add module doc string with installation instructions to train_mace.py

fix pyright possibly unbound variable errors

.pre-commit-config.yaml

@@ -7,7 +7,7 @@ default_install_hook_types: [pre-commit, commit-msg]
 
 repos:
   - repo: https://github.com/astral-sh/ruff-pre-commit
-    rev: v0.3.0
+    rev: v0.3.3
     hooks:
       - id: ruff
         args: [--fix]
@@ -30,7 +30,7 @@ repos:
       - id: trailing-whitespace
 
   - repo: https://github.com/pre-commit/mirrors-mypy
-    rev: v1.8.0
+    rev: v1.9.0
     hooks:
       - id: mypy
         additional_dependencies: [types-pyyaml, types-requests]
@@ -56,7 +56,7 @@ repos:
         exclude: ^(site/src/figs/.+\.svelte|data/wbm/20.+\..+|site/src/(routes|figs).+\.(yaml|json)|changelog.md)$
 
   - repo: https://github.com/pre-commit/mirrors-eslint
-    rev: v9.0.0-beta.1
+    rev: v9.0.0-beta.2
     hooks:
       - id: eslint
         types: [file]
@@ -80,7 +80,7 @@ repos:
       - id: check-github-actions
 
   - repo: https://github.com/RobertCraigie/pyright-python
-    rev: v1.1.352
+    rev: v1.1.355
     hooks:
       - id: pyright
         args: [--level, error]

data/mp/build_phase_diagram.py

@@ -105,13 +105,14 @@
 
 
 # %%
-df_mp["our_mp_e_form"] = [
+e_form_us = "e_form_us"
+df_mp[e_form_us] = [
     get_e_form_per_atom(mp_computed_entries[mp_id]) for mp_id in df_mp.index
 ]
 
 
 # make sure get_form_energy_per_atom() reproduces MP formation energies
-ax = pymatviz.density_scatter(df_mp[Key.form_energy], df_mp["our_mp_e_form"])
+ax = pymatviz.density_scatter(df_mp[Key.form_energy], df_mp[e_form_us])
 ax.set(
     title="MP Formation Energy Comparison",
     xlabel="MP Formation Energy (eV/atom)",

data/wbm/compare_cse_vs_ce_mp_2020_corrections.py

@@ -18,7 +18,7 @@
 
 """
 NOTE MaterialsProject2020Compatibility takes structural information into account when
-correcting energies (only applies to certain oxides and sulfides). Always use
+correcting energies (for certain oxides and sulfides). Always use
 ComputedStructureEntry, not ComputedEntry when applying corrections.
 """
 

data/wbm/compile_wbm_test_set.py

@@ -161,8 +161,8 @@ def increment_wbm_material_id(wbm_id: str) -> str:
 assert df_wbm.index[-1] == "wbm-5-23308"
 
 df_wbm[Key.init_struct] = df_wbm.pop("org")
-df_wbm["final_structure"] = df_wbm.pop("opt")
-assert list(df_wbm.columns) == [Key.init_struct, "final_structure"]
+df_wbm[Key.final_struct] = df_wbm.pop("opt")
+assert list(df_wbm.columns) == [Key.init_struct, Key.final_struct]
 
 
 # %% download WBM ComputedStructureEntries from
@@ -247,7 +247,7 @@ def increment_wbm_material_id(wbm_id: str) -> str:
 ]
 
 df_wbm["composition_from_final_struct"] = [
-    Structure.from_dict(struct).composition for struct in tqdm(df_wbm.final_structure)
+    Structure.from_dict(struct).composition for struct in tqdm(df_wbm[Key.final_struct])
 ]
 
 # all but 1 composition matches between CSE and final structure
@@ -499,7 +499,9 @@ def fix_bad_struct_index_mismatch(material_id: str) -> str:
 for mat_id, cse in df_wbm[Key.cse].items():
     assert mat_id == cse["entry_id"], f"{mat_id} != {cse['entry_id']}"
 
