Fix chunking issues in sum_AMEL and reduce_damages

CHANGELOG.md

@@ -6,6 +6,7 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ## [0.4.0] - Unreleased
 ### Added
+- Functions to concatenate input damages across batches. ([PR #83](https://github.com/ClimateImpactLab/dscim/pull/83), [@davidrzhdu](https://github.com/davidrzhdu))
 - New unit tests for [dscim/utils/input_damages.py](https://github.com/ClimateImpactLab/dscim/blob/main/src/dscim/preprocessing/input_damages.py). ([PR #68](https://github.com/ClimateImpactLab/dscim/pull/68), [@davidrzhdu](https://github.com/davidrzhdu))
 - New unit tests for [dscim/utils/rff.py](https://github.com/ClimateImpactLab/dscim/blob/main/src/dscim/utils/rff.py). ([PR #73](https://github.com/ClimateImpactLab/dscim/pull/73), [@JMGilbert](https://github.com/JMGilbert))
 - New unit tests for [dscim/dscim/preprocessing.py](https://github.com/ClimateImpactLab/dscim/blob/main/src/dscim/preprocessing/preprocessing.py). ([PR #67](https://github.com/ClimateImpactLab/dscim/pull/67), [@JMGilbert](https://github.com/JMGilbert))
@@ -23,6 +24,7 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 - Remove old/unnecessary files. ([PR #57](https://github.com/ClimateImpactLab/dscim/pull/57), [@JMGilbert](https://github.com/JMGilbert))
 - Remove unused “save_path” and “ec_cls” from `read_energy_files_parallel()`. ([PR #56](https://github.com/ClimateImpactLab/dscim/pull/56), [@davidrzhdu](https://github.com/davidrzhdu))
 ### Fixed
+- Make all input damages output files with correct chunksizes. ([PR #83](https://github.com/ClimateImpactLab/dscim/pull/83), [@JMGilbert](https://github.com/JMGilbert))
 - Add `.load()` to every loading of population data from EconVars. ([PR #82](https://github.com/ClimateImpactLab/dscim/pull/82), [@davidrzhdu](https://github.com/davidrzhdu))
 - Make `compute_ag_damages` function correctly save outputs in float32. ([PR #72](https://github.com/ClimateImpactLab/dscim/pull/72) and [PR #82](https://github.com/ClimateImpactLab/dscim/pull/82), [@davidrzhdu](https://github.com/davidrzhdu))
 - Make rff damage functions read in and save out in the proper filepath structure. ([PR #79](https://github.com/ClimateImpactLab/dscim/pull/79), [@JMGilbert](https://github.com/JMGilbert))

src/dscim/preprocessing/input_damages.py

@@ -95,6 +95,46 @@ def _parse_projection_filesys(input_path, query="exists==True"):
     return df.query(query)
 
 
+def concatenate_damage_output(damage_dir, basename, save_path):
+    """Concatenate labor/energy damage output across batches.
+
+    Parameters
+    ----------
+    damage_dir str
+        Directory containing separate labor/energy damage output files by batches.
+    basename str
+        Prefix of the damage output filenames (ex. {basename}_batch0.zarr)
+    save_path str
+        Path to save concatenated file in .zarr format
+    """
+    paths = glob.glob(f"{damage_dir}/{basename}*")
+    data = xr.open_mfdataset(paths=paths, engine="zarr")
+
+    for v in data:
+        del data[v].encoding["chunks"]
+
+    chunkies = {
+        "batch": 15,
+        "rcp": 1,
+        "gcm": 1,
+        "model": 1,
+        "ssp": 1,
+        "region": -1,
+        "year": 10,
+    }
+
+    data = data.chunk(chunkies)
+
+    for v in list(data.coords.keys()):
+        if data.coords[v].dtype == object:
+            data.coords[v] = data.coords[v].astype("unicode")
+    for v in list(data.variables.keys()):
+        if data[v].dtype == object:
+            data[v] = data[v].astype("unicode")
+
+    data.to_zarr(save_path, mode="w")
+
+
 def calculate_labor_impacts(input_path, file_prefix, variable, val_type):
     """Calculate impacts for labor results.
 
