def test_regrid():
# use conservative regridding as an example,
# since it is the most well-tested studied one in papers
# TODO: possible to break this long test into smaller tests?
# not easy due to strong dependencies.
grid_in = esmf_grid(lon_in.T, lat_in.T)
grid_out = esmf_grid(lon_out.T, lat_out.T)
# no corner info yet, should not be able to use conservative
with pytest.raises(ValueError):
esmf_regrid_build(grid_in, grid_out, 'conservative')
# now add corners
add_corner(grid_in, lon_b_in.T, lat_b_in.T)
add_corner(grid_out, lon_b_out.T, lat_b_out.T)
# also write to file for scipy regridding
filename = 'test_weights.nc'
if os.path.exists(filename):
os.remove(filename)
regrid = esmf_regrid_build(grid_in,
grid_out,
'conservative',
filename=filename)
assert regrid.regrid_method is ESMF.RegridMethod.CONSERVE
# apply regridding using ESMPy's native method
data_out_esmpy = esmf_regrid_apply(regrid, data_in.T).T
rel_err = (data_out_esmpy - data_ref) / data_ref # relative error
assert np.max(np.abs(rel_err)) < 0.05
# apply regridding using scipy
weights = read_weights(filename, lon_in.size, lon_out.size)
shape_in = lon_in.shape
shape_out = lon_out.shape
data_out_scipy = apply_weights(weights, data_in, shape_in, shape_out)
# must be exactly the same as esmpy's result!
# TODO: this fails once but I cannot replicate it.
# Maybe assert_equal is too strict for scipy vs esmpy comparision
assert_equal(data_out_scipy, data_out_esmpy)
# finally, test broadcasting with scipy
# TODO: need to test broadcasting with ESMPy backend?
# We only use Scipy in frontend, and ESMPy is just for backend benchmark
# However, it is useful to compare performance and show scipy is 3x faster
data4D_out = apply_weights(weights, data4D_in, shape_in, shape_out)
# data over broadcasting dimensions should agree
assert_almost_equal(data4D_in.mean(axis=(2, 3)),
data4D_out.mean(axis=(2, 3)),
decimal=10)
# clean-up
esmf_regrid_finalize(regrid)
os.remove(filename)
def test_read_weights(tmp_path):
fn = tmp_path / 'weights.nc'
grid_in = Grid.from_xarray(lon_in.T, lat_in.T)
grid_out = Grid.from_xarray(lon_out.T, lat_out.T)
regrid_memory = esmf_regrid_build(grid_in, grid_out, method='bilinear')
esmf_regrid_build(grid_in, grid_out, method='bilinear', filename=str(fn))
w = regrid_memory.get_weights_dict(deep_copy=True)
sm = read_weights(w, lon_in.size, lon_out.size)
# Test Path and string to netCDF file against weights dictionary
np.testing.assert_array_equal(
read_weights(fn, lon_in.size, lon_out.size).todense(), sm.todense())
np.testing.assert_array_equal(
read_weights(str(fn), lon_in.size, lon_out.size).todense(),
sm.todense())
# Test xr.Dataset
np.testing.assert_array_equal(
read_weights(xr.open_dataset(fn), lon_in.size, lon_out.size).todense(),
sm.todense(),
)
# Test COO matrix
np.testing.assert_array_equal(
read_weights(sm, lon_in.size, lon_out.size).todense(), sm.todense())
# Test failures
with pytest.raises(IOError):
read_weights(tmp_path / 'wrong_file.nc', lon_in.size, lon_out.size)
with pytest.raises(ValueError):
read_weights({}, lon_in.size, lon_out.size)
with pytest.raises(ValueError):
ds = xr.open_dataset(fn)
read_weights(ds.drop_vars('col'), lon_in.size, lon_out.size)