No public description

PiperOrigin-RevId: 631148246 No public description PiperOrigin-RevId: 629465861 No public description PiperOrigin-RevId: 629538727 No public description PiperOrigin-RevId: 629803430 No public description PiperOrigin-RevId: 629838772 No public description PiperOrigin-RevId: 629875841 No public description PiperOrigin-RevId: 630194303 No public description PiperOrigin-RevId: 630212078 No public description PiperOrigin-RevId: 630215811 No public description PiperOrigin-RevId: 630402195 No public description PiperOrigin-RevId: 630452190 No public description PiperOrigin-RevId: 630511573 No public description PiperOrigin-RevId: 630706657 No public description PiperOrigin-RevId: 630712978 No public description PiperOrigin-RevId: 630735187
2024-04-30 10:58:13 -07:00
commit 81dc60c086
19 changed files with 2776 additions and 0 deletions
@@ -0,0 +1,26 @@
+# Extended Benchmarks
+
+The benchmark setting has been borrowed from Nixtla's original [benchmarking](https://github.com/AzulGarza/nixtla/tree/main/experiments/amazon-chronos) of time-series foundation models against a strong statistical ensemble. Later more datasets were added by the Chronos team in this [pull request](https://github.com/shchur/nixtla/tree/chronos-full-eval/experiments/amazon-chronos). We compare on all the datasets in this extended benchmarks.
+
+All experiments were performed on a [g2-standard-32](https://cloud.google.com/compute/docs/gpus). 
+
+## Running TimesFM on the benchmark
+
+Install the environment and the package as detailed in the main README and then follow the steps from the base directory.
+
+```
+conda activate tfm_env
+TF_CPP_MIN_LOG_LEVEL=2 XLA_PYTHON_CLIENT_PREALLOCATE=false python3 -m experiments.extended_benchmarks.run_timesfm --model_path=<model_path> --backend="gpu"
+```
+
+In the above, `<model_path>` should point to the checkpoint directory that can be downloaded from HuggingFace. 
+
+Note: In the current version of TimesFM we focus on point forecasts and therefore the mase, smape have been calculated using the quantile head corresponding to the median i.e 0.5 quantile. We do offer 10 quantile heads but they have not been calibrated after pretraining. We recommend using them with caution or calibrate/conformalize them on a hold out for your applications. More to follow on later versions.
+
+## Benchmark Results
+
+![Benchmark Results Table](./tfm_results.png)
+
+We can see that TimesFM performs the best in terms of both mase and smape. More importantly it is much faster than the other methods, in particular it is more than 600x faster than StatisticalEnsemble and 80x faster than Chronos (Large).
+
+Note: This benchmark only compares on `one` small horizon window for long horizon datasets like ETT hourly and 15 minutes. More in depth comparison on longer horizon rolling validation tasks are presented in our long horizon benchmarks.
@@ -0,0 +1,146 @@
+# Copyright 2024 Google LLC
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""Evaluation script for timesfm."""
+
+import os
+import sys
+import time
+
+from absl import flags
+import numpy as np
+import pandas as pd
+from paxml import checkpoints
+import timesfm
+
+from .utils import ExperimentHandler
+
+
+dataset_names = [
+    "m1_monthly",
+    "m1_quarterly",
+    "m1_yearly",
+    "m3_monthly",
+    "m3_other",
+    "m3_quarterly",
+    "m3_yearly",
+    "m4_quarterly",
+    "m4_yearly",
+    "tourism_monthly",
+    "tourism_quarterly",
+    "tourism_yearly",
+    "nn5_daily_without_missing",
+    "m5",
+    "nn5_weekly",
+    "traffic",
+    "weather",
+    "dominick",
+    "australian_electricity_demand",
+    "car_parts_without_missing",
+    "cif_2016",
+    "covid_deaths",
+    "ercot",
+    "ett_small_15min",
+    "ett_small_1h",
+    "exchange_rate",
+    "fred_md",
+    "hospital",
+]
+
+context_dict = {
+    "cif_2016": 32,
+    "tourism_yearly": 64,
+    "covid_deaths": 64,
+    "tourism_quarterly": 64,
+    "tourism_monthly": 64,
+    "m1_monthly": 64,
+    "m1_quarterly": 64,
+    "m1_yearly": 64,
+    "m3_monthly": 64,
+    "m3_other": 64,
+    "m3_quarterly": 64,
+    "m3_yearly": 64,
+    "m4_quarterly": 64,
+    "m4_yearly": 64,
+}
+
+_MODEL_PATH = flags.DEFINE_string(
+    "model_path", "/home/timesfm_q10_20240501", "Path to model"
+)
+_BATCH_SIZE = flags.DEFINE_integer("batch_size", 64, "Batch size")
+_HORIZON = flags.DEFINE_integer("horizon", 128, "Horizon")
+_BACKEND = flags.DEFINE_string("backend", "gpu", "Backend")
+_NUM_JOBS = flags.DEFINE_integer("num_jobs", 1, "Number of jobs")
+_SAVE_DIR = flags.DEFINE_string("save_dir", "./results", "Save directory")
+
+
+QUANTILES = list(np.arange(1, 10) / 10.0)
+
+
+def main():
+  results_list = []
+  tfm = timesfm.TimesFm(
+      context_len=512,
+      horizon_len=_HORIZON.value,
+      input_patch_len=32,
+      output_patch_len=128,
+      num_layers=20,
+      model_dims=1280,
+      backend=_BACKEND.value,
+      per_core_batch_size=_BATCH_SIZE.value,
+      quantiles=QUANTILES,
+  )
+  tfm.load_from_checkpoint(
+      _MODEL_PATH.value,
+      checkpoint_type=checkpoints.CheckpointType.FLAX,
+  )
+  run_id = np.random.randint(100000)
+  model_name = "timesfm"
+  for dataset in dataset_names:
+    print(f"Evaluating model {model_name} on dataset {dataset}", flush=True)
+    exp = ExperimentHandler(dataset, quantiles=QUANTILES)
+
+    if dataset in context_dict:
+      context_len = context_dict[dataset]
+    else:
+      context_len = 512
+    train_df = exp.train_df
+    freq = exp.freq
+    init_time = time.time()
+    fcsts_df = tfm.forecast_on_df(
+        inputs=train_df,
+        freq=freq,
+        value_name="y",
+        model_name=model_name,
+        forecast_context_len=context_len,
+        num_jobs=_NUM_JOBS.value,
+    )
+    total_time = time.time() - init_time
+    time_df = pd.DataFrame({"time": [total_time], "model": model_name})
+    results = exp.evaluate_from_predictions(
+        models=[model_name], fcsts_df=fcsts_df, times_df=time_df
+    )
+    print(results, flush=True)
+    results_list.append(results)
+    results_full = pd.concat(results_list)
+    save_path = os.path.join(_SAVE_DIR.value, str(run_id))
+    print(f"Saving results to {save_path}", flush=True)
+    os.makedirs(save_path, exist_ok=True)
+    results_full.to_csv(f"{save_path}/results.csv")
+
+
+if __name__ == "__main__":
+  FLAGS = flags.FLAGS
+  FLAGS(sys.argv)
+  main()
@@ -0,0 +1,278 @@
+# Copyright 2024 Google LLC
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""Forked from https://github.com/Nixtla/nixtla/blob/main/experiments/amazon-chronos/src/utils.py."""
