder_positional_encoding): PositionalEncoding(
    (dropout): Dropout(p=0.1, inplace=False)
  )
  (transformer_encoder): TransformerEncoder(
    (layers): ModuleList(
      (0): TransformerEncoderLayer(
        (self_attn): MultiheadAttention(
          (out_proj): _LinearWithBias(in_features=16, out_features=16, bias=True)
        )
        (linear1): Linear(in_features=16, out_features=32, bias=True)
        (dropout): Dropout(p=0.1, inplace=False)
        (linear2): Linear(in_features=32, out_features=16, bias=True)
        (norm1): LayerNorm((16,), eps=1e-05, elementwise_affine=True)
        (norm2): LayerNorm((16,), eps=1e-05, elementwise_affine=True)
        (dropout1): Dropout(p=0.1, inplace=False)
        (dropout2): Dropout(p=0.1, inplace=False)
      )
      (1): TransformerEncoderLayer(
        (self_attn): MultiheadAttention(
          (out_proj): _LinearWithBias(in_features=16, out_features=16, bias=True)
        )
        (linear1): Linear(in_features=16, out_features=32, bias=True)
        (dropout): Dropout(p=0.1, inplace=False)
        (linear2): Linear(in_features=32, out_features=16, bias=True)
        (norm1): LayerNorm((16,), eps=1e-05, elementwise_affine=True)
        (norm2): LayerNorm((16,), eps=1e-05, elementwise_affine=True)
        (dropout1): Dropout(p=0.1, inplace=False)
        (dropout2): Dropout(p=0.1, inplace=False)
      )
    )
    (norm): LayerNorm((16,), eps=1e-05, elementwise_affine=True)
  )
  (transformer_decoder): TransformerDecoder(
    (layers): ModuleList(
      (0): TransformerDecoderLayer(
        (self_attn): MultiheadAttention(
          (out_proj): _LinearWithBias(in_features=16, out_features=16, bias=True)
        )
        (multihead_attn): MultiheadAttention(
          (out_proj): _LinearWithBias(in_features=16, out_features=16, bias=True)
        )
        (linear1): Linear(in_features=16, out_features=32, bias=True)
        (dropout): Dropout(p=0.1, inplace=False)
        (linear2): Linear(in_features=32, out_features=16, bias=True)
        (norm1): LayerNorm((16,), eps=1e-05, elementwise_affine=True)
        (norm2): LayerNorm((16,), eps=1e-05, elementwise_affine=True)
        (norm3): LayerNorm((16,), eps=1e-05, elementwise_affine=True)
        (dropout1): Dropout(p=0.1, inplace=False)
        (dropout2): Dropout(p=0.1, inplace=False)
        (dropout3): Dropout(p=0.1, inplace=False)
      )
      (1): TransformerDecoderLayer(
        (self_attn): MultiheadAttention(
          (out_proj): _LinearWithBias(in_features=16, out_features=16, bias=True)
        )
        (multihead_attn): MultiheadAttention(
          (out_proj): _LinearWithBias(in_features=16, out_features=16, bias=True)
        )
        (linear1): Linear(in_features=16, out_features=32, bias=True)
        (dropout): Dropout(p=0.1, inplace=False)
        (linear2): Linear(in_features=32, out_features=16, bias=True)
        (norm1): LayerNorm((16,), eps=1e-05, elementwise_affine=True)
        (norm2): LayerNorm((16,), eps=1e-05, elementwise_affine=True)
        (norm3): LayerNorm((16,), eps=1e-05, elementwise_affine=True)
        (dropout1): Dropout(p=0.1, inplace=False)
        (dropout2): Dropout(p=0.1, inplace=False)
        (dropout3): Dropout(p=0.1, inplace=False)
      )
    )
    (norm): LayerNorm((16,), eps=1e-05, elementwise_affine=True)
  )
  (out_linear): Linear(in_features=16, out_features=1, bias=True)
  (embeddings): ModuleDict(
    (month): Embedding(12, 6)
  )
)>

In [8]:

model.hparams

model.hparams

Out[8]:

"activation":                        relu
"categorical_groups":                {}
"d_model":                           16
"dim_feedforward":                   32
"dropout":                           0.1
"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]}
"learning_rate":                     0.001
"log_interval":                      -1
"log_val_interval":                  None
"logging_metrics":                   ModuleList()
"loss":                              L1Loss()
"max_encoder_length":                18
"max_prediction_length":             12
"monotone_constaints":               {}
"n_head":                            1
"n_layers":                          2
"output_size":                       1
"output_transformer":                TorchNormalizer()
"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 [10]:

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 = TransformerModel.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, x_index = 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 = TransformerModel.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, x_index = 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)

  | Name                        | Type               | Params
-------------------------------------------------------------------
0 | loss                        | L1Loss             | 0     
1 | logging_metrics             | ModuleList         | 0     
2 | encoder_input_linear        | Linear             | 128   
3 | decoder_input_linear        | Linear             | 112   
4 | encoder_positional_encoding | PositionalEncoding | 0     
5 | decoder_positional_encoding | PositionalEncoding | 0     
6 | transformer_encoder         | TransformerEncoder | 4.5 K 
7 | transformer_decoder         | TransformerDecoder | 6.7 K 
8 | out_linear                  | Linear             | 17    
9 | embeddings                  | ModuleDict         | 72    
-------------------------------------------------------------------
11.5 K    Trainable params
0         Non-trainable params
11.5 K    Total params
0.046     Total estimated model params size (MB)
D:\Anaconda3\envs\DeepTS_Forecasting\lib\site-packages\pytorch_lightning\callbacks\model_checkpoint.py:631: UserWarning: Checkpoint directory lightning_logs\default\version_17\checkpoints exists and is not empty.
  rank_zero_warn(f"Checkpoint directory {dirpath} exists and is not empty.")

Global seed set to 123

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]],

        [[350.6286],
         [331.8389],
         [368.7483],
         [367.6466],
         [376.8046],
         [437.5001],
         [456.6611],
         [470.3661],
         [442.6349],
         [358.7311],
         [334.2704],
         [360.6251]]], dtype=torch.float64)
MAE is: tensor(88.1287, dtype=torch.float64)