-df_wbm["cse"] = [ComputedStructureEntry.from_dict(dct) for dct in tqdm(df_wbm[Key.cse])]
+df_wbm[Key.cse] = [
+    ComputedStructureEntry.from_dict(dct) for dct in tqdm(df_wbm[Key.cse])
+]
 # raw WBM ComputedStructureEntries have no energy corrections applied:
 assert all(cse.uncorrected_energy == cse.energy for cse in df_wbm.cse)
 # summary and CSE n_sites match
@@ -548,7 +550,7 @@ def fix_bad_struct_index_mismatch(material_id: str) -> str:
 # takes ~20 min at 200 it/s for 250k entries in WBM
 assert Key.each_true not in df_summary
 
-for mat_id, cse in tqdm(df_wbm.cse.items(), total=len(df_wbm)):
+for mat_id, cse in tqdm(df_wbm[Key.cse].items(), total=len(df_wbm)):
     assert mat_id == cse.entry_id, f"{mat_id=} != {cse.entry_id=}"
     assert cse.entry_id in df_summary.index, f"{cse.entry_id=} not in df_summary"
 
@@ -562,7 +564,7 @@ def fix_bad_struct_index_mismatch(material_id: str) -> str:
 assert sum(df_wbm.index != df_summary.index) == 0
 
 for row in tqdm(df_wbm.itertuples(), total=len(df_wbm), desc="ML energies to CSEs"):
-    mat_id, cse, formula = row.Index, row.cse, row.formula_from_cse
+    mat_id, cse, formula = row.Index, row[Key.cse], row.formula_from_cse
     assert mat_id == cse.entry_id, f"{mat_id=} != {cse.entry_id=}"
     assert mat_id in df_summary.index, f"{mat_id=} not in df_summary"
 
@@ -665,12 +667,9 @@ def fix_bad_struct_index_mismatch(material_id: str) -> str:
 
 # %% only here to load data for later inspection
 if False:
-    wbm_summary_path = f"{WBM_DIR}/2022-10-19-wbm-summary.csv.gz"
-    df_summary = pd.read_csv(wbm_summary_path).set_index(Key.mat_id)
-    df_wbm = pd.read_json(
-        f"{WBM_DIR}/2022-10-19-wbm-computed-structure-entries+init-structs.json.bz2"
-    ).set_index(Key.mat_id)
+    df_summary = pd.read_csv(DATA_FILES.wbm_summary).set_index(Key.mat_id)
+    df_wbm = pd.read_json(DATA_FILES.wbm_cses_plus_init_structs).set_index(Key.mat_id)
 
-    df_wbm["cse"] = [
+    df_wbm[Key.cse] = [
         ComputedStructureEntry.from_dict(dct) for dct in tqdm(df_wbm[Key.cse])
     ]

matbench_discovery/plots.py

@@ -106,6 +106,7 @@ def hist_classified_stable_vs_hull_dist(
     clf_col, value_name = "classified", "count"
 
     df_plot = pd.DataFrame()
+    each_true_pos = each_true_neg = each_false_neg = each_false_pos = None
 
     for facet, df_group in (
         df.groupby(kwargs["facet_col"]) if "facet_col" in kwargs else [(None, df)]

models/alignn_ff/readme.md

@@ -8,7 +8,7 @@ This effort was aborted for the following reasons:
 1. **Training difficulties**: ALIGNN-FF proved to be very resource-hungry. [12 GB of MPtrj training data](https://figshare.com/articles/dataset/23713842) turned into 600 GB of ALIGNN graph data. This forces small batch size even on nodes with large GPU memory, which slowed down training.
 1. **Ineffectiveness of fine-tuning**: Efforts to fine-tune the ALIGNN-FF WT10 model on the CHGNet data suffered high initial loss, even worse than the untrained model, indicating significant dataset incompatibility.
 