+
+from functools import partial
+from itertools import repeat
+import multiprocessing
+import os
+from pathlib import Path
+from typing import List
+
+from gluonts.dataset import Dataset
+from gluonts.dataset.repository.datasets import (
+    dataset_names as gluonts_datasets,
+    get_dataset,
+)
+from gluonts.time_feature.seasonality import get_seasonality
+import numpy as np
+import pandas as pd
+from utilsforecast.evaluation import evaluate
+from utilsforecast.losses import mae, mase, smape
+
+
+def parallel_transform(inp):
+  ts, last_n = inp[0], inp[1]
+  return ExperimentHandler._transform_gluonts_instance_to_df(ts, last_n=last_n)
+
+
+def quantile_loss(
+    df: pd.DataFrame,
+    models: list,
+    q: float = 0.5,
+    id_col: str = "unique_id",
+    target_col: str = "y",
+) -> pd.DataFrame:
+  delta_y = df[models].sub(df[target_col], axis=0)
+  res = (
+      np.maximum(q * delta_y, (q - 1) * delta_y)
+      .groupby(df[id_col], observed=True)
+      .mean()
+  )
+  res.index.name = id_col
+  res = res.reset_index()
+  return res
+
+
+class ExperimentHandler:
+
+  def __init__(
+      self,
+      dataset: str,
+      quantiles: List[float] = list(np.arange(1, 10) / 10.0),
+      results_dir: str = "./results",
+      models_dir: str = "./models",
+  ):
+    if dataset not in gluonts_datasets:
+      raise Exception(
+          f"dataset {dataset} not found in gluonts "
+          f"available datasets: {', '.join(gluonts_datasets)}"
+      )
+    self.dataset = dataset
+    self.quantiles = quantiles
+    self.level = self._transform_quantiles_to_levels(quantiles)
+    self.results_dir = results_dir
+    self.models_dir = models_dir
+    # defining datasets
+    self._maybe_download_m3_or_m5_file(self.dataset)
+    gluonts_dataset = get_dataset(self.dataset)
+    self.horizon = gluonts_dataset.metadata.prediction_length
+    if self.horizon is None:
+      raise Exception(
+          f"horizon not found for dataset {self.dataset} "
+          "experiment cannot be run"
+      )
+    self.freq = gluonts_dataset.metadata.freq
+    # get_seasonality() returns 1 for freq='D', override this to 7. This significantly improves the accuracy of
+    # statistical models on datasets like m5/nn5_daily. The models like AutoARIMA/AutoETS can still set
+    # seasonality=1 internally on datasets like weather by choosing non-seasonal models during model selection.
+    if self.freq == "D":
+      self.seasonality = 7
+    else:
+      self.seasonality = get_seasonality(self.freq)
+    self.gluonts_train_dataset = gluonts_dataset.train
+    self.gluonts_test_dataset = gluonts_dataset.test
+    self._create_dir_if_not_exists(self.results_dir)
+    try:
+      multiprocessing.set_start_method("spawn")
+    except RuntimeError:
+      print("Multiprocessing context has already been set.")
