pytorch-ts/examples/m5.ipynb

In [17]:

import matplotlib.pyplot as plt
import json

In [2]:

import torch

In [3]:

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

In [4]:

from pts.dataset.repository import get_dataset
from pts.dataset.utils import to_pandas

In [5]:

dataset = get_dataset("m5", regenerate=False)

In [25]:

entry = next(iter(dataset.train))
train_series = to_pandas(entry)
train_series.plot()
plt.grid(which="both")
plt.legend(["train series"], loc="upper left")
plt.title(entry['item_id'])
plt.show()

In [24]:

entry = next(iter(dataset.test))
test_series = to_pandas(entry)
test_series.plot()
plt.axvline(train_series.index[-1], color='r') # end of train dataset
plt.grid(which="both")
plt.legend(["test series", "end of train series"], loc="upper left")
plt.title(entry['item_id'])
plt.show()

In [8]:

print(f"Recommended prediction horizon: {dataset.metadata.prediction_length}")
print(f"Frequency of the time series: {dataset.metadata.freq}")

Recommended prediction horizon: 28
Frequency of the time series: D

In [9]:

from pts.model.deepar import DeepAREstimator
from pts.modules import ZeroInflatedNegativeBinomialOutput
from pts import Trainer

In [10]:

estimator = DeepAREstimator(
    distr_output=ZeroInflatedNegativeBinomialOutput(),
    cell_type='GRU',
    input_size=72,
    num_cells=64,
    num_layers=3,
    dropout_rate=0.2,
    use_feat_dynamic_real=True,
    use_feat_static_cat=True,
    cardinality=[int(cat_feat_info.cardinality) for cat_feat_info in dataset.metadata.feat_static_cat],
    embedding_dimension = [4, 4, 4, 4, 16],
    prediction_length=dataset.metadata.prediction_length,
    context_length=dataset.metadata.prediction_length*2,
    freq=dataset.metadata.freq,
    scaling=True,
    trainer=Trainer(device=device,
                    epochs=50,
                    learning_rate=1e-3,
                    num_batches_per_epoch=120,
                    batch_size=256,
                    num_workers=8,
                    pin_memory=True,
                   )
)

In [11]:

predictor = estimator.train(dataset.train)

119it [00:29,  4.09it/s, avg_epoch_loss=1.17, epoch=0]
119it [00:31,  3.81it/s, avg_epoch_loss=1.14, epoch=1]
119it [00:27,  4.29it/s, avg_epoch_loss=1.12, epoch=2]
119it [00:29,  4.06it/s, avg_epoch_loss=1.11, epoch=3]
119it [00:28,  4.17it/s, avg_epoch_loss=1.1, epoch=4]
119it [00:28,  4.14it/s, avg_epoch_loss=1.1, epoch=5]
119it [00:30,  3.93it/s, avg_epoch_loss=1.1, epoch=6] 
119it [00:27,  4.27it/s, avg_epoch_loss=1.11, epoch=7]
119it [00:28,  4.16it/s, avg_epoch_loss=1.09, epoch=8]
119it [00:29,  4.09it/s, avg_epoch_loss=1.11, epoch=9]
119it [00:27,  4.40it/s, avg_epoch_loss=1.1, epoch=10]
119it [00:28,  4.23it/s, avg_epoch_loss=1.1, epoch=11]
119it [00:28,  4.12it/s, avg_epoch_loss=1.1, epoch=12] 
119it [00:29,  4.06it/s, avg_epoch_loss=1.1, epoch=13]
119it [00:29,  4.10it/s, avg_epoch_loss=1.11, epoch=14]
119it [00:28,  4.24it/s, avg_epoch_loss=1.1, epoch=15] 
119it [00:30,  3.95it/s, avg_epoch_loss=1.1, epoch=16]
119it [00:27,  4.28it/s, avg_epoch_loss=1.09, epoch=17]
119it [00:27,  4.26it/s, avg_epoch_loss=1.1, epoch=18]
119it [00:29,  4.07it/s, avg_epoch_loss=1.1, epoch=19] 
119it [00:29,  3.98it/s, avg_epoch_loss=1.09, epoch=20]
119it [00:27,  4.33it/s, avg_epoch_loss=1.1, epoch=21]
119it [00:29,  4.08it/s, avg_epoch_loss=1.09, epoch=22]
119it [00:29,  4.09it/s, avg_epoch_loss=1.09, epoch=23]
119it [00:28,  4.22it/s, avg_epoch_loss=1.09, epoch=24]
119it [00:27,  4.26it/s, avg_epoch_loss=1.09, epoch=25]
119it [00:31,  3.81it/s, avg_epoch_loss=1.1, epoch=26]
119it [00:31,  3.73it/s, avg_epoch_loss=1.09, epoch=27]
119it [00:27,  4.32it/s, avg_epoch_loss=1.08, epoch=28]
119it [00:28,  4.14it/s, avg_epoch_loss=1.09, epoch=29]
119it [00:30,  3.87it/s, avg_epoch_loss=1.08, epoch=30]
119it [00:28,  4.19it/s, avg_epoch_loss=1.09, epoch=31]
119it [00:28,  4.17it/s, avg_epoch_loss=1.08, epoch=32]
119it [00:29,  4.09it/s, avg_epoch_loss=1.1, epoch=33] 
119it [00:27,  4.39it/s, avg_epoch_loss=1.09, epoch=34]
119it [00:28,  4.21it/s, avg_epoch_loss=1.09, epoch=35]
119it [00:28,  4.16it/s, avg_epoch_loss=1.09, epoch=36]
119it [00:27,  4.31it/s, avg_epoch_loss=1.08, epoch=37]
119it [00:29,  4.07it/s, avg_epoch_loss=1.09, epoch=38]
119it [00:28,  4.19it/s, avg_epoch_loss=1.09, epoch=39]
119it [00:29,  4.06it/s, avg_epoch_loss=1.09, epoch=40]
119it [00:28,  4.14it/s, avg_epoch_loss=1.08, epoch=41]
119it [00:28,  4.16it/s, avg_epoch_loss=1.09, epoch=42]
119it [00:27,  4.25it/s, avg_epoch_loss=1.09, epoch=43]
119it [00:27,  4.26it/s, avg_epoch_loss=1.1, epoch=44]
119it [00:26,  4.41it/s, avg_epoch_loss=1.09, epoch=45]
119it [00:27,  4.25it/s, avg_epoch_loss=1.08, epoch=46]
119it [00:28,  4.20it/s, avg_epoch_loss=1.09, epoch=47]
119it [00:30,  3.92it/s, avg_epoch_loss=1.09, epoch=48]
119it [00:30,  3.96it/s, avg_epoch_loss=1.09, epoch=49]

