Dependency cleanup: remove pandas, torchvision, and onnx as core dependencies

CHANGELOG.md

@@ -114,6 +114,9 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ### Dependencies
 
+- Drops `onnx`, `torchvision` and `pandas` from the required dependencies.
+  `pandas` is now optional via `datapipes-extras` or `model-extras`.
+
 ## [2.2.1] - 2026-XX-YY
 
 ### Fixed

physicsnemo/core/version_check.py

@@ -214,6 +214,11 @@ def _format_install_hint(
         "h5py",
         group="datapipes-extras",
     ),
+    "pandas": _format_install_hint(
+        "pandas",
+        group="datapipes-extras",
+        direct_hint="pip install pandas  (also included in [model-extras])",
+    ),
     "netCDF4": _format_install_hint(
         "netCDF4",
         group="model-extras",

physicsnemo/datapipes/gnn/drivaernet_dataset.py

@@ -18,7 +18,6 @@
 from pathlib import Path
 from typing import Iterable
 
-import pandas as pd
 import torch
 from torch import Tensor
 from torch.utils.data import Dataset
@@ -28,6 +27,8 @@
 from physicsnemo.datapipes.meta import DatapipeMetaData
 from physicsnemo.nn.module.gnn_layers.utils import PyGData
 
+pd = OptionalImport("pandas")
+
 # Lazy imports for optional dependencies
 pyg = OptionalImport("torch_geometric")
 pv = OptionalImport("pyvista")

physicsnemo/datapipes/healpix/coupledtimeseries_dataset.py

@@ -23,7 +23,6 @@
 from typing import Optional, Sequence, Union
 
 import numpy as np
-import pandas as pd
 import torch
 from omegaconf import DictConfig, OmegaConf
 
@@ -34,6 +33,8 @@
 from . import couplers
 from .timeseries_dataset import TimeSeriesDataset
 
+pd = OptionalImport("pandas")
+
 xr = OptionalImport("xarray")
 
 logger = logging.getLogger(__name__)

physicsnemo/datapipes/healpix/couplers.py

@@ -20,11 +20,12 @@
 from typing import Sequence
 
 import numpy as np
-import pandas as pd
 import torch as th
 
 from physicsnemo.core.version_check import OptionalImport
 
+pd = OptionalImport("pandas")
+
 xr = OptionalImport("xarray")
 
 logger = logging.getLogger(__name__)
@@ -46,7 +47,7 @@ def __init__(
         presteps: int = 0,
         input_time_dim: int = 2,
         output_time_dim: int = 2,
-        input_times: Sequence = [pd.Timedelta("24h"), pd.Timedelta("48h")],
+        input_times: Sequence = ["24h", "48h"],
         prepared_coupled_data=True,
     ):
         """
@@ -68,8 +69,8 @@ def __init__(
         output_time_dim: int, optional
             number of output times for each model step, default 2
         input_times: Sequence, optional
-            sequence of pandas Timedelta objects that indicate which times are to be coupled,
-            default [pd.Timedelta("24h"), pd.Timedelta("48h")]
+            sequence of pandas Timedelta objects (or strings accepted by ``pd.Timedelta``)
+            that indicate which times are to be coupled, default ["24h", "48h"]
         prepared_coupled_data: boolean, optional
             If True assumes data in dataset has been prepared approiately for training:
             averages have already been calculated so that each time step denotes
@@ -289,7 +290,7 @@ def __init__(
         input_time_dim: int = 2,
         output_time_dim: int = 2,
         averaging_window: str = "24h",
-        input_times: Sequence = [pd.Timedelta("24h"), pd.Timedelta("48h")],
+        input_times: Sequence = ["24h", "48h"],
         prepared_coupled_data=True,
     ):
         """
@@ -313,8 +314,8 @@ def __init__(
         averaging_window: str, optional
             period over which coupled data is averaged before sent back to model, default "24h"
         input_times: Sequence, optional
-            sequence of pandas Timedelta objects that indicate which times are to be coupled,
-            default [pd.Timedelta("24h"), pd.Timedelta("48h")]
+            sequence of pandas Timedelta objects (or strings accepted by ``pd.Timedelta``)
+            that indicate which times are to be coupled, default ["24h", "48h"]
         prepared_coupled_data: boolean, optional
             If True assumes data in dataset has been prepared approiately for training:
             averages have already been calculated so that each time step denotes