+
+  @staticmethod
+  def _maybe_download_m3_or_m5_file(dataset: str):
+    if dataset[:2] == "m3":
+      m3_file = Path.home() / ".gluonts" / "datasets" / "M3C.xls"
+      if not m3_file.exists():
+        from datasetsforecast.m3 import M3
+        from datasetsforecast.utils import download_file
+
+        download_file(m3_file.parent, M3.source_url)
+    elif dataset == "m5":
+      m5_raw_dir = Path.home() / ".gluonts" / "m5"
+      if not m5_raw_dir.exists():
+        import zipfile
+        from datasetsforecast.m5 import M5
+        from datasetsforecast.utils import download_file
+
+        download_file(m5_raw_dir, M5.source_url)
+        with zipfile.ZipFile(m5_raw_dir / "m5.zip", "r") as zip_ref:
+          zip_ref.extractall(m5_raw_dir)
+
+  @staticmethod
+  def _transform_quantiles_to_levels(quantiles: List[float]) -> List[int]:
+    level = [
+        int(100 - 200 * q) for q in quantiles if q < 0.5
+    ]  # in this case mean=mediain
+    level = sorted(list(set(level)))
+    return level
+
+  @staticmethod
+  def _create_dir_if_not_exists(directory: str):
+    Path(directory).mkdir(parents=True, exist_ok=True)
+
+  @staticmethod
+  def _transform_gluonts_instance_to_df(
+      ts: dict,
+      last_n: int | None = None,
+  ) -> pd.DataFrame:
+    start_period = ts["start"]
+    start_ds, freq = start_period.to_timestamp(), start_period.freq
+    target = ts["target"]
+    ds = pd.date_range(start=start_ds, freq=freq, periods=len(target))
+    if last_n is not None:
+      target = target[-last_n:]
+      ds = ds[-last_n:]
+    ts_df = pd.DataFrame({"unique_id": ts["item_id"], "ds": ds, "y": target})
+    return ts_df
+
+  @staticmethod
+  def _transform_gluonts_dataset_to_df(
+      gluonts_dataset: Dataset,
+      last_n: int | None = None,
+  ) -> pd.DataFrame:
+    with multiprocessing.Pool(os.cpu_count()) as pool:  # Create a process pool
+      results = pool.map(
+          parallel_transform, zip(gluonts_dataset, repeat(last_n))
+      )
+    df = pd.concat(results)
+    df = df.reset_index(drop=True)
+    return df
+
+  @property
+  def train_df(self) -> pd.DataFrame:
+    train_df = self._transform_gluonts_dataset_to_df(self.gluonts_train_dataset)
+    return train_df
+
+  @property
+  def test_df(self) -> pd.DataFrame:
+    test_df = self._transform_gluonts_dataset_to_df(
+        self.gluonts_test_dataset,
+        last_n=self.horizon,
+    )
+    # Make sure that only the first backtest window is used for evaluation on `traffic` / `exchange_rate` datasets
+    return test_df.groupby("unique_id", sort=False).head(self.horizon)
+
+  def save_dataframe(self, df: pd.DataFrame, file_name: str):
+    df.to_csv(f"{self.results_dir}/{file_name}", index=False)
+
+  def save_results(
+      self, fcst_df: pd.DataFrame, total_time: float, model_name: str
+  ):
+    self.save_dataframe(
+        fcst_df,
+        f"{model_name}-{self.dataset}-fcst.csv",
+    )
+    time_df = pd.DataFrame({"time": [total_time], "model": model_name})
+    self.save_dataframe(
+        time_df,
+        f"{model_name}-{self.dataset}-time.csv",
+    )
+
+  def fcst_from_level_to_quantiles(
+      self,
+      fcst_df: pd.DataFrame,
+      model_name: str,
+  ) -> pd.DataFrame:
+    fcst_df = fcst_df.copy()
+    cols = ["unique_id", "ds", model_name]
+    for q in self.quantiles:
+      if q == 0.5:
+        col = f"{model_name}"
+      else:
+        lv = int(100 - 200 * q)
+        hi_or_lo = "lo" if lv > 0 else "hi"
+        lv = abs(lv)
+        col = f"{model_name}-{hi_or_lo}-{lv}"
+      q_col = f"{model_name}-q-{q}"
+      fcst_df[q_col] = fcst_df[col].values
+      cols.append(q_col)
+    return fcst_df[cols]
+
+  def evaluate_models(self, models: List[str]) -> pd.DataFrame:
+    fcsts_df = []
+    times_df = []
+    for model in models:
+      fcst_method_df = pd.read_csv(
+          f"{self.results_dir}/{model}-{self.dataset}-fcst.csv"
+      ).set_index(["unique_id", "ds"])
+      fcsts_df.append(fcst_method_df)
+      time_method_df = pd.read_csv(
+          f"{self.results_dir}/{model}-{self.dataset}-time.csv"
+      )
+      times_df.append(time_method_df)
+    fcsts_df = pd.concat(fcsts_df, axis=1).reset_index()
+    fcsts_df["ds"] = pd.to_datetime(fcsts_df["ds"])
+    times_df = pd.concat(times_df)
+    return self.evaluate_from_predictions(
+        models=models, fcsts_df=fcsts_df, times_df=times_df
+    )
+
+  def evaluate_from_predictions(
+      self, models: List[str], fcsts_df: pd.DataFrame, times_df: pd.DataFrame
+  ) -> pd.DataFrame:
+    test_df = self.test_df
+    train_df = self.train_df
+    test_df = test_df.merge(fcsts_df, how="left")
+    assert test_df.isna().sum().sum() == 0, "merge contains nas"
+    # point evaluation
+    point_fcsts_cols = ["unique_id", "ds", "y"] + models
+    test_df["unique_id"] = test_df["unique_id"].astype(str)
+    train_df["unique_id"] = train_df["unique_id"].astype(str)
+    mase_seas = partial(mase, seasonality=self.seasonality)
+    eval_df = evaluate(
+        test_df[point_fcsts_cols],
+        train_df=train_df,
+        metrics=[smape, mase_seas, mae],
+    )
+    # probabilistic evaluation
+    eval_prob_df = []
+    for q in self.quantiles:
+      prob_cols = [f"{model}-q-{q}" for model in models]
+      eval_q_df = quantile_loss(test_df, models=prob_cols, q=q)
+      eval_q_df[prob_cols] = eval_q_df[prob_cols] * self.horizon
+      eval_q_df = eval_q_df.rename(columns=dict(zip(prob_cols, models)))
+      eval_q_df["metric"] = f"quantile-loss-{q}"
+      eval_prob_df.append(eval_q_df)
+    eval_prob_df = pd.concat(eval_prob_df)
+    eval_prob_df = eval_prob_df.groupby("metric").sum().reset_index()
+    total_y = test_df["y"].sum()
+    eval_prob_df[models] = eval_prob_df[models] / total_y
+    eval_prob_df["metric"] = "scaled_crps"
+    eval_df = pd.concat([eval_df, eval_prob_df]).reset_index(drop=True)
+    eval_df = eval_df.groupby("metric").mean(numeric_only=True).reset_index()
+    eval_df = eval_df.melt(
+        id_vars="metric", value_name="value", var_name="model"
+    )
+    times_df.insert(0, "metric", "time")
+    times_df = times_df.rename(columns={"time": "value"})
+    eval_df = pd.concat([eval_df, times_df])
+    eval_df.insert(0, "dataset", self.dataset)
+    eval_df = eval_df.sort_values(["dataset", "metric", "model"])
+    eval_df = eval_df.reset_index(drop=True)
+    return eval_df
+
+
+if __name__ == "__main__":
+  multiprocessing.set_start_method("spawn")
@@ -0,0 +1,36 @@
+# Extended Benchmarks
+
+We benchmark on the original test set for ETT datasets as per long horizon benchmark papers (see [here](https://openreview.net/forum?id=pCbC3aQB5W) for example.) In the original benchmark, rolling validation task on all test windows (with a stride of 1) is considered. While we can easily run our method on this task, the baselines can take a very long time to run. Therefore we present results on a modified task with stride between windows set to Horizon length i.e all disjoint horizons in the test period is considered.
+
+All experiments were performed on a [g2-standard-32](https://cloud.google.com/compute/docs/gpus). We compare TimesFM with [Amazon-Chronos](https://github.com/amazon-science/chronos-forecasting).
+
+## Running TimesFM on the benchmark
+
+Install the environment and the package as detailed in the main README and then follow the steps from the base directory.
+
+```
+conda activate tfm_env
+TF_CPP_MIN_LOG_LEVEL=2 XLA_PYTHON_CLIENT_PREALLOCATE=false python3 -m experiments.long_horizon_benchmarks.run_eval \
+--model_path=<model_path> --backend="gpu" \
+--pred_len=96 --context_len=512 --dataset=etth1
+```
+
+In the above, `<model_path>` should point to the checkpoint directory that can be downloaded from HuggingFace. 
+
+For running chronos on the same benchmark you can run the command,
+
+```
+TF_CPP_MIN_LOG_LEVEL=2 XLA_PYTHON_CLIENT_PREALLOCATE=false python3 -m experiments.long_horizon_benchmarks.run_eval \
+--model_path=amazon/chronos-t5-mini --backend="gpu" \
+--pred_len=96 --context_len=512 --dataset=etth1
+```
+
+You can change the model size from "mini" to "large" as required. The datasets we benchmark on are etth1, etth2, ettm1 and ettm2.
+
+## Benchmark Results
+
+![Benchmark Results Table](./tfm_long_horizon.png)
+
+We compare the performance on horizon lengths of 96, 192 and 336, while context length is held fixed at 512.
+
+We can see that TimesFM performs the best in terms of both wape and smape. More importantly it is much faster than the other methods, in particular it is more than 1000x faster than Chronos (Large).
@@ -0,0 +1,261 @@
+# Copyright 2024 The Google Research Authors.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""TF dataloaders for general timeseries datasets.
+
+The expected input format is csv file with a datetime index.
+"""
+
+
+from absl import logging
+import numpy as np
+import pandas as pd
+from sklearn.preprocessing import StandardScaler
+import tensorflow as tf
+from . import time_features
+
+
+class TimeSeriesdata(object):
+  """Data loader class."""
+
+  def __init__(
+      self,
+      data_path,
+      datetime_col,
+      num_cov_cols,
+      cat_cov_cols,
+      ts_cols,
+      train_range,
+      val_range,
+      test_range,
+      hist_len,
+      pred_len,
+      batch_size,
+      freq='H',
+      normalize=True,
+      epoch_len=None,
+      holiday=False,
+      permute=True,
+  ):
+    """Initialize objects.
+
+    Args:
+      data_path: path to csv file
+      datetime_col: column name for datetime col
+      num_cov_cols: list of numerical global covariates
+      cat_cov_cols: list of categorical global covariates
+      ts_cols: columns corresponding to ts
+      train_range: tuple of train ranges
+      val_range: tuple of validation ranges
+      test_range: tuple of test ranges
+      hist_len: historical context
+      pred_len: prediction length
+      batch_size: batch size (number of ts in a batch)
+      freq: freq of original data
+      normalize: std. normalize data or not
+      epoch_len: num iters in an epoch
+      holiday: use holiday features or not
+      permute: permute ts in train batches or not
+
+    Returns:
+      None
+    """
+    self.data_df = pd.read_csv(open(data_path, 'r'))
+    if not num_cov_cols:
+      self.data_df['ncol'] = np.zeros(self.data_df.shape[0])
+      num_cov_cols = ['ncol']
+    if not cat_cov_cols:
+      self.data_df['ccol'] = np.zeros(self.data_df.shape[0])
+      cat_cov_cols = ['ccol']
+    self.data_df.fillna(0, inplace=True)
+    self.data_df.set_index(
+        pd.DatetimeIndex(self.data_df[datetime_col]), inplace=True
+    )
+    self.num_cov_cols = num_cov_cols
+    self.cat_cov_cols = cat_cov_cols
+    self.ts_cols = ts_cols
+    self.train_range = train_range
+    self.val_range = val_range
+    self.test_range = test_range
+    data_df_idx = self.data_df.index
+    date_index = data_df_idx.union(
+        pd.date_range(
+            data_df_idx[-1] + pd.Timedelta(1, freq=freq),
+            periods=pred_len + 1,
+            freq=freq,
+        )
+    )
+    self.time_df = time_features.TimeCovariates(
+        date_index, holiday=holiday
+    ).get_covariates()
+    self.hist_len = hist_len
+    self.pred_len = pred_len
+    self.batch_size = batch_size
+    self.freq = freq
+    self.normalize = normalize
+    self.data_mat = self.data_df[self.ts_cols].to_numpy().transpose()
+    self.data_mat = self.data_mat[:, 0 : self.test_range[1]]
+    self.time_mat = self.time_df.to_numpy().transpose()
+    self.num_feat_mat = self.data_df[num_cov_cols].to_numpy().transpose()
+    self.cat_feat_mat, self.cat_sizes = self._get_cat_cols(cat_cov_cols)
+    self.normalize = normalize
+    if normalize:
+      self._normalize_data()
+    logging.info(
+        'Data Shapes: %s, %s, %s, %s',
+        self.data_mat.shape,
+        self.time_mat.shape,
+        self.num_feat_mat.shape,
+        self.cat_feat_mat.shape,
+    )
+    self.epoch_len = epoch_len
+    self.permute = permute
+
+  def _get_cat_cols(self, cat_cov_cols):
+    """Get categorical columns."""
+    cat_vars = []
+    cat_sizes = []
+    for col in cat_cov_cols:
+      dct = {x: i for i, x in enumerate(self.data_df[col].unique())}
+      cat_sizes.append(len(dct))
+      mapped = self.data_df[col].map(lambda x: dct[x]).to_numpy().transpose()  # pylint: disable=cell-var-from-loop
+      cat_vars.append(mapped)
+    return np.vstack(cat_vars), cat_sizes
+
+  def _normalize_data(self):
+    self.scaler = StandardScaler()
+    train_mat = self.data_mat[:, self.train_range[0] : self.train_range[1]]
+    self.scaler = self.scaler.fit(train_mat.transpose())
+    self.data_mat = self.scaler.transform(self.data_mat.transpose()).transpose()
+
+  def train_gen(self):
+    """Generator for training data."""
+    num_ts = len(self.ts_cols)
+    perm = np.arange(
+        self.train_range[0] + self.hist_len,
+        self.train_range[1] - self.pred_len,
+    )
+    perm = np.random.permutation(perm)
+    hist_len = self.hist_len
+    logging.info('Hist len: %s', hist_len)
+    if not self.epoch_len:
+      epoch_len = len(perm)
+    else:
+      epoch_len = self.epoch_len
+    for idx in perm[0:epoch_len]:
+      for _ in range(num_ts // self.batch_size + 1):
+        if self.permute:
+          tsidx = np.random.choice(num_ts, size=self.batch_size, replace=False)
+        else:
+          tsidx = np.arange(num_ts)
+        dtimes = np.arange(idx - hist_len, idx + self.pred_len)
+        (
+            bts_train,
+            bts_pred,
+            bfeats_train,
+            bfeats_pred,
+            bcf_train,
+            bcf_pred,
+        ) = self._get_features_and_ts(dtimes, tsidx, hist_len)
+
+        all_data = [
+            bts_train,
+            bfeats_train,
+            bcf_train,
+            bts_pred,
+            bfeats_pred,
+            bcf_pred,
+            tsidx,
+        ]
+        yield tuple(all_data)
+
+  def test_val_gen(self, mode='val', shift=1):
+    """Generator for validation/test data."""
+    if mode == 'val':
+      start = self.val_range[0]
+      end = self.val_range[1] - self.pred_len + 1
+    elif mode == 'test':
+      start = self.test_range[0]
+      end = self.test_range[1] - self.pred_len + 1
+    else:
+      raise NotImplementedError('Eval mode not implemented')
+    num_ts = len(self.ts_cols)
+    hist_len = self.hist_len
+    logging.info('Hist len: %s', hist_len)
+    perm = np.arange(start, end)
+    if self.epoch_len:
+      epoch_len = self.epoch_len
+    else:
+      epoch_len = len(perm)
+    for i in range(0, epoch_len, shift):
+      idx = perm[i]
+      for batch_idx in range(0, num_ts, self.batch_size):
+        tsidx = np.arange(batch_idx, min(batch_idx + self.batch_size, num_ts))
+        dtimes = np.arange(idx - hist_len, idx + self.pred_len)
+        (
+            bts_train,
+            bts_pred,
+            bfeats_train,
+            bfeats_pred,
+            bcf_train,
+            bcf_pred,
+        ) = self._get_features_and_ts(dtimes, tsidx, hist_len)
+        all_data = [
+            bts_train,
+            bfeats_train,
+            bcf_train,
+            bts_pred,
+            bfeats_pred,
+            bcf_pred,
+            tsidx,
+        ]
+        yield tuple(all_data)
+
+  def _get_features_and_ts(self, dtimes, tsidx, hist_len=None):
+    """Get features and ts in specified windows."""
+    if hist_len is None:
+      hist_len = self.hist_len
+    data_times = dtimes[dtimes < self.data_mat.shape[1]]
+    bdata = self.data_mat[:, data_times]
+    bts = bdata[tsidx, :]
+    bnf = self.num_feat_mat[:, data_times]
+    bcf = self.cat_feat_mat[:, data_times]
+    btf = self.time_mat[:, dtimes]
+    if bnf.shape[1] < btf.shape[1]:
+      rem_len = btf.shape[1] - bnf.shape[1]
+      rem_rep = np.repeat(bnf[:, [-1]], repeats=rem_len)
+      rem_rep_cat = np.repeat(bcf[:, [-1]], repeats=rem_len)
+      bnf = np.hstack([bnf, rem_rep.reshape(bnf.shape[0], -1)])
+      bcf = np.hstack([bcf, rem_rep_cat.reshape(bcf.shape[0], -1)])
+    bfeats = np.vstack([btf, bnf])
+    bts_train = bts[:, 0:hist_len]
+    bts_pred = bts[:, hist_len:]
+    bfeats_train = bfeats[:, 0:hist_len]
+    bfeats_pred = bfeats[:, hist_len:]
+    bcf_train = bcf[:, 0:hist_len]
+    bcf_pred = bcf[:, hist_len:]
+    return bts_train, bts_pred, bfeats_train, bfeats_pred, bcf_train, bcf_pred
+
+  def tf_dataset(self, mode='train', shift=1):
+    """Tensorflow Dataset."""