In [12]:

from pts.evaluation import make_evaluation_predictions, Evaluator

In [13]:

forecast_it, ts_it = make_evaluation_predictions(
    dataset=dataset.test,  # test dataset
    predictor=predictor,  # predictor
    num_samples=100,  # number of sample paths we want for evaluation
)

In [14]:

forecasts = list(forecast_it)
tss = list(ts_it)

In [15]:

evaluator = Evaluator()
agg_metrics, item_metrics = evaluator(iter(tss), iter(forecasts), num_series=len(dataset.test))

Running evaluation: 100%|██████████| 30490/30490 [00:01<00:00, 23152.64it/s]

In [18]:

print(json.dumps(agg_metrics, indent=4))

{
    "MSE": 4.439620313754265,
    "abs_error": 807089.0,
    "abs_target_sum": 1231764.0,
    "abs_target_mean": 1.4428196598416343,
    "seasonal_error": 1.1272178349378457,
    "MASE": 0.8789472000957106,
    "MAPE": 0.30587227335898637,
    "sMAPE": 0.6816909686747539,
    "OWA": NaN,
    "MSIS": 7.28829495943688,
    "QuantileLoss[0.1]": 228315.8,
    "Coverage[0.1]": 0.0042929766199690765,
    "QuantileLoss[0.2]": 422650.8,
    "Coverage[0.2]": 0.01732535257461463,
    "QuantileLoss[0.3]": 586642.4,
    "Coverage[0.3]": 0.042479970013587595,
    "QuantileLoss[0.4]": 716891.6,
    "Coverage[0.4]": 0.08317012603663966,
    "QuantileLoss[0.5]": 807089.0,
    "Coverage[0.5]": 0.14288174108607035,
    "QuantileLoss[0.6]": 854345.2,
    "Coverage[0.6]": 0.2176439582064377,
    "QuantileLoss[0.7]": 842037.0,
    "Coverage[0.7]": 0.3306306517359322,
    "QuantileLoss[0.8]": 755156.7999999999,
    "Coverage[0.8]": 0.48669352949444783,
    "QuantileLoss[0.9]": 547328.7999999999,
    "Coverage[0.9]": 0.7026999484608537,
    "RMSE": 2.1070406530853325,
    "NRMSE": 1.4603631429007586,
    "ND": 0.655230222672525,
    "wQuantileLoss[0.1]": 0.185356772888313,
    "wQuantileLoss[0.2]": 0.34312644305240286,
    "wQuantileLoss[0.3]": 0.47626201122942385,
    "wQuantileLoss[0.4]": 0.5820040202506324,
    "wQuantileLoss[0.5]": 0.655230222672525,
    "wQuantileLoss[0.6]": 0.6935948769407126,
    "wQuantileLoss[0.7]": 0.6836025407464417,
    "wQuantileLoss[0.8]": 0.6130693866682253,
    "wQuantileLoss[0.9]": 0.44434550774336634,
    "mean_wQuantileLoss": 0.5196213091324493,
    "MAE_Coverage": 0.2746868606412719
}

In [19]:

item_metrics.plot(x='MSIS', y='MASE', kind='scatter')
plt.grid(which="both")
plt.show()

In [ ]: