tmp: m4

baselines/MLP/MASE.py

@@ -0,0 +1,20 @@
+import numpy as np
+import torch
+
+def mase(prediction: torch.Tensor, inputs:torch.Tensor, target: torch.Tensor, frequency: int = 1, null_val: float = np.nan) -> np.ndarray:
+    """
+    MASE loss as defined in "Scaled Errors" https://robjhyndman.com/papers/mase.pdf
+
+    :param forecast: Forecast values. Shape: batch, time_o
+    :param insample: Insample values. Shape: batch, time_i
+    :param outsample: Target values. Shape: batch, time_o
+    :param frequency: Frequency value
+    :return: Same shape array with error calculated for each time step
+    """
+    prediction = prediction.detach().cpu().numpy()
+    target = target.detach().cpu().numpy()
+    inputs = inputs.detach().cpu().numpy()
+    a = np.mean(np.abs(prediction - target))
+    b = np.mean(np.abs(inputs[:-frequency] - inputs[frequency:]))
+    result = torch.Tensor([a / b])
+    return result

baselines/MLP/MLP_M4_Monthly.py

@@ -0,0 +1,114 @@
+import os
+import sys
+
+# TODO: remove it when basicts can be installed by pip
+sys.path.append(os.path.abspath(__file__ + "/../../.."))
+from easydict import EasyDict
+from basicts.runners import M4Runner
+from basicts.losses import masked_mae
+from basicts.data import M4ForecastingDataset
+from basicts.utils import partial
+
+from .mlp_arch import MultiLayerPerceptron
+from .MASE import mase
+
+seasonal_pattern = "Monthly"
+
+frequence = {"Yearly": 1, "Quarterly": 4, "Monthly": 12, "Weekly": 1, "Daily": 1, "Hourly": 24}[seasonal_pattern]
+lookback_size = 2 # [2, 3, 4, 5, 6, 7]
+CFG = EasyDict()
+
+# ================= general ================= #
+CFG.DESCRIPTION = "Multi-layer perceptron model configuration"
+CFG.RUNNER = M4Runner
+CFG.DATASET_CLS = M4ForecastingDataset
+CFG.DATASET_NAME = "M4_" + seasonal_pattern
+CFG.DATASET_OUTPUT_LEN = {"Yearly": 6, "Quarterly": 8, "Monthly": 18, "Weekly": 13, "Daily": 14, "Hourly": 48}[seasonal_pattern]
+CFG.DATASET_INPUT_LEN = lookback_size * CFG.DATASET_OUTPUT_LEN
+CFG.GPU_NUM = 1
+CFG.METRICS = {"MASE": partial(mase, frequency=frequence)}
+CFG.TARGET_METRICS = "MASE"
+
+# ================= environment ================= #
+CFG.ENV = EasyDict()
+CFG.ENV.SEED = 1
+CFG.ENV.CUDNN = EasyDict()
+CFG.ENV.CUDNN.ENABLED = True
+
+# ================= model ================= #
+CFG.MODEL = EasyDict()
+CFG.MODEL.NAME = "MultiLayerPerceptron"
+CFG.MODEL.ARCH = MultiLayerPerceptron
+CFG.MODEL.PARAM = {
+    "history_seq_len": CFG.DATASET_INPUT_LEN,
+    "prediction_seq_len": CFG.DATASET_OUTPUT_LEN,
+    "hidden_dim": 32
+}
+CFG.MODEL.FORWARD_FEATURES = [0]
+CFG.MODEL.TARGET_FEATURES = [0]
+
+# ================= optim ================= #
+CFG.TRAIN = EasyDict()
+CFG.TRAIN.LOSS = masked_mae
+CFG.TRAIN.OPTIM = EasyDict()
+CFG.TRAIN.OPTIM.TYPE = "Adam"
+CFG.TRAIN.OPTIM.PARAM = {
+    "lr": 0.002,
+    "weight_decay": 1.0e-5,
+    "eps": 1.0e-8
+}
+CFG.TRAIN.LR_SCHEDULER = EasyDict()
+CFG.TRAIN.LR_SCHEDULER.TYPE = "MultiStepLR"
+CFG.TRAIN.LR_SCHEDULER.PARAM = {
+    "milestones": [1, 30, 38, 46, 54, 62, 70, 80],
+    "gamma": 0.5
+}
+
+# ================= train ================= #
+CFG.TRAIN.CLIP_GRAD_PARAM = {
+    "max_norm": 5.0
+}
+CFG.TRAIN.NUM_EPOCHS = 100
+CFG.TRAIN.CKPT_SAVE_DIR = os.path.join(
+    "checkpoints",
+    "_".join([CFG.MODEL.NAME, str(CFG.TRAIN.NUM_EPOCHS)])
+)
+# train data
+CFG.TRAIN.DATA = EasyDict()
+# read data
+CFG.TRAIN.DATA.DIR = "datasets/" + CFG.DATASET_NAME
+# dataloader args, optional
+CFG.TRAIN.DATA.BATCH_SIZE = 32
+CFG.TRAIN.DATA.PREFETCH = False
+CFG.TRAIN.DATA.SHUFFLE = True
+CFG.TRAIN.DATA.NUM_WORKERS = 2
+CFG.TRAIN.DATA.PIN_MEMORY = False
+
+# ================= validate ================= #
+CFG.VAL = EasyDict()
+CFG.VAL.INTERVAL = 1
+# validating data
+CFG.VAL.DATA = EasyDict()
+# read data
+CFG.VAL.DATA.DIR = "datasets/" + CFG.DATASET_NAME
+# dataloader args, optional
+CFG.VAL.DATA.BATCH_SIZE = 32
+CFG.VAL.DATA.PREFETCH = False
+CFG.VAL.DATA.SHUFFLE = False
+CFG.VAL.DATA.NUM_WORKERS = 2
+CFG.VAL.DATA.PIN_MEMORY = False
+
+# ================= test ================= #
+CFG.TEST = EasyDict()
+CFG.TEST.INTERVAL = 1
+# evluation
+# test data
+CFG.TEST.DATA = EasyDict()
+# read data
+CFG.TEST.DATA.DIR = "datasets/" + CFG.DATASET_NAME
+# dataloader args, optional
+CFG.TEST.DATA.BATCH_SIZE = 32
+CFG.TEST.DATA.PREFETCH = False
+CFG.TEST.DATA.SHUFFLE = False
+CFG.TEST.DATA.NUM_WORKERS = 2
+CFG.TEST.DATA.PIN_MEMORY = False