+    if mode == 'train':
+      gen_fn = self.train_gen
+    else:
+      gen_fn = lambda: self.test_val_gen(mode, shift)
+    output_types = tuple(
+        [tf.float32] * 2 + [tf.int32] + [tf.float32] * 2 + [tf.int32] * 2
+    )
+    dataset = tf.data.Dataset.from_generator(gen_fn, output_types)
+    dataset = dataset.prefetch(tf.data.experimental.AUTOTUNE)
+    return dataset
@@ -0,0 +1,255 @@
+# Copyright 2024 The Google Research Authors.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""Eval pipeline."""
+
+import json
+import os
+import sys
+import time
+from absl import flags
+import chronos
+import numpy as np
+import pandas as pd
+from paxml import checkpoints
+import timesfm
+import torch
+import tqdm
+from . import data_loader
+
+
+FLAGS = flags.FLAGS
+
+_BATCH_SIZE = flags.DEFINE_integer(
+    "batch_size", 64, "Batch size for the randomly sampled batch"
+)
+_DATASET = flags.DEFINE_string("dataset", "etth1", "The name of the dataset.")
+_MODEL_PATH = flags.DEFINE_string(
+    "model_path", "./timesfm_q10_20240501", "The name of the dataset."
+)
+_DATETIME_COL = flags.DEFINE_string(
+    "datetime_col", "date", "Column having datetime."
+)
+_NUM_COV_COLS = flags.DEFINE_list(
+    "num_cov_cols", None, "Column having numerical features."
+)
+_CAT_COV_COLS = flags.DEFINE_list(
+    "cat_cov_cols", None, "Column having categorical features."
+)
+_TS_COLS = flags.DEFINE_list("ts_cols", None, "Columns of time-series features")
+_NORMALIZE = flags.DEFINE_bool(
+    "normalize", True, "normalize data for eval or not"
+)
+_CONTEXT_LEN = flags.DEFINE_integer(
+    "context_len", 512, "Length of the context window"
+)
+_PRED_LEN = flags.DEFINE_integer("pred_len", 96, "prediction length.")
+_BACKEND = flags.DEFINE_string("backend", "gpu", "backend to use")
+_RESULTS_DIR = flags.DEFINE_string(
+    "results_dir", "./results/long_horizon", "results directory"
+)
+
+
+DATA_DICT = {
+    "ettm2": {
+        "boundaries": [34560, 46080, 57600],
+        "data_path": "./datasets/ETT-small/ETTm2.csv",
+        "freq": "15min",
+    },
+    "ettm1": {
+        "boundaries": [34560, 46080, 57600],
+        "data_path": "./datasets/ETT-small/ETTm1.csv",
+        "freq": "15min",
+    },
+    "etth2": {
+        "boundaries": [8640, 11520, 14400],
+        "data_path": "./datasets/ETT-small/ETTh2.csv",
+        "freq": "H",
+    },
+    "etth1": {
+        "boundaries": [8640, 11520, 14400],
+        "data_path": "./datasets/ETT-small/ETTh1.csv",
+        "freq": "H",
+    },
+    "elec": {
+        "boundaries": [18413, 21044, 26304],
+        "data_path": "./datasets/electricity/electricity.csv",
+        "freq": "H",
+    },
+    "traffic": {
+        "boundaries": [12280, 14036, 17544],
+        "data_path": "./datasets/traffic/traffic.csv",
+        "freq": "H",
+    },
+    "weather": {
+        "boundaries": [36887, 42157, 52696],
+        "data_path": "./datasets/weather/weather.csv",
+        "freq": "10min",
+    },
+}
+
+QUANTILES = list(np.arange(1, 10) / 10.0)
+EPS = 1e-7
+
+
+def get_forecasts(model_path, model, past, freq, pred_len):
+  """Get forecasts."""
+  if model_path.startswith("amazon"):
+    out = model.predict(
+        torch.tensor(past),
+        prediction_length=pred_len,
+        limit_prediction_length=False,
+    )
+    out = out.numpy()
+    out = np.median(out, axis=1)
+  else:
+    lfreq = [freq] * past.shape[0]
+    _, out = model.forecast(list(past), lfreq)
+    out = out[:, :, 5]
+  return out
+
+
+def _mse(y_pred, y_true):
+  """mse loss."""
+  return np.square(y_pred - y_true)
+
+
+def _mae(y_pred, y_true):
+  """mae loss."""
+  return np.abs(y_pred - y_true)
+
+
+def _smape(y_pred, y_true):
+  """_smape loss."""
+  abs_diff = np.abs(y_pred - y_true)
+  abs_val = (np.abs(y_true) + np.abs(y_pred)) / 2
+  abs_val = np.where(abs_val > EPS, abs_val, 1.0)
+  abs_diff = np.where(abs_val > EPS, abs_diff, 0.0)
+  return abs_diff / abs_val
+
+
+def eval():
+  """Eval pipeline."""
