Feature/hfhub utils

CHANGELOG.md

@@ -6,6 +6,10 @@ but cannot always guarantee backwards compatibility. Changes that may **break co
 
 ## [Unreleased](https://github.com/unit8co/darts/tree/master)
 
+- Added utility functions for Huggingface Hub integration. Upload/download Darts TimeSeries and ForecastingModel instances. [#2201](https://github.com/unit8co/darts/pull/2201)
+  by [Ivelin Ivanov](https://github.com/ivelin).
+
+
 [Full Changelog](https://github.com/unit8co/darts/compare/0.27.2...master)
 
 ### For users of the library:

darts/utils/hfhub.py

@@ -0,0 +1,104 @@
+import pandas as pd
+from dotenv import load_dotenv
+import os
+import tempfile
+from typing import Optional
+from darts import TimeSeries
+from darts.models.forecasting.forecasting_model import ForecastingModel
+from huggingface_hub import snapshot_download, upload_folder, create_repo
+from darts.logging import get_logger
+
+logger = get_logger(__name__)
+
+class HFHub:
+    """
+    HuggingFace Hub integration using official HF API.
+    https://huggingface.co/docs/huggingface_hub/v0.20.3/en/guides/integrations
+    """
+
+    def __init__(self, api_key: Optional[str] = None):
+        if api_key is None:
+            # load from .env file or OS vars if available
+            load_dotenv(override=True)
+            api_key = os.getenv("HF_TOKEN")
+            assert (
+                api_key is not None
+            ), "Could not find HF_TOKEN in OS environment. Cannot interact with HF Hub."
+        self.HF_TOKEN = api_key
+
+    def upload_model(
+        self,
+        repo_id: str = None,
+        model: ForecastingModel = None,
+        private: Optional[bool] = True,
+    ):
+        # Create repo if not existing yet and get the associated repo_id
+        create_repo(repo_id=repo_id, repo_type="model", private=private, exist_ok=True,
+                    token=self.HF_TOKEN)
+
+        with tempfile.TemporaryDirectory() as tmpdirname:
+            # print("created temporary directory for model", tmpdirname)
+            model.save(path=f"{tmpdirname}/{model.model_name}")
+            upload_folder(repo_id=repo_id, folder_path=tmpdirname, token=self.HF_TOKEN)
+
+    def download_model(
+        self,
+        repo_id: str = None,
+        model_name: str = None,
+        model_class: object = None,
+    ) -> ForecastingModel:
+        with tempfile.TemporaryDirectory() as tmpdirname:
+            logger.info(
+                "HFHub model files downloaded to local temp dir: ",
+                 os.listdir(tmpdirname)
+            )            
+            snapshot_download(
+                repo_id=repo_id, local_dir=tmpdirname, token=self.HF_TOKEN
+            )
+            model = model_class.load(path=f"{tmpdirname}/{model_name}")
+            return model
+
+    def upload_timeseries(
+        self,
+        repo_id: str = None,
+        series: TimeSeries = None,
+        series_name: str = None,
+        private: Optional[bool] = True,
+    ):
+        # Create repo if not existing yet and get the associated repo_id
+        repo_info = create_repo(
+            repo_id=repo_id, repo_type="dataset", private=private, exist_ok=True,
+            token=self.HF_TOKEN
+        )
+        # print(f"repo_info: ", repo_info)
+        df = series.pd_dataframe()
+        with tempfile.TemporaryDirectory() as tmpdirname:
+            df.to_parquet(path=f"{tmpdirname}/{series_name}.parquet")
+            upload_folder(
+                repo_id=repo_id,
+                repo_type="dataset",
+                folder_path=tmpdirname,
+                token=self.HF_TOKEN,
+            )
+
+    def download_timeseries(
+        self,
+        repo_id: str = None,
+        series_name: str = None,
+    ) -> TimeSeries:
+        with tempfile.TemporaryDirectory() as tmpdirname:
+            snapshot_download(
+                repo_id=repo_id,
+                repo_type="dataset",
+                local_dir=tmpdirname,
+                token=self.HF_TOKEN,
+            )
+            logger.info(
+                "HFHub data files downloaded to local temp dir: ",
+                 os.listdir(tmpdirname)
+            )            
+            df = pd.read_parquet(
+                f"{tmpdirname}/{series_name}.parquet", engine="pyarrow"
+            )
+            ts = TimeSeries.from_dataframe(df)
+            return ts