@@ -371,7 +411,7 @@ def process_batch(g):
     batches = [ds for ds in batches if ds is not None]
     chunkies = {
         "rcp": 1,
-        "region": 24378,
+        "region": -1,
         "gcm": 1,
         "year": 10,
         "model": 1,
@@ -738,12 +778,21 @@ def prep(
         ).expand_dims({"gcm": [gcm]})
 
         damages = damages.chunk(
-            {"batch": 15, "ssp": 1, "model": 1, "rcp": 1, "gcm": 1, "year": 10}
+            {
+                "batch": 15,
+                "ssp": 1,
+                "model": 1,
+                "rcp": 1,
+                "gcm": 1,
+                "year": 10,
+                "region": -1,
+            }
         )
         damages.coords.update({"batch": [f"batch{i}" for i in damages.batch.values]})
 
         # convert to EPA VSL
         damages = damages * 0.90681089
+        damages = damages.astype(np.float32)
 
         for v in list(damages.coords.keys()):
             if damages.coords[v].dtype == object:
@@ -790,6 +839,15 @@ def coastal_inputs(
             )
         else:
             d = d.sel(adapt_type=adapt_type, vsl_valuation=vsl_valuation, drop=True)
+            chunkies = {
+                "batch": 15,
+                "ssp": 1,
+                "model": 1,
+                "slr": 1,
+                "year": 10,
+                "region": -1,
+            }
+            d = d.chunk(chunkies)
             d.to_zarr(
                 f"{path}/coastal_damages_{version}-{adapt_type}-{vsl_valuation}.zarr",
                 consolidated=True,

src/dscim/preprocessing/preprocessing.py

@@ -102,6 +102,24 @@ def reduce_damages(
                 xr.open_zarr(damages).chunks["batch"][0] == 15
             ), "'batch' dim on damages does not have chunksize of 15. Please rechunk."
 
+            if "coastal" not in sector:
+                chunkies = {
+                    "rcp": 1,
+                    "region": -1,
+                    "gcm": 1,
+                    "year": 10,
+                    "model": 1,
+                    "ssp": 1,
+                }
+            else:
+                chunkies = {
+                    "region": -1,
+                    "slr": 1,
+                    "year": 10,
+                    "model": 1,
+                    "ssp": 1,
+                }
+
             ce_batch_dims = [i for i in gdppc.dims] + [
                 i for i in ds.dims if i not in gdppc.dims and i != "batch"
             ]
@@ -110,15 +128,14 @@ def reduce_damages(
                 i for i in gdppc.region.values if i in ce_batch_coords["region"]
             ]
             ce_shapes = [len(ce_batch_coords[c]) for c in ce_batch_dims]
-            ce_chunks = [xr.open_zarr(damages).chunks[c][0] for c in ce_batch_dims]
 
     template = xr.DataArray(
-        da.empty(ce_shapes, chunks=ce_chunks),
+        da.empty(ce_shapes),
         dims=ce_batch_dims,
         coords=ce_batch_coords,
-    )
+    ).chunk(chunkies)
 
-    other = xr.open_zarr(damages)
+    other = xr.open_zarr(damages).chunk(chunkies)
 
     out = other.map_blocks(
         ce_from_chunk,
@@ -205,7 +222,21 @@ def sum_AMEL(
         for sector in sectors:
             print(f"Opening {sector},{params[sector]['sector_path']}")
             ds = xr.open_zarr(params[sector]["sector_path"], consolidated=True)
-            ds = ds[params[sector][var]].rename(var)
+            ds = (
+                ds[params[sector][var]]
+                .rename(var)
+                .chunk(
+                    {
+                        "batch": 15,
+                        "ssp": 1,
+                        "model": 1,
+                        "rcp": 1,
+                        "gcm": 1,
+                        "year": 10,
+                        "region": -1,
+                    }
+                )
+            )
             ds = xr.where(np.isinf(ds), np.nan, ds)
             datasets.append(ds)
 

tests/test_input_damages.py

@@ -9,6 +9,7 @@
 from dscim.menu.simple_storage import EconVars
 from dscim.preprocessing.input_damages import (
     _parse_projection_filesys,
+    concatenate_damage_output,
     calculate_labor_impacts,
     concatenate_labor_damages,
     calculate_labor_batch_damages,
@@ -31,7 +32,7 @@ def test_parse_projection_filesys(tmp_path):
     """
     Test that parse_projection_filesys correctly retrieves projection system output structure
     """
-    rcp = ["rcp85", "rcp45"]
+    rcp = ["rcp45", "rcp85"]
     gcm = ["ACCESS1-0", "GFDL-CM3"]
     model = ["high", "low"]
     ssp = [f"SSP{n}" for n in range(2, 4)]
@@ -45,14 +46,14 @@ def test_parse_projection_filesys(tmp_path):
                         os.makedirs(os.path.join(tmp_path, b, r, g, m, s))
 