+  dataset = _DATASET.value
+  data_path = DATA_DICT[dataset]["data_path"]
+  freq = DATA_DICT[dataset]["freq"]
+  int_freq = timesfm.freq_map(freq)
+  boundaries = DATA_DICT[dataset]["boundaries"]
+
+  data_df = pd.read_csv(open(data_path, "r"))
+
+  if _TS_COLS.value is not None:
+    ts_cols = DATA_DICT[dataset]["ts_cols"]
+    num_cov_cols = DATA_DICT[dataset]["num_cov_cols"]
+    cat_cov_cols = DATA_DICT[dataset]["cat_cov_cols"]
+  else:
+    ts_cols = [col for col in data_df.columns if col != _DATETIME_COL.value]
+    num_cov_cols = None
+    cat_cov_cols = None
+  batch_size = min(_BATCH_SIZE.value, len(ts_cols))
+  dtl = data_loader.TimeSeriesdata(
+      data_path=data_path,
+      datetime_col=_DATETIME_COL.value,
+      num_cov_cols=num_cov_cols,
+      cat_cov_cols=cat_cov_cols,
+      ts_cols=np.array(ts_cols),
+      train_range=[0, boundaries[0]],
+      val_range=[boundaries[0], boundaries[1]],
+      test_range=[boundaries[1], boundaries[2]],
+      hist_len=_CONTEXT_LEN.value,
+      pred_len=_PRED_LEN.value,
+      batch_size=batch_size,
+      freq=freq,
+      normalize=_NORMALIZE.value,
+      epoch_len=None,
+      holiday=False,
+      permute=False,
+  )
+  eval_itr = dtl.tf_dataset(
+      mode="test", shift=_PRED_LEN.value
+  ).as_numpy_iterator()
+  model_path = _MODEL_PATH.value
+  if model_path.startswith("amazon"):
+    model = chronos.ChronosPipeline.from_pretrained(
+        model_path,
+        device_map="auto",
+        torch_dtype=torch.bfloat16,
+    )
+  else:
+    model = timesfm.TimesFm(
+        context_len=_CONTEXT_LEN.value,
+        horizon_len=_PRED_LEN.value,
+        input_patch_len=32,
+        output_patch_len=128,
+        num_layers=20,
+        model_dims=1280,
+        backend=_BACKEND.value,
+        per_core_batch_size=batch_size,
+        quantiles=QUANTILES,
+    )
+    model.load_from_checkpoint(
+        model_path,
+        checkpoint_type=checkpoints.CheckpointType.FLAX,
+    )
+  smape_run_losses = []
+  mse_run_losses = []
+  mae_run_losses = []
+
+  num_elements = 0
+  abs_sum = 0
+  start_time = time.time()
+
+  for batch in tqdm.tqdm(eval_itr):
+    past = batch[0]
+    actuals = batch[3]
+    forecasts = get_forecasts(
+        model_path, model, past, int_freq, _PRED_LEN.value
+    )
+    forecasts = forecasts[:, 0 : actuals.shape[1]]
+    mae_run_losses.append(_mae(forecasts, actuals).sum())
+    mse_run_losses.append(_mse(forecasts, actuals).sum())
+    smape_run_losses.append(_smape(forecasts, actuals).sum())
+    num_elements += actuals.shape[0] * actuals.shape[1]
+    abs_sum += np.abs(actuals).sum()
+
+  mse_val = np.sum(mse_run_losses) / num_elements
+
+  result_dict = {
+      "mse": mse_val,
+      "smape": np.sum(smape_run_losses) / num_elements,
+      "mae": np.sum(mae_run_losses) / num_elements,
+      "wape": np.sum(mae_run_losses) / abs_sum,
+      "nrmse": np.sqrt(mse_val) / (abs_sum / num_elements),
+      "num_elements": num_elements,
+      "abs_sum": abs_sum,
+      "total_time": time.time() - start_time,
+      "model_path": model_path,
+      "dataset": dataset,
+      "freq": freq,
+      "pred_len": _PRED_LEN.value,
+      "context_len": _CONTEXT_LEN.value,
+  }
+  run_id = np.random.randint(10000)
+  save_path = os.path.join(_RESULTS_DIR.value, str(run_id))
+  print(f"Saving results to {save_path}", flush=True)
+  os.makedirs(save_path, exist_ok=True)
+  with open(os.path.join(save_path, "results.json"), "w") as f:
+    json.dump(result_dict, f)
+  print(result_dict, flush=True)
+
+
+if __name__ == "__main__":
+  FLAGS = flags.FLAGS
+  FLAGS(sys.argv)
+  eval()
@@ -0,0 +1,215 @@
+# Copyright 2024 The Google Research Authors.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""Directory to extract time covariates.
+
+Extract time covariates from datetime.
+"""
+
+import numpy as np
+import pandas as pd
+from pandas.tseries.holiday import EasterMonday
+from pandas.tseries.holiday import GoodFriday
+from pandas.tseries.holiday import Holiday
+from pandas.tseries.holiday import SU
+from pandas.tseries.holiday import TH
+from pandas.tseries.holiday import USColumbusDay
+from pandas.tseries.holiday import USLaborDay
+from pandas.tseries.holiday import USMartinLutherKingJr
+from pandas.tseries.holiday import USMemorialDay
+from pandas.tseries.holiday import USPresidentsDay
+from pandas.tseries.holiday import USThanksgivingDay
+from pandas.tseries.offsets import DateOffset
+from pandas.tseries.offsets import Day
+from pandas.tseries.offsets import Easter
+from sklearn.preprocessing import StandardScaler
+from tqdm import tqdm
+
+
+# This is 183 to cover half a year (in both directions), also for leap years
+# + 17 as Eastern can be between March, 22 - April, 25
+MAX_WINDOW = 183 + 17
+
+
+def _distance_to_holiday(holiday):
+  """Return distance to given holiday."""
+
+  def _distance_to_day(index):
+    holiday_date = holiday.dates(
+        index - pd.Timedelta(days=MAX_WINDOW),
+        index + pd.Timedelta(days=MAX_WINDOW),
+    )
+    assert (
+        len(holiday_date) != 0  # pylint: disable=g-explicit-length-test
+    ), f"No closest holiday for the date index {index} found."
+    # It sometimes returns two dates if it is exactly half a year after the
+    # holiday. In this case, the smaller distance (182 days) is returned.
+    return (index - holiday_date[0]).days
+
+  return _distance_to_day
+
+
+EasterSunday = Holiday(
+    "Easter Sunday", month=1, day=1, offset=[Easter(), Day(0)]
+)
+NewYearsDay = Holiday("New Years Day", month=1, day=1)
+SuperBowl = Holiday(
+    "Superbowl", month=2, day=1, offset=DateOffset(weekday=SU(1))
+)
+MothersDay = Holiday(
+    "Mothers Day", month=5, day=1, offset=DateOffset(weekday=SU(2))
+)
+IndependenceDay = Holiday("Independence Day", month=7, day=4)
+ChristmasEve = Holiday("Christmas", month=12, day=24)
+ChristmasDay = Holiday("Christmas", month=12, day=25)
+NewYearsEve = Holiday("New Years Eve", month=12, day=31)
+BlackFriday = Holiday(
+    "Black Friday",
+    month=11,
+    day=1,
+    offset=[pd.DateOffset(weekday=TH(4)), Day(1)],
+)
+CyberMonday = Holiday(
+    "Cyber Monday",
+    month=11,
+    day=1,
+    offset=[pd.DateOffset(weekday=TH(4)), Day(4)],
+)
+
+HOLIDAYS = [
+    EasterMonday,
+    GoodFriday,
+    USColumbusDay,
+    USLaborDay,
+    USMartinLutherKingJr,
+    USMemorialDay,
+    USPresidentsDay,
+    USThanksgivingDay,
+    EasterSunday,
+    NewYearsDay,
+    SuperBowl,
+    MothersDay,
+    IndependenceDay,
+    ChristmasEve,
+    ChristmasDay,
+    NewYearsEve,
+    BlackFriday,
+    CyberMonday,
+]
+
+
+class TimeCovariates(object):
+  """Extract all time covariates except for holidays."""
+
+  def __init__(
+      self,
+      datetimes,
+      normalized=True,
+      holiday=False,
+  ):
+    """Init function.
+
+    Args:
+      datetimes: pandas DatetimeIndex (lowest granularity supported is min)
+      normalized: whether to normalize features or not
+      holiday: fetch holiday features or not
+
+    Returns:
+      None
+    """
+    self.normalized = normalized
+    self.dti = datetimes
+    self.holiday = holiday
+
+  def _minute_of_hour(self):
+    minutes = np.array(self.dti.minute, dtype=np.float32)
+    if self.normalized:
+      minutes = minutes / 59.0 - 0.5
+    return minutes
+
+  def _hour_of_day(self):
+    hours = np.array(self.dti.hour, dtype=np.float32)
+    if self.normalized:
+      hours = hours / 23.0 - 0.5
+    return hours
+
+  def _day_of_week(self):
+    day_week = np.array(self.dti.dayofweek, dtype=np.float32)
+    if self.normalized:
+      day_week = day_week / 6.0 - 0.5
+    return day_week
+
+  def _day_of_month(self):
+    day_month = np.array(self.dti.day, dtype=np.float32)
+    if self.normalized:
+      day_month = day_month / 30.0 - 0.5
+    return day_month
+
+  def _day_of_year(self):
+    day_year = np.array(self.dti.dayofyear, dtype=np.float32)
+    if self.normalized:
+      day_year = day_year / 364.0 - 0.5
+    return day_year
+
+  def _month_of_year(self):
+    month_year = np.array(self.dti.month, dtype=np.float32)
+    if self.normalized:
+      month_year = month_year / 11.0 - 0.5
+    return month_year
+
+  def _week_of_year(self):
+    week_year = np.array(self.dti.strftime("%U").astype(int), dtype=np.float32)
+    if self.normalized:
+      week_year = week_year / 51.0 - 0.5
+    return week_year
+
+  def _get_holidays(self):
+    dti_series = self.dti.to_series()
+    hol_variates = np.vstack([
+        dti_series.apply(_distance_to_holiday(h)).values for h in tqdm(HOLIDAYS)
+    ])
+    # hol_variates is (num_holiday, num_time_steps), the normalization should be
+    # performed in the num_time_steps dimension.
+    return StandardScaler().fit_transform(hol_variates.T).T
+
+  def get_covariates(self):
+    """Get all time covariates."""
+    moh = self._minute_of_hour().reshape(1, -1)
+    hod = self._hour_of_day().reshape(1, -1)
+    dom = self._day_of_month().reshape(1, -1)
+    dow = self._day_of_week().reshape(1, -1)
+    doy = self._day_of_year().reshape(1, -1)
+    moy = self._month_of_year().reshape(1, -1)
+    woy = self._week_of_year().reshape(1, -1)
+
+    all_covs = [
+        moh,
+        hod,
+        dom,
+        dow,
+        doy,
+        moy,
+        woy,
+    ]
+    columns = ["moh", "hod", "dom", "dow", "doy", "moy", "woy"]
+    if self.holiday:
+      hol_covs = self._get_holidays()
+      all_covs.append(hol_covs)
+      columns += [f"hol_{i}" for i in range(len(HOLIDAYS))]
+
+    return pd.DataFrame(
+        data=np.vstack(all_covs).transpose(),
+        columns=columns,
+        index=self.dti,
+    )