from asaplib.data.xyz import ASAPXYZ
from ai2_kit.core.artifact import Artifact, ArtifactDict
from ai2_kit.core.log import get_logger
from ai2_kit.core.util import dump_json, flush_stdio, limit
from ai2_kit.core.pydantic import BaseModel
from typing import List, Optional, Tuple, Dict
from io import StringIO
from dataclasses import dataclass
import pandas as pd
from tabulate import tabulate
from itertools import groupby
from functools import lru_cache
import ase.io
import os
from .data import get_data_format, DataFormat, artifacts_to_ase_atoms
from .iface import ICllSelectorOutput, BaseCllContext
from .constant import LAMMPS_DUMP_DIR, LAMMPS_DUMP_SUFFIX, DEFAULT_ASAP_SOAP_DESC, DEFAULT_ASAP_PCA_REDUCER
from .asap import get_descriptor, reduce_dimension, get_trainer, get_cluster
logger = get_logger(__name__)
[docs]@dataclass
class CllModelDevSelectorContext(BaseCllContext):
...
[docs]@dataclass
class CllModelDeviSelectorOutput(ICllSelectorOutput):
candidates: List[Artifact]
passing_rate: float
new_explore_systems: List[Artifact]
[docs] def get_model_devi_dataset(self):
return self.candidates
[docs] def get_passing_rate(self) -> float:
return self.passing_rate
[docs] def get_new_explore_systems(self) -> List[Artifact]:
return self.new_explore_systems
[docs]async def cll_model_devi_selector(input: CllModelDeviSelectorInput, ctx: CllModelDevSelectorContext):
executor = ctx.resource_manager.default_executor
work_dir = os.path.join(executor.work_dir, ctx.path_prefix)
executor.mkdir(work_dir)
f_trust_lo = input.config.f_trust_lo
f_trust_hi = input.config.f_trust_hi
results = executor.run_python_fn(bulk_select_structures_by_model_devi)(
model_devi_outputs=[a.to_dict() for a in input.model_devi_data],
model_devi_file=input.model_devi_file,
f_trust_lo=f_trust_lo, f_trust_hi=f_trust_hi,
type_map=input.type_map, work_dir=work_dir,
new_explore_system_q=input.config.new_explore_system_q,
max_decent_per_traj=input.config.max_decent_per_traj,
screening_fn=input.config.screening_fn,
)
candidates = [ result['decent'] for result, _ in results if 'decent' in result ]
new_systems = [ result['next'] for result, _ in results if 'next' in result ]
stats = [ stats for _, stats in results ]
# group next_structures by `attrs.source` and keep the first one
# so that the total number of explore structures will be the same as the original one
new_systems = [next(group) for _source, group in groupby(
new_systems, key=lambda s: s['attrs']['source'])]
# write model_deviation stats report
# TODO: refactor into a function
def _get_row(stat: dict):
url = stat['src']
url = f'...{url[-30:]}' if len(url) > 30 else url # wrap url if it is too long
total, good, decent, poor = stat['total'], stat['good'], stat['decent'], stat['poor']
return [
url, total, good, decent, poor,
f'{good / total * 100:.2f}%', f'{decent / total * 100:.2f}%', f'{poor / total * 100:.2f}%',
]
headers = ['file', 'total', 'good', 'decent', 'poor', 'good%', 'decent%', 'poor%']
table = [ _get_row(stat) for stat in stats]
total = sum(row[1] for row in table)
total_good = sum(row[2] for row in table)
stats_report = tabulate(table, headers=headers, tablefmt='tsv')
logger.info('stats report: \n%s\n', stats_report)
executor.dump_text(stats_report, os.path.join(work_dir, 'stats.tsv'))
# further select candidates by ASAP
if input.config.asap_options and not input.config.asap_options.disable:
asap_options = input.config.asap_options
candidates = executor.run_python_fn(bulk_select_distinct_structures)(
candidates=candidates,
descriptor_opt=asap_options.descriptor,