baselines/MLP/MLP_M4_Weekly.py

@@ -0,0 +1,113 @@
+import os
+import sys
+
+# TODO: remove it when basicts can be installed by pip
+sys.path.append(os.path.abspath(__file__ + "/../../.."))
+from easydict import EasyDict
+from basicts.runners import M4Runner
+from basicts.losses import masked_mae
+from basicts.data import M4ForecastingDataset
+from basicts.utils import partial
+
+from .mlp_arch import MultiLayerPerceptron
+from .MASE import mase
+
+seasonal_pattern = "Weekly"
+
+frequence = {"Yearly": 1, "Quarterly": 4, "Monthly": 12, "Weekly": 1, "Daily": 1, "Hourly": 24}[seasonal_pattern]
+CFG = EasyDict()
+
+# ================= general ================= #
+CFG.DESCRIPTION = "Multi-layer perceptron model configuration"
+CFG.RUNNER = M4Runner
+CFG.DATASET_CLS = M4ForecastingDataset
+CFG.DATASET_NAME = "M4_" + seasonal_pattern
+CFG.DATASET_INPUT_LEN = 26
+CFG.DATASET_OUTPUT_LEN = {"Yearly": 6, "Quarterly": 8, "Monthly": 18, "Weekly": 13, "Daily": 14, "Hourly": 48}[seasonal_pattern]
+CFG.GPU_NUM = 1
+CFG.METRICS = {"MASE": partial(mase, frequency=frequence)}
+CFG.TARGET_METRICS = "MASE"
+
+# ================= environment ================= #
+CFG.ENV = EasyDict()
+CFG.ENV.SEED = 1
+CFG.ENV.CUDNN = EasyDict()
+CFG.ENV.CUDNN.ENABLED = True
+
+# ================= model ================= #
+CFG.MODEL = EasyDict()
+CFG.MODEL.NAME = "MultiLayerPerceptron"
+CFG.MODEL.ARCH = MultiLayerPerceptron
+CFG.MODEL.PARAM = {
+    "history_seq_len": CFG.DATASET_INPUT_LEN,
+    "prediction_seq_len": CFG.DATASET_OUTPUT_LEN,
+    "hidden_dim": 32
+}
+CFG.MODEL.FORWARD_FEATURES = [0]
+CFG.MODEL.TARGET_FEATURES = [0]
+
+# ================= optim ================= #
+CFG.TRAIN = EasyDict()
+CFG.TRAIN.LOSS = masked_mae
+CFG.TRAIN.OPTIM = EasyDict()
+CFG.TRAIN.OPTIM.TYPE = "Adam"
+CFG.TRAIN.OPTIM.PARAM = {
+    "lr": 0.002,
+    "weight_decay": 1.0e-5,
+    "eps": 1.0e-8
+}
+CFG.TRAIN.LR_SCHEDULER = EasyDict()
+CFG.TRAIN.LR_SCHEDULER.TYPE = "MultiStepLR"
+CFG.TRAIN.LR_SCHEDULER.PARAM = {
+    "milestones": [1, 30, 38, 46, 54, 62, 70, 80],
+    "gamma": 0.5
+}
+
+# ================= train ================= #
+CFG.TRAIN.CLIP_GRAD_PARAM = {
+    "max_norm": 5.0
+}
+CFG.TRAIN.NUM_EPOCHS = 100
+CFG.TRAIN.CKPT_SAVE_DIR = os.path.join(
+    "checkpoints",
+    "_".join([CFG.MODEL.NAME, str(CFG.TRAIN.NUM_EPOCHS)])
+)
+# train data
+CFG.TRAIN.DATA = EasyDict()
+# read data
+CFG.TRAIN.DATA.DIR = "datasets/" + CFG.DATASET_NAME
+# dataloader args, optional
+CFG.TRAIN.DATA.BATCH_SIZE = 32
+CFG.TRAIN.DATA.PREFETCH = False
+CFG.TRAIN.DATA.SHUFFLE = True
+CFG.TRAIN.DATA.NUM_WORKERS = 2
+CFG.TRAIN.DATA.PIN_MEMORY = False
+
+# ================= validate ================= #
+CFG.VAL = EasyDict()
+CFG.VAL.INTERVAL = 1
+# validating data
+CFG.VAL.DATA = EasyDict()
+# read data
+CFG.VAL.DATA.DIR = "datasets/" + CFG.DATASET_NAME
+# dataloader args, optional
+CFG.VAL.DATA.BATCH_SIZE = 32
+CFG.VAL.DATA.PREFETCH = False
+CFG.VAL.DATA.SHUFFLE = False
+CFG.VAL.DATA.NUM_WORKERS = 2
+CFG.VAL.DATA.PIN_MEMORY = False
+
+# ================= test ================= #
+CFG.TEST = EasyDict()
+CFG.TEST.INTERVAL = 1
+# evluation
+# test data
+CFG.TEST.DATA = EasyDict()
+# read data
+CFG.TEST.DATA.DIR = "datasets/" + CFG.DATASET_NAME
+# dataloader args, optional
+CFG.TEST.DATA.BATCH_SIZE = 32
+CFG.TEST.DATA.PREFETCH = False
+CFG.TEST.DATA.SHUFFLE = False
+CFG.TEST.DATA.NUM_WORKERS = 2
+CFG.TEST.DATA.PIN_MEMORY = False

basicts/__init__.py

@@ -1,6 +1,6 @@
 from .launcher import launch_training, launch_runner
 from .runners import BaseRunner
 
-__version__ = "0.3.5"
+__version__ = "0.3.6"
 
 __all__ = ["__version__", "launch_training", "launch_runner", "BaseRunner"]

basicts/data/__init__.py

@@ -4,8 +4,9 @@
 
 from .registry import SCALER_REGISTRY
 from .dataset_zoo.simple_tsf_dataset import TimeSeriesForecastingDataset
+from .dataset_zoo.m4_dataset import M4ForecastingDataset
 
-__all__ = ["SCALER_REGISTRY", "TimeSeriesForecastingDataset"]
+__all__ = ["SCALER_REGISTRY", "TimeSeriesForecastingDataset", "M4ForecastingDataset"]
 