-The decision to abort adding ALIGNN FF to Matbench Discovery v1 was made after weeks of work due to ongoing technical challenges and resource limitations. See the [PR discussion](https://github.com/janosh/matbench-discovery/pull/47) for further details.
+The decision to abort testing ALIGNN FF was made after weeks of work due to ongoing technical challenges and resource limitations. See the [PR discussion](https://github.com/janosh/matbench-discovery/pull/47) for further details.
 
 ## Fine-tuning
 

models/mace/readme.md

@@ -1,7 +1,7 @@
 ## MACE formation energy predictions on WBM test set
 
 The original MACE submission used the 2M parameter checkpoint [`2023-08-14-mace-yuan-trained-mptrj-04.model`](https://figshare.com/ndownloader/files/42374049) trained by Yuan Chiang on the [MPtrj dataset](https://figshare.com/articles/dataset/23713842).
-We initially tested the `2023-07-14-mace-universal-2-big-128-6.model` checkpoint trained on the much smaller [original M3GNet training set](https://figshare.com/articles/dataset/MPF_2021_2_8/19470599) which we received directly from Ilyes Batatia. MPtrj-trained MACE performed better and was used for the Matbench Discovery v1 submission.
+We initially tested the `2023-07-14-mace-universal-2-big-128-6.model` checkpoint trained on the much smaller [original M3GNet training set](https://figshare.com/articles/dataset/MPF_2021_2_8/19470599) which we received directly from Ilyes Batatia. MPtrj-trained MACE performed better and was used for the Matbench Discovery submission.
 
 In late October (received 2023-10-29), Philipp Benner trained a much larger 16M parameter MACE for over 100 epochs in MPtrj which achieved an (at the time SOTA) F1 score of 0.64 and DAF of 3.13.
 
@@ -21,12 +21,7 @@ MACE relaxed each test set structure until the maximum force in the training set
 
 #### Training
 
-- `loss="uip"`
-- `energy_weight=1`
-- `forces_weight=1`
-- `stress_weight=0.01`
-- `r_max=6.0`
-- `lr=0.005`
-- `batch_size=10`
+See the module doc string in `train_mace.py` for how to install MACE for multi-GPU training.
+A single-GPU training script that works with the current [MACE PyPI release](https://pypi.org/project/mace-torch) (v0.3.4 as of 2024-03-21) could be provided if there's interest.
 
-We used conditional loss weighting. We did _not_ use MACE's newest attention block feature which in our testing performed significantly worse than `RealAgnosticResidualInteractionBlock`.
+Our training used conditional loss weighting. We did _not_ use MACE's newest attention block feature which in our testing performed significantly worse than `RealAgnosticResidualInteractionBlock`.

models/mace/train_mace.py

@@ -1,3 +1,17 @@
+"""
+This script requires installing the as-yet unmerged multi-GPU branch
+in the MACE repo.
+pip install git+https://github.com/ACEsuit/mace@multi-GPU
+Plan is to merge it into main and then release to PyPI. At that point,
+the install command will be:
+pip install mace-torch
+
+If you want to fine-tune an existing MACE checkpoint rather than train a
+model from scratch, install the foundations branch instead which has an interface
+just for that.
+pip install git+https://github.com/ACEsuit/mace@foundations
+"""
+
 from __future__ import annotations
 
 import ast
@@ -36,9 +50,6 @@
 __date__ = "2023-09-18"
 
 
-# This script requires installing MACE.
-# pip install git+https://github.com/ACEsuit/mace
-
 module_dir = os.path.dirname(__file__)
 
 slurm_vars = slurm_submit(
@@ -77,8 +88,8 @@ def main(**kwargs: Any) -> None:
     if args.distributed:
         try:
             distr_env = DistributedEnvironment()
-        except Exception as e:
-            print(f"Error specifying environment for distributed training: {e}")
+        except Exception as exc:
+            print(f"Error specifying environment for distributed training: {exc}")
             return
         world_size = distr_env.world_size
         local_rank = distr_env.local_rank
@@ -122,10 +133,10 @@ def main(**kwargs: Any) -> None:
 
