can be used from materials-toolkit

download-and-process.py

@@ -11,10 +11,17 @@ import multiprocessing as mp
 import json
 import re
 import tarfile
+import resource
+
+rlimit = resource.getrlimit(resource.RLIMIT_NOFILE)
+resource.setrlimit(resource.RLIMIT_NOFILE, (1 << 16, rlimit[1]))
 
 from ase.io import read
 import numpy as np
+import torch
 import h5py
+from materials_toolkit.data import HDF5Dataset, StructureData
+from materials_toolkit.data.datasets import MaterialsProjectData
 
 zip_file = "mp.2019.04.01.json.zip"
 url = "https://figshare.com/ndownloader/articles/8097992/versions/2"
@@ -97,108 +104,41 @@ def gen_structure_from_json(filename: str, chunksize: Optional[int] = 1 << 20):
     yield "{" + stack + "}"
 
 
-def parse_structure(json_str: str) -> tuple:
+def parse_structure(json_str: str) -> MaterialsProjectData:
     data = json.loads(json_str)
     struct = read(io.StringIO(data["structure"]), format="cif")
 
-    cell = struct.cell.array
-    natoms = len(struct)
-    x = struct.get_scaled_positions()
-    formula = struct.get_chemical_formula()
-    z = struct.get_atomic_numbers()
-
-    material_id = data["material_id"]
-    energy_pa = data["formation_energy_per_atom"]
-    return material_id, natoms, formula, cell, x, z, energy_pa
-
+    cell = torch.from_numpy(struct.cell.array).unsqueeze(0).float()
+    x = torch.from_numpy(struct.get_scaled_positions()).float()
+    z = torch.from_numpy(struct.get_atomic_numbers()).int()
+    material_id = torch.tensor(
+        [int(data["material_id"].split("-")[1])], dtype=torch.long
+    )
+    energy_pa = torch.tensor([data["formation_energy_per_atom"]], dtype=torch.float)
 
-def process_job(input_queue, output_queue):
-    while True:
-        job = input_queue.get()
+    return MaterialsProjectData(
+        pos=x, z=z, cell=cell, material_id=material_id, energy_pa=energy_pa
+    )
 
-        if job is None:
-            break
 
-        output_queue.put(parse_structure(job))
+def process(path: Optional[str] = "."):
+    mp_dir = os.path.join(path, "materials-project")
 
+    os.makedirs(mp_dir, exist_ok=True)
 
-def process(path: Optional[str] = ".", workers: Optional[int] = max(4, os.cpu_count())):
-    if os.path.exists(os.path.join(path, "index.json")) and os.path.exists(
-        os.path.join(path, "data.hdf5")
+    if os.path.exists(os.path.join(mp_dir, "batching.json")) and os.path.exists(
+        os.path.join(mp_dir, "data.hdf5")
     ):
         return
 
-    input_queue = mp.Queue()
-    output_queue = mp.Queue()
-    results = []
-
-    processes = []
-    for _ in range(workers):
-        p = mp.Process(target=process_job, args=(input_queue, output_queue))
-        p.start()
-        processes.append(p)
-
-    for elem in gen_structure_from_json(
-        os.path.join(path, "unzipped/mp.2019.04.01.json")
-    ):
-        input_queue.put(elem)
-
-        if input_queue.qsize() > 2 * workers:
-            results.append(output_queue.get())
-        while not output_queue.empty():
-            results.append(output_queue.get())
-
-    for _ in range(workers):
-        input_queue.put(None)
-
-    for process in processes:
-        process.join()
-
-    while not output_queue.empty():
-        results.append(output_queue.get())
-
-    material_id = [material_id for material_id, _, _, _, _, _, _ in results]
-    formula = [formula for _, _, formula, _, _, _, _ in results]
-
-    natoms = np.array([natoms for _, natoms, _, _, _, _, _ in results], dtype=np.int64)
-    atoms_ptr = np.pad(natoms.cumsum(0), (1, 0)).astype(np.int64)
-    idx = np.arange(len(results), dtype=np.int64)
-
-    cell = np.stack(
-        [cell for _, _, _, cell, _, _, _ in results], axis=0, dtype=np.float32
-    )
-    x = np.concatenate([x for _, _, _, _, x, _, _ in results], axis=0, dtype=np.float32)
-    z = np.concatenate([z for _, _, _, _, _, z, _ in results], axis=0, dtype=np.int64)
-    energy_pa = np.array(
-        [energy_pa for _, _, _, _, _, _, energy_pa in results], dtype=np.float32
-    )
-
-    index = [
-        {
-            "index": int(i),
-            "id": str(m_id),
-            "formula": str(f),
-            "natoms": int(n),
-            "energy_pa": float(e),
-        }
-        for i, m_id, f, n, e in zip(idx, material_id, formula, natoms, energy_pa)
+    results = [
+        parse_structure(elem)
+        for elem in gen_structure_from_json(
+            os.path.join(path, "unzipped/mp.2019.04.01.json")
+        )
     ]
-    with open(os.path.join(path, "index.json"), "w") as fp:
-        json.dump(index, fp)
-
-    f = h5py.File("data.hdf5", "w")
-
-    structures = f.create_group("structures")
-    structures.create_dataset("cell", data=cell, dtype=np.float32)
-    structures.create_dataset("natoms", data=natoms, dtype=np.int32)
-    structures.create_dataset("energy_pa", data=energy_pa, dtype=np.float32)
-    structures.create_dataset("atoms_ptr", data=atoms_ptr, dtype=np.int64)
-
-    atoms = f.create_group("atoms")
-    atoms.create_dataset("positions", data=x, dtype=np.float32)
-    atoms.create_dataset("atomic_number", data=z, dtype=np.uint8)
 
-    f.close()
+    HDF5Dataset.create_dataset(mp_dir, results)
 
 
 def compress(path: Optional[str] = "."):
@@ -209,8 +149,8 @@ def compress(path: Optional[str] = "."):
 
     print("compress into materials-project.tar.gz")
     with tarfile.open(output_file, "w:gz") as tar:
-        tar.add(os.path.join(path, "index.json"))
-        tar.add(os.path.join(path, "data.hdf5"))
+        tar.add(os.path.join(path, "materials-project/batching.json"))
+        tar.add(os.path.join(path, "materials-project/data.hdf5"))
 
 
 download_raw_mp()

materials-project.tar.gz

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:bcf27bb6a544f3cb28ab686ca98ba0977b8ef0d4445e14a63990c8cb91a4158b
-size 40789247
+oid sha256:d27d6d4570823fff5a2482ff1e33df223dcffe33a550a0facbd7a4878ad2d37e
+size 38793411