physicsnemo/datapipes/healpix/timeseries_dataset.py

@@ -23,7 +23,6 @@
 from typing import Optional, Sequence, Union
 
 import numpy as np
-import pandas as pd
 import torch
 from omegaconf import DictConfig, OmegaConf
 
@@ -32,6 +31,8 @@
 from physicsnemo.datapipes.meta import DatapipeMetaData
 from physicsnemo.utils.insolation import insolation
 
+pd = OptionalImport("pandas")
+
 xr = OptionalImport("xarray")
 
 logger = logging.getLogger(__name__)

physicsnemo/experimental/datapipes/healda/configs/sensors.py

@@ -24,7 +24,9 @@
 from dataclasses import dataclass, field
 
 import numpy as np
-import pandas as pd
+from physicsnemo.core.version_check import OptionalImport
+
+pd = OptionalImport("pandas")
 
 # Recipe-side directory containing per-sensor `*_normalizations.csv` files
 # (and the ERA5 stats CSV consumed by `loaders.era5`). When unset, sensor

physicsnemo/experimental/datapipes/healda/dataset.py

@@ -52,14 +52,16 @@
 from typing import Union
 
 import numpy as np
-import pandas as pd
 import torch
 
 from physicsnemo.experimental.datapipes.healda.indexing import get_flat_indexer
 from physicsnemo.experimental.datapipes.healda.loaders.era5 import get_batch_info
 from physicsnemo.experimental.datapipes.healda.protocols import ObsLoader, Transform
 from physicsnemo.experimental.datapipes.healda.time_utils import as_cftime
 from physicsnemo.experimental.datapipes.healda.types import VariableConfig
+from physicsnemo.core.version_check import OptionalImport
+
+pd = OptionalImport("pandas")
 
 # HEALPix level-6 pixel count: 12 * 4^6
 NPIX_HPX6 = 12 * 4**6

physicsnemo/experimental/datapipes/healda/loaders/era5.py

@@ -25,7 +25,6 @@
 from typing import Optional
 
 import numpy as np
-import pandas as pd
 
 from physicsnemo.core.version_check import OptionalImport
 
@@ -35,6 +34,8 @@
 from physicsnemo.experimental.datapipes.healda.loaders.zarr_loader import NO_LEVEL, ZarrLoader
 from physicsnemo.experimental.datapipes.healda.types import BatchInfo, TimeUnit, VariableConfig
 
+pd = OptionalImport("pandas")
+
 __all__ = ["ERA5Loader", "get_batch_info"]
 
 SST_LAND_FILL_VALUE = 290
@@ -174,7 +175,7 @@ def _encode_channel(channel) -> str:
         return name
 
 
-def _load_raw_stats(config: VariableConfig) -> pd.DataFrame:
+def _load_raw_stats(config: VariableConfig) -> "pd.DataFrame":
     if config.name == "ufs":
         file_name = "ufs_v0_stats.csv"
     elif config.name == "era5":

physicsnemo/experimental/datapipes/healda/loaders/ufs_obs.py

@@ -38,7 +38,6 @@
 
 import fsspec
 import numpy as np
-import pandas as pd
 from physicsnemo.core.version_check import OptionalImport
 
 pa = OptionalImport("pyarrow")
@@ -53,6 +52,8 @@
 from physicsnemo.experimental.datapipes.healda.configs.sensors import SENSOR_CONFIGS
 from physicsnemo.experimental.datapipes.healda.transforms.obs_filtering import filter_observations
 
