def test_mixed_nan_to_num(self):
array = np.array([1, np.nan, 1, np.nan])
# case 1 - the array has less nans than would be triggered
num_array = BaseEngineer.maxed_nan_to_num(array, nan=0, max_ratio=0.75)
assert (num_array == np.array([1, 0, 1, 0])).all()
# case 2 - the array has more nans than would be triggered
num_array = BaseEngineer.maxed_nan_to_num(array, nan=0, max_ratio=0.25)
assert num_array is None
def tif_to_np(
path_to_dataset: Path,
add_ndvi: bool,
add_ndwi: bool,
nan: float,
normalizing_dict: Optional[Dict[str, np.ndarray]],
days_per_timestep: int,
) -> TestInstance:
_, start_date, _ = cast(
Tuple[str, datetime, datetime],
BaseEngineer.process_filename(path_to_dataset.name,
include_extended_filenames=True),
)
x = BaseEngineer.load_tif(path_to_dataset,
days_per_timestep=days_per_timestep,
start_date=start_date)
lon, lat = np.meshgrid(x.x.values, x.y.values)
flat_lat, flat_lon = (
np.squeeze(lat.reshape(-1, 1), -1),
np.squeeze(lon.reshape(-1, 1), -1),
)
x_np = x.values
x_np = x_np.reshape(x_np.shape[0], x_np.shape[1],
x_np.shape[2] * x_np.shape[3])
x_np = np.moveaxis(x_np, -1, 0)
if add_ndvi:
x_np = BaseEngineer.calculate_ndvi(x_np, num_dims=3)
if add_ndwi:
x_np = BaseEngineer.calculate_ndwi(x_np, num_dims=3)
x_np = BaseEngineer.maxed_nan_to_num(x_np, nan=nan)
if normalizing_dict is not None:
x_np = (x_np - normalizing_dict["mean"]) / normalizing_dict["std"]
return TestInstance(x=x_np, lat=flat_lat, lon=flat_lon)
def test_randomly_selected_latlons_not_labels(self, monkeypatch):
def return_const_latlons(array, size, replace):
return array[:size]
monkeypatch.setattr(np.random, "choice", return_const_latlons)
lat_array = np.array([1, 2, 3])
lon_array = np.array([4, 5, 6])
# test case 1 - both the lat and lon don't match - should skip
label_lat, label_lon = 1, 4
random_lat, random_lon = BaseEngineer.randomly_select_latlon(
lat_array, lon_array, label_lat, label_lon)
assert (random_lat == 2) and (random_lon == 5)
# test case 2 - one of the lat and lon do match, but the other doesn't
# - shouldn't skip
label_lat, label_lon = 1, 2
random_lat, random_lon = BaseEngineer.randomly_select_latlon(
lat_array, lon_array, label_lat, label_lon)
assert (random_lat == 1) and (random_lon == 4)
def test_find_nearest(self):
array = np.array([1, 2, 3, 4, 5])
target = 1.1
assert BaseEngineer.find_nearest(array, target) == 1
def plot_results(model_preds: xr.Dataset,
tci_path: Path,
savepath: Path,
prefix: str = "") -> None:
multi_output = len(model_preds.data_vars) > 1
tci = sentinel_as_tci(
BaseEngineer.load_tif(tci_path,
start_date=datetime(2020, 1, 1),
days_per_timestep=30),
scale=False,
).isel(time=-1)
tci = tci.sortby("x").sortby("y")
model_preds = model_preds.sortby("lat").sortby("lon")
plt.clf()
fig, ax = plt.subplots(1,
3,
figsize=(20, 7.5),
subplot_kw={"projection": ccrs.PlateCarree()})
fig.suptitle(
f"Model results for tile with bottom left corner:"
f"\nat latitude {float(model_preds.lat.min())}"
f"\n and longitude {float(model_preds.lon.min())}",
fontsize=15,
)
# ax 1 - original
img_extent_1 = (tci.x.min(), tci.x.max(), tci.y.min(), tci.y.max())
img = np.clip(np.moveaxis(tci.values, 0, -1), 0, 1)
ax[0].set_title("True colour image")
ax[0].imshow(img,
origin="upper",
extent=img_extent_1,
transform=ccrs.PlateCarree())
args_dict = {
"origin": "upper",
"extent": img_extent_1,
"transform": ccrs.PlateCarree(),
}
if multi_output:
mask = np.argmax(model_preds.to_array().values, axis=0)
# currently, we have 10 classes (at most). It seems unlikely we will go
# above 20
args_dict["cmap"] = plt.cm.get_cmap("tab20",
len(model_preds.data_vars))
else:
mask = model_preds.prediction_0
args_dict.update({"vmin": 0, "vmax": 1})
# ax 2 - mask
ax[1].set_title("Mask")
im = ax[1].imshow(mask, **args_dict)
# finally, all together
ax[2].set_title("Mask on top of the true colour image")
ax[2].imshow(img,
origin="upper",
extent=img_extent_1,
transform=ccrs.PlateCarree())
args_dict["alpha"] = 0.3
if not multi_output:
mask = mask > 0.5
ax[2].imshow(mask, **args_dict)
colorbar_args = {
"ax": ax.ravel().tolist(),
}
if multi_output:
# This function formatter will replace integers with target names
formatter = plt.FuncFormatter(
lambda val, loc: list(model_preds.data_vars)[val])
colorbar_args.update({
"ticks": range(len(model_preds.data_vars)),
"format": formatter
})
# We must be sure to specify the ticks matching our target names
fig.colorbar(im, **colorbar_args)
plt.savefig(savepath / f"results_{prefix}{tci_path.name}.png",
bbox_inches="tight",
dpi=300)
plt.close()