 # fix bugs on Windows systems and on jupyter
 project_dir = os.path.dirname(os.path.dirname(os.path.dirname(os.path.abspath(__file__))))

basicts/data/dataset_zoo/m4_dataset.py

@@ -0,0 +1,84 @@
+import os
+import random
+
+import torch
+from torch.utils.data import Dataset
+
+from ...utils import load_pkl
+
+
+class M4ForecastingDataset(Dataset):
+    """
+    BasicTS tries its best to follow the commonly-used processing approaches of M4 dataset, while also providing more flexible interfaces.
+    M4 dataset differs from general MTS datasets in the following aspects:
+        - M4 dataset is a univariate time series dataset, which does not sample in a synchronized manner.
+            In the state-of-the-art M4 prediction solutions, NBeats [1], the authors first sample ids of the time series and then randomly sample the time series data for each time series.
+        - Padding and masking are used to make training more flexible and robust.
+        - There is no normalization in M4 dataset.
+        - There is no validation dataset in M4 dataset.
+        - The test data is the last sample of each time series.
+        - The future sequence length is fixed for different subsets.
+
+    Reference:
+        [1] N-BEATS: Neural basis expansion analysis for interpretable time series forecasting
+        [2] https://github.com/ServiceNow/N-BEATS/blob/master/common/sampler.py
+    """
+
+    def __init__(self, data_file_path: str, index_file_path: str, mask_file_path: str, mode: str) -> None:
+        super().__init__()
+        assert mode in ["train", "test"], "error mode"
+        self._check_if_file_exists(data_file_path, index_file_path, mask_file_path)
+        # read raw data (normalized)
+        self.data = load_pkl(data_file_path)[mode] # padded data: List[List]
+        self.mask = load_pkl(mask_file_path)[mode] # padded mask: List[List]
+        # read index
+        self.index = load_pkl(index_file_path)[mode] # train/test index of each time series: List[List]
+
+    def _check_if_file_exists(self, data_file_path: str, index_file_path: str, mask_file_path: str):
+        """Check if data file and index file exist.
+
+        Args:
+            data_file_path (str): data file path
+            index_file_path (str): index file path
+
+        Raises:
+            FileNotFoundError: no data file
+            FileNotFoundError: no index file
+        """
+
+        if not os.path.isfile(data_file_path):
+            raise FileNotFoundError("BasicTS can not find data file {0}".format(data_file_path))
+        if not os.path.isfile(index_file_path):
+            raise FileNotFoundError("BasicTS can not find index file {0}".format(index_file_path))
+        if not os.path.isfile(mask_file_path):
+            raise FileNotFoundError("BasicTS can not find mask file {0}".format(mask_file_path))
+
+    def __getitem__(self, ts_id: int) -> tuple:
+        """Get a sample.
+
+        Args:
+            ts_id (int): the iteration index, i.e., the time series id (not the self.index).
+
+        Returns:
+            tuple: future_data, history_data, future_mask, history_mask, where the shape of data is L x C and mask is L.
+        """
+
+        ts_idxs = list(self.index[ts_id])
+        # random select a time series sample
+        idx = ts_idxs[random.randint(0, len(ts_idxs)-1)]
+
+        history_data = torch.Tensor(self.data[ts_id][idx[0]:idx[1]]).unsqueeze(1).float()
+        future_data = torch.Tensor(self.data[ts_id][idx[1]:idx[2]]).unsqueeze(1).float()
+        history_mask = torch.Tensor(self.mask[ts_id][idx[0]:idx[1]]).unsqueeze(1).float()
+        future_mask = torch.Tensor(self.mask[ts_id][idx[1]:idx[2]]).unsqueeze(1).float()
+
+        return future_data, history_data, future_mask, history_mask
+
+    def __len__(self):
+        """Dataset length (=number of time series)
+
+        Returns:
+            int: dataset length
+        """
+
+        return len(self.data)

basicts/runners/__init__.py

@@ -2,6 +2,8 @@
 from .base_tsf_runner import BaseTimeSeriesForecastingRunner
 from .runner_zoo.simple_tsf_runner import SimpleTimeSeriesForecastingRunner
 from .runner_zoo.no_bp_runner import NoBPRunner
+from .runner_zoo.m4_tsf_runner import M4Runner
 
 __all__ = ["BaseRunner", "BaseTimeSeriesForecastingRunner",
-           "SimpleTimeSeriesForecastingRunner", "NoBPRunner"]
+           "SimpleTimeSeriesForecastingRunner", "NoBPRunner",
+           "M4Runner"]