     out_expected = {
-        "batch": list(chain(repeat("batch9", 16), repeat("batch6", 16))),
-        "rcp": list(chain(repeat("rcp85", 8), repeat("rcp45", 8))) * 2,
+        "batch": list(chain(repeat("batch6", 16), repeat("batch9", 16))),
+        "rcp": list(chain(repeat("rcp45", 8), repeat("rcp85", 8))) * 2,
         "gcm": list(chain(repeat("ACCESS1-0", 4), repeat("GFDL-CM3", 4))) * 4,
         "model": list(chain(repeat("high", 2), repeat("low", 2))) * 8,
         "ssp": ["SSP2", "SSP3"] * 16,
         "path": [
             os.path.join(tmp_path, b, r, g, m, s)
-            for b in ["batch9", "batch6"]
+            for b in ["batch6", "batch9"]
             for r in rcp
             for g in gcm
             for m in model
@@ -64,11 +65,81 @@ def test_parse_projection_filesys(tmp_path):
     df_out_expected = pd.DataFrame(out_expected)
 
     df_out_actual = _parse_projection_filesys(input_path=tmp_path)
+    df_out_actual = df_out_actual.sort_values(
+        by=["batch", "rcp", "gcm", "model", "ssp"]
+    )
     df_out_actual.reset_index(drop=True, inplace=True)
 
     pd.testing.assert_frame_equal(df_out_expected, df_out_actual)
 
 
+def test_concatenate_damage_output(tmp_path):
+    """
+    Test that concatenate_damage_output correctly concatenates damages across batches and saves to a single zarr file
+    """
+    d = os.path.join(tmp_path, "concatenate_in")
+    if not os.path.exists(d):
+        os.makedirs(d)
+
+    for b in ["batch6", "batch9"]:
+        ds_in = xr.Dataset(
+            {
+                "delta_rebased": (
+                    ["ssp", "rcp", "model", "gcm", "batch", "year", "region"],
+                    np.full((2, 2, 2, 2, 1, 2, 2), 1),
+                ),
+                "histclim_rebased": (
+                    ["ssp", "rcp", "model", "gcm", "batch", "year", "region"],
+                    np.full((2, 2, 2, 2, 1, 2, 2), 2),
+                ),
+            },
+            coords={
+                "batch": (["batch"], [b]),
+                "gcm": (["gcm"], ["ACCESS1-0", "BNU-ESM"]),
+                "model": (["model"], ["IIASA GDP", "OECD Env-Growth"]),
+                "rcp": (["rcp"], ["rcp45", "rcp85"]),
+                "region": (["region"], ["ZWE.test_region", "USA.test_region"]),
+                "ssp": (["ssp"], ["SSP2", "SSP3"]),
+                "year": (["year"], [2020, 2099]),
+            },
+        )
+
+        infile = os.path.join(d, f"test_insuffix_{b}.zarr")
+
+        ds_in.to_zarr(infile)
+
+    ds_out_expected = xr.Dataset(
+        {
+            "delta_rebased": (
+                ["ssp", "rcp", "model", "gcm", "batch", "year", "region"],
+                np.full((2, 2, 2, 2, 2, 2, 2), 1),
+            ),
+            "histclim_rebased": (
+                ["ssp", "rcp", "model", "gcm", "batch", "year", "region"],
+                np.full((2, 2, 2, 2, 2, 2, 2), 2),
+            ),
+        },
+        coords={
+            "batch": (["batch"], ["batch6", "batch9"]),
+            "gcm": (["gcm"], ["ACCESS1-0", "BNU-ESM"]),
+            "model": (["model"], ["IIASA GDP", "OECD Env-Growth"]),
+            "rcp": (["rcp"], ["rcp45", "rcp85"]),
+            "region": (["region"], ["ZWE.test_region", "USA.test_region"]),
+            "ssp": (["ssp"], ["SSP2", "SSP3"]),
+            "year": (["year"], [2020, 2099]),
+        },
+    )
+
+    concatenate_damage_output(
+        damage_dir=d,
+        basename="test_insuffix",
+        save_path=os.path.join(d, "concatenate.zarr"),
+    )
+    ds_out_actual = xr.open_zarr(os.path.join(d, "concatenate.zarr"))
+
+    xr.testing.assert_equal(ds_out_expected, ds_out_actual)
+
+
 @pytest.fixture
 def labor_in_val_fixture(tmp_path):
     """
@@ -1030,11 +1101,11 @@ def test_prep_mortality_damages(
         {
             "delta": (
                 ["gcm", "batch", "ssp", "rcp", "model", "year", "region"],
-                np.full((2, 2, 2, 2, 2, 2, 2), -0.90681089),
+                np.float32(np.full((2, 2, 2, 2, 2, 2, 2), -0.90681089)),
             ),
             "histclim": (
                 ["gcm", "batch", "ssp", "rcp", "model", "year", "region"],
-                np.full((2, 2, 2, 2, 2, 2, 2), 2 * 0.90681089),
+                np.float32(np.full((2, 2, 2, 2, 2, 2, 2), 2 * 0.90681089)),
             ),
         },
         coords={