fix models/bowsr/test_bowsr.py passing unrelaxed, not bowsr-relaxed structure to megnet

Author: janosh

matbench_discovery/__init__.py

@@ -15,7 +15,7 @@
 # directory to store model checkpoints downloaded from wandb cloud storage
 CHECKPOINT_DIR = f"{ROOT}/wandb/checkpoints"
 # wandb <entity>/<project name> to record new runs to
-WANDB_PATH = "materialsproject/matbench-discovery"
+WANDB_PATH = "janosh/matbench-discovery"
 
 timestamp = f"{datetime.now():%Y-%m-%d@%H-%M-%S}"
 today = timestamp.split("@")[0]

models/bowsr/test_bowsr.py

@@ -32,10 +32,10 @@
 #     Some of your processes may have been killed by the cgroup out-of-memory handler.
 slurm_mem_per_node = 12000
 # set large job array size for fast testing/debugging
-slurm_array_task_count = 1000
+slurm_array_task_count = 500
 # see https://stackoverflow.com/a/55431306 for how to change array throttling
 # post submission
-slurm_max_parallel = 50
+slurm_max_parallel = 100
 energy_model = "megnet"
 job_name = f"bowsr-{energy_model}-wbm-{task_type}{'-debug' if DEBUG else ''}"
 out_dir = os.environ.get("SBATCH_OUTPUT", f"{module_dir}/{today}-{job_name}")
@@ -137,7 +137,9 @@
         structure_bowsr, energy_bowsr = optimizer.get_optimized_structure_and_energy()
 
         results = {
-            f"e_form_per_atom_bowsr_{energy_model}": model.predict_energy(structure),
+            f"e_form_per_atom_bowsr_{energy_model}": model.predict_energy(
+                structure_bowsr
+            ),
             "structure_bowsr": structure_bowsr,
             "energy_bowsr": energy_bowsr,
         }

models/cgcnn/test_cgcnn.py

@@ -70,12 +70,16 @@
 df[input_col] = [Structure.from_dict(x) for x in tqdm(df[input_col], disable=None)]
 
 filters = {
-    "created_at": {"$gt": "2022-12-03", "$lt": "2022-12-04"},
-    "display_name": {"$regex": "^train-cgcnn-robust-augment=3-"},
+    # "display_name": {"$regex": "^train-cgcnn-robust-augment=3-"},
+    # "created_at": {"$gt": "2022-12-03", "$lt": "2022-12-04"},
+    "display_name": {"$regex": "^train-cgcnn-robust-augment=0-"},
+    "created_at": {"$gt": "2023-01-09", "$lt": "2023-01-10"},
 }
 runs = wandb.Api().runs(WANDB_PATH, filters=filters)
+assert (
+    len(runs) == 10
+), f"Expected 10 runs, got {len(runs)} filtering {WANDB_PATH=} with {filters=}"
 
-assert len(runs) == 10, f"Expected 10 runs, got {len(runs)} for {filters=}"
 for idx, run in enumerate(runs):
     for key, val in run.config.items():
         if val == runs[0].config[key] or key.startswith(("slurm_", "timestamp")):
@@ -96,6 +100,7 @@
     input_col=input_col,
     wandb_run_filters=filters,
     slurm_vars=slurm_vars,
+    training_run_ids=[run.id for run in runs],
 )
 
 wandb.init(project="matbench-discovery", name=job_name, config=run_params)

models/cgcnn/train_cgcnn.py

@@ -28,7 +28,7 @@
 target_col = "formation_energy_per_atom"
 input_col = "structure"
 id_col = "material_id"
-augment = 1  # 0 for no augmentation, n>1 means train on n perturbations of each crystal
+augment = 0  # 0 for no augmentation, n>1 means train on n perturbations of each crystal
 # in the training set all assigned the same original target energy
 job_name = f"train-cgcnn-robust-{augment=}{'-debug' if DEBUG else ''}"
 print(f"{job_name=}")

models/wrenformer/test_wrenformer.py

@@ -56,8 +56,10 @@
     "display_name": {"$regex": "wrenformer-robust"},
 }
 runs = wandb.Api().runs(WANDB_PATH, filters=filters)
+assert (
+    len(runs) == 10
+), f"Expected 10 runs, got {len(runs)} filtering {WANDB_PATH=} with {filters=}"
 
-assert len(runs) == 10, f"Expected 10 runs, got {len(runs)} for {filters=}"
 for idx, run in enumerate(runs):
     for key, val in run.config.items():
         if val == runs[0].config[key] or key.startswith(("slurm_", "timestamp")):
@@ -78,6 +80,7 @@
     input_col=input_col,
     wandb_run_filters=filters,
     slurm_vars=slurm_vars,
+    training_run_ids=[run.id for run in runs],
 )
 
 wandb.init(project="matbench-discovery", name=job_name, config=run_params)

scripts/metrics_table.py

@@ -27,7 +27,7 @@
         ),
     ),
     "Voronoi RF": dict(
-        n_runs=70,
+        n_runs=68,
         filters=dict(
             created_at={"$gt": "2022-11-17", "$lt": "2022-11-28"},
             display_name={"$regex": "voronoi-features"},
@@ -83,12 +83,20 @@
 
     assert len(runs) == n_runs, f"found {len(runs)=} for {model}, expected {n_runs}"
 
-    run_time = sum(run.summary.get("_wandb", {}).get("runtime", 0) for run in runs)
+    each_run_time = [run.summary.get("_wandb", {}).get("runtime", 0) for run in runs]
+
+    run_time_total = sum(each_run_time)
     # NOTE we assume all jobs have the same metadata here
     metadata = requests.get(runs[0].file("wandb-metadata.json").url).json()
 
     n_gpu, n_cpu = metadata.get("gpu_count", 0), metadata.get("cpu_count", 0)
-    run_times[model] = {"Run time": run_time, "Hardware": f"GPU: {n_gpu}, CPU: {n_cpu}"}
+    run_times[model] = {
+        "Run time": run_time_total,
+        "Hardware": f"GPU: {n_gpu}, CPU: {n_cpu}",
+    }
+
+
+ax = (pd.Series(each_run_time) / 3600).hist(bins=100)
 
 
 # on 2022-11-28:

site/src/app.css

@@ -30,6 +30,7 @@ body > div {
 main {
   padding: calc(1ex + 2vw);
   flex: 1;
+  container-type: inline-size;
 }
 button {
   color: var(--text-color);

site/svelte.config.js

@@ -35,7 +35,7 @@ export default {
 
   preprocess: [
     {
-      // preprocess markdown citations @auth_1stwordtitle_yyyy into superscript
+      // preprocess markdown citations @auth_1st-word-title_yyyy into superscript
       // links to bibliography items, href must match References.svelte
       markup: (file) => {
         if (file.filename.endsWith(`paper/+page.svx`)) {