+pd = OptionalImport("pandas")
+
 LOCAL_CHANNEL_ID = pa.field("local_channel_id", pa.uint16())
 
 
@@ -163,7 +164,7 @@ def channel_table(self) -> pa.Table:
         array = pa.array(local_channel_ids).cast(LOCAL_CHANNEL_ID.type)
         return table.append_column(LOCAL_CHANNEL_ID, array)
 
-    def _get_interval_times(self, dt: datetime) -> pd.DatetimeIndex:
+    def _get_interval_times(self, dt: datetime) -> "pd.DatetimeIndex":
         start, end = self.obs_context_hours
         start += self.data_spacing
         return pd.date_range(
@@ -172,7 +173,7 @@ def _get_interval_times(self, dt: datetime) -> pd.DatetimeIndex:
             freq=f"{self.data_spacing}h",
         )
 
-    def _get_parquet_files_to_read(self, interval_times: pd.DatetimeIndex):
+    def _get_parquet_files_to_read(self, interval_times: "pd.DatetimeIndex"):
         required_dates = {t.strftime("%Y%m%d") for t in interval_times}
         for sensor in self.sensors:
             for date in required_dates:
@@ -247,7 +248,7 @@ def _add_channel_metadata(self, table):
             GLOBAL_CHANNEL_ID.name,
         )
 
-    async def sel_time(self, times: pd.DatetimeIndex) -> dict:
+    async def sel_time(self, times: "pd.DatetimeIndex") -> dict:
         """Load observation data for specified times.
 
         Args:

physicsnemo/experimental/datapipes/healda/loaders/zarr_loader.py

@@ -27,10 +27,11 @@
 
 import cftime
 import numpy as np
-import pandas as pd
 
 from physicsnemo.core.version_check import OptionalImport
 
+pd = OptionalImport("pandas")
+
 xr = OptionalImport("xarray")
 zarr = OptionalImport("zarr")
 _zarr_sync = OptionalImport("zarr.core.sync")

physicsnemo/experimental/datapipes/healda/protocols.py

@@ -27,8 +27,10 @@
 from typing import Any, Protocol, runtime_checkable
 
 import cftime
-import pandas as pd
 import torch
+from physicsnemo.core.version_check import OptionalImport
+
+pd = OptionalImport("pandas")
 
 
 @runtime_checkable
@@ -51,7 +53,7 @@ async def sel_time(self, times):
                 return {"obs": tables}
     """
 
-    async def sel_time(self, times: pd.DatetimeIndex) -> dict[str, list[Any]]:
+    async def sel_time(self, times: "pd.DatetimeIndex") -> dict[str, list[Any]]:
         """Load observation data for the given timestamps.
 
         Args:

physicsnemo/experimental/datapipes/healda/time_utils.py

@@ -19,7 +19,9 @@
 
 import cftime
 import numpy as np
-import pandas as pd
+from physicsnemo.core.version_check import OptionalImport
+
+pd = OptionalImport("pandas")
 
 
 def as_pydatetime(time) -> datetime.datetime:

physicsnemo/models/dlwp_healpix/HEALPixRecUNet.py

@@ -27,13 +27,13 @@
 from dataclasses import dataclass
 from typing import Any, Dict, Sequence
 
-import pandas as pd
 import torch
 from hydra.utils import instantiate
 from omegaconf import DictConfig
 
 from physicsnemo.core.meta import ModelMetaData
 from physicsnemo.core.module import Module
+from physicsnemo.core.version_check import OptionalImport
 from physicsnemo.nn.module.hpx import HEALPixFoldFaces, HEALPixUnfoldFaces
 
 from .layers import (
@@ -43,6 +43,8 @@
     _remap_obj,
 )
 
+pd = OptionalImport("pandas")
+
 logger = logging.getLogger(__name__)
 
 

physicsnemo/utils/insolation.py

@@ -14,8 +14,53 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+import datetime
+import re
+
 import numpy as np
-import pandas as pd
+
+_LEADING_YEAR = re.compile(r"\s*(-?\d{4,})")
+
+
+def _calendar_year(d) -> int:
+    """Calendar year of one date-like value, as ``pandas.Timestamp(d).year`` reports it.
+
+    For timezone-aware datetimes and ISO strings that carry a UTC offset, this is
+    the year of the *local* wall-clock time, not of the UTC instant numpy stores.
+    """
+    if isinstance(
+        d, datetime.date
+    ):  # datetime.datetime, datetime.date, pandas.Timestamp
+        return d.year
+    if isinstance(d, str):
+        match = _LEADING_YEAR.match(d)
+        if match:
+            return int(match.group(1))
+    return int(np.datetime64(d).astype("datetime64[Y]").astype(np.int64)) + 1970
+
+
+def _days_since_year_start(dates) -> np.ndarray:
+    """Fractional days from January 1st 00:00 of each date's calendar year.
+
+    Reproduces ``(np.array(dates, dtype="datetime64") - Timestamp(year, 1, 1)) / 1 day``
+    from the earlier pandas-based implementation exactly: instants are numpy's
+    conversion of ``dates`` (UTC for timezone-aware input), the year is the local
+    calendar year, and year starts are held at microsecond resolution, which is the
+    unit numpy inferred from the ``Timestamp`` objects, so unit promotion in the
+    subtraction is unchanged for every input dtype.
+    """
+    instants = np.array(dates, dtype="datetime64")
+    if np.isnat(instants).any():
+        raise ValueError("insolation: 'dates' contains NaT or None entries")
+    if isinstance(dates, np.ndarray) and np.issubdtype(dates.dtype, np.datetime64):
+        # Naive datetime64 arrays: the calendar year is unambiguous, stay vectorized.
+        years = instants.astype("datetime64[Y]")
+    else:
+        flat = np.asarray(dates, dtype=object).ravel()
+        years = np.array([_calendar_year(d) for d in flat], dtype=np.int64) - 1970
+        years = years.astype("datetime64[Y]").reshape(instants.shape)
+    start_years = years.astype("datetime64[us]")
+    return (instants - start_years) / np.timedelta64(1, "D")
 
 
 def insolation(
@@ -69,12 +114,7 @@ def insolation(
     beta = np.sqrt(1 - ecc**2.0)
 
     # Get the day of year as a float.
-    start_years = np.array(
-        [pd.Timestamp(pd.Timestamp(d).year, 1, 1) for d in dates], dtype="datetime64"
-    )
-    days_arr = (np.array(dates, dtype="datetime64") - start_years) / np.timedelta64(
-        1, "D"
-    )
+    days_arr = _days_since_year_start(dates)
     for d in range(n_dim):
         days_arr = np.expand_dims(days_arr, -1)
     # For daily max values, set the day to 0.5 and the longitude everywhere to 0 (this is approx noon)

pyproject.toml

@@ -21,14 +21,11 @@ classifiers = [
 dynamic = ["version", "optional_dependencies"]
 
 dependencies = [
-    "onnx>=1.14.0",
     "warp-lang>=1.14.0",
-    "pandas>=2.2.0",
     "nvtx>=0.2.10",
     "treelib>=1.2.5",
     "numpy>=1.22.4",
     "torch>=2.10.0",
-    "torchvision>=0.25.0a0",
     "tqdm>=4.60.0",
     "requests>=2.32.2",
     # urllib3 is a transitive dependency (via requests/botocore/etc.); pin to
@@ -304,10 +301,12 @@ nn-extras = [
 ]
 model-extras = [
     "nvidia-physicsnemo[nn-extras]",
+    "pandas>=2.2.0",
     "pyvista>=0.46.4",
     "vtk",
 ]
 datapipes-extras = [
+    "pandas>=2.2.0",
     "tfrecord",
     "dask",
     "netCDF4",