     # Data preparation
     if args.train_file.endswith(".xyz"):
-        if args.valid_file is not None:
-            assert args.valid_file.endswith(
-                ".xyz"
-            ), "valid_file if given must be same format as train_file"
+        if args.valid_file is not None and not args.valid_file.endswith(".xyz"):
+            raise RuntimeError(
+                f"valid_file must be .xyz if train_file is .xyz, got {args.valid_file}"
+            )
         config_type_weights = get_config_type_weights(args.config_type_weights)
         collections, atomic_energies_dict = get_dataset_from_xyz(
             train_path=args.train_file,
@@ -150,7 +161,7 @@ def main(**kwargs: Any) -> None:
             f"{len(collections.valid)}, tests=[{test_config_lens}]"
         )
     elif args.train_file.endswith(".h5"):
-        atomic_energies_dict = None
+        atomic_energies_dict = collections = None
     else:
         raise RuntimeError(
             f"train_file must be either .xyz or .h5, got {args.train_file}"
@@ -485,8 +496,8 @@ def main(**kwargs: Any) -> None:
             f"{args.swa_forces_weight}, learning rate : {args.swa_lr}"
         )
         if args.loss == "forces_only":
-            print("Can not select swa with forces only loss.")
-        elif args.loss == "virials":
+            raise RuntimeError("Can not select SWA with forces-only loss.")
+        if args.loss == "virials":
             loss_fn_energy = modules.WeightedEnergyForcesVirialsLoss(
                 energy_weight=args.swa_energy_weight,
                 forces_weight=args.swa_forces_weight,

pyproject.toml

@@ -55,7 +55,7 @@ running-models = [
   # when attempting PyPI publish
   # "aviary@git+https://github.com/CompRhys/aviary",
   "alignn",
-  "chgnet",
+  "chgnet>=0.3.0",
   "jarvis-tools",
   "m3gnet",
   "mace-torch",
@@ -93,6 +93,7 @@ ignore = [
   "D205",    # blank-line-after-summary
   "DTZ005",
   "E731",    # lambda-assignment
+  "EM101",
   "EM102",
   "FBT001",
   "FBT002",

scripts/update_wandb_runs.py

@@ -22,12 +22,12 @@
 
 
 # %%
-df = pd.DataFrame([run.config | dict(run.summary) for run in runs])
-df[["display_name", "id"]] = [(run.display_name, run.id) for run in runs]
+df_runs = pd.DataFrame([run.config | dict(run.summary) for run in runs])
+df_runs[["display_name", "id"]] = [(run.display_name, run.id) for run in runs]
 
 
 # %%
-df.isna().sum()
+df_runs.isna().sum()
 
 
 # %% --- Update run metadata ---
@@ -41,9 +41,9 @@
         "mace-wbm-IS2RE-debug-", "mace-wbm-IS2RE-"
     )
 
-    for x in (Task.IS2RE, "ES2RE"):
-        if x in run.display_name:
-            new_config["task_type"] = x
+    for key in (Task.IS2RE, Task.RS2RE):
+        if key in run.display_name:
+            new_config["task_type"] = key
 
     if "SLURM_JOB_ID" in new_config:
         new_config["slurm_job_id"] = new_config.pop("SLURM_JOB_ID")

site/src/routes/preprint/references.yaml

@@ -1426,7 +1426,7 @@ references:
   URL: https://www.nature.com/articles/s41467-021-23339-x
   volume: '12'
 
-- id: mok_direction-based_2022
+- id: mok_directionbased_2022
   accessed:
     - year: 2022
       month: 10
@@ -1438,7 +1438,7 @@ references:
       given: Jongseung
     - family: Back
       given: Seoin
-  citation-key: mok_direction-based_2022
+  citation-key: mok_directionbased_2022
   DOI: 10.26434/chemrxiv-2022-dp58c
   genre: preprint
   issued: