Merge pull request #125 from quantifyearth/mwd-import-speed

Load heavier imports lazily

Author: mdales

CHANGES.md

@@ -1,3 +1,9 @@
+## v1.12.5 (18/02/2026)
+
+### Changed
+
+* Reduce initial import time.
+
 ## v1.12.4 (12/02/2026)
 
 ### Added

pyproject.toml

@@ -6,7 +6,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "yirgacheffe"
-version = "1.12.4"
+version = "1.12.5"
 description = "Abstraction of gdal datasets for doing basic math operations"
 readme = "README.md"
 authors = [{ name = "Michael Dales", email = "mwd24@cam.ac.uk" }]
@@ -31,6 +31,7 @@ dependencies = [
     "tomli",
     "h3",
     "pyproj",
+    "lazy_loader"
 ]
 requires-python = ">=3.10"
 

yirgacheffe/_backends/numpy.py

@@ -4,7 +4,6 @@
 
 from osgeo import gdal
 import numpy as np
-import torch
 
 from .enumeration import operators as op
 from .enumeration import dtype
@@ -63,15 +62,18 @@ def init() -> None:
     gdal.SetCacheMax(constants.GDAL_CACHE_LIMIT)
 
 def conv2d_op(data, weights):
+    # Pytorch is very slow to import, and currently a niche use case
+    # for just this operator. As such we defer import to a local one
+    # to improve initial load times.
+    import torch # pylint: disable=C0415
+
     # torch wants to process dimensions of channels of width of height
     # Which is why both the data and weights get nested into two arrays here,
     # and then we have to unpack it from that nesting.
-
     preped_weights = np.array([[weights]])
     conv = torch.nn.Conv2d(1, 1, weights.shape, bias=False)
     conv.weight = torch.nn.Parameter(torch.from_numpy(preped_weights))
     preped_data = torch.from_numpy(np.array([[data]]))
-
     res = conv(preped_data)
     return res.detach().numpy()[0][0]
 

yirgacheffe/_datatypes/mapprojection.py

@@ -2,11 +2,14 @@
 import math
 from functools import lru_cache
 
-import pyproj
-from pyproj import CRS
+import lazy_loader as lazy # type: ignore
 
 from .pixelscale import PixelScale
 
+# Pyproj is relatively slow to import, which was adding reasonably to
+# the import time for yirgacheffe, so make it lazy
+pyproj = lazy.load('pyproj')
+
 # As per https://xkcd.com/2170/, we need to stop caring about floating point
 # accuracy at some point as it becomes problematic.
 # The value here is 1 meter, given that geo data that we've been working with
@@ -27,7 +30,7 @@
 # that makes 1600 rasters take 72 seconds without this fix, and 2 seconds with it.
 @lru_cache(maxsize=128)
 def _get_projection_string(provided_name: str) -> str:
-    crs = CRS.from_string(provided_name)
+    crs = pyproj.CRS.from_string(provided_name)
     epsg = crs.to_epsg()
     if epsg is not None:
         return f"EPSG:{epsg}"
@@ -64,7 +67,7 @@ class MapProjection:
 
     def __init__(self, projection_string: str, xstep: float, ystep: float) -> None:
         try:
-            self.crs = CRS.from_string(projection_string)
+            self.crs = pyproj.CRS.from_string(projection_string)
         except pyproj.exceptions.CRSError as exc:
             raise ValueError(f"Invalid projection: {projection_string}") from exc
         self.xstep = xstep

yirgacheffe/layers/h3layer.py

@@ -3,14 +3,18 @@
 from math import ceil, floor
 from typing import Any
 
-import h3
+import lazy_loader as lazy # type: ignore
 import numpy as np
 
 from .._datatypes import Area, MapProjection, Window
 from .base import YirgacheffeLayer
 from .._backends import backend
 from .._backends.enumeration import dtype as DataType
 
+# H3 is relatively slow to import, which was adding reasonably to
+# the import time for yirgacheffe, so make it lazy
+h3 = lazy.load('h3')
+
 class H3CellLayer(YirgacheffeLayer):
 
     def __init__(self, cell_id: str, projection: MapProjection):