dim_reducer_opt=asap_options.dim_reducer,
cluster_opt=asap_options.cluster,
type_map=input.type_map,
work_dir=work_dir,
limit_per_cluster=asap_options.limit_per_cluster,
sort_by_energy=asap_options.sort_by_ssw_energy,
)
return CllModelDeviSelectorOutput(
candidates=[Artifact.of(**a) for a in candidates],
new_explore_systems=[Artifact.of(**a) for a in new_systems],
passing_rate=total_good / total,
)
def __export_remote_functions():
def bulk_select_structures_by_model_devi(model_devi_outputs: List[ArtifactDict],
model_devi_file: str,
f_trust_lo: float,
f_trust_hi: float,
new_explore_system_q: float,
type_map: List[str],
work_dir: str,
max_decent_per_traj: int,
screening_fn: Optional[str],
workers: int = 4,
) -> List[Tuple[Dict[str, ArtifactDict], dict]]:
import joblib
return joblib.Parallel(n_jobs=workers)(
joblib.delayed(select_structures_by_model_devi)(
model_devi_output=output,
model_devi_file=model_devi_file,
f_trust_lo=f_trust_lo, f_trust_hi=f_trust_hi,
type_map=type_map,
work_dir=os.path.join(work_dir, 'model_devi', f'{i:06}'),
max_decent_per_traj=max_decent_per_traj,
new_explore_system_q=new_explore_system_q,
screening_fn=screening_fn,
)
for i, output in enumerate(model_devi_outputs)
) # type: ignore
def select_structures_by_model_devi(model_devi_output: ArtifactDict,
model_devi_file: str,
f_trust_lo: float,
f_trust_hi: float,
type_map: List[str],
work_dir: str,
new_explore_system_q: float,
max_decent_per_traj: int,
screening_fn: Optional[str],
) -> Tuple[Dict[str, ArtifactDict], dict]:
"""
analysis the model_devi output of explore stage and select candidates
:param next_explore_system_q: the quantile of model_devi score to select the structure for next round of exploration
"""
os.makedirs(work_dir, exist_ok=True)
dump_json(model_devi_output, os.path.join(work_dir, 'model_devi_output.debug.json'))
model_devi_dir = model_devi_output['url']
model_devi_file = model_devi_output['attrs'].pop('model_devi_file', model_devi_file)
force_col = 'max_devi_f'
print(f'criteria: {f_trust_lo} <= {force_col} < {f_trust_hi}')
# get path of model_devi file
data_format = get_data_format(model_devi_output) # type: ignore
if data_format in (DataFormat.LAMMPS_OUTPUT_DIR, DataFormat.LASP_LAMMPS_OUT_DIR):
model_devi_file = os.path.join(model_devi_dir, model_devi_file)
else:
raise ValueError('unknown model_devi_data types')
logger.info('start to analysis file: %s', model_devi_file)
# load model_devi data
with open(model_devi_file, 'r') as f:
text = f.read()
df = pd.read_csv(StringIO(text.lstrip('#')), delim_whitespace=True)
# layout:
# step max_devi_v min_devi_v avg_devi_v max_devi_f min_devi_f avg_devi_f
# 0 0 0.006793 0.000672 0.003490 0.143317 0.005612 0.026106
# 1 100 0.006987 0.000550 0.003952 0.128178 0.006042 0.022608
# load structures
atoms_list = []
if data_format == DataFormat.LAMMPS_OUTPUT_DIR:
lammps_dump_dir = model_devi_output['attrs'].pop('lammps_dump_dir', LAMMPS_DUMP_DIR)
for frame_id in df.step:
dump_file = os.path.join(model_devi_dir, lammps_dump_dir, f'{frame_id}{LAMMPS_DUMP_SUFFIX}')
atoms_list += ase.io.read(dump_file, ':', format='lammps-dump-text', specorder=type_map)
elif data_format == DataFormat.LASP_LAMMPS_OUT_DIR:
structures_file = os.path.join(model_devi_dir, 'structures.xyz')
atoms_list += ase.io.read(structures_file, ':', format='extxyz')
else:
raise ValueError('unknown model_devi_data types')
# screening structure before model_devi analysis
if screening_fn is not None:
if 'ssw_energy' in atoms_list[0].info:
s_ssw_energy = pd.Series(map(lambda atoms: atoms.info['ssw_energy'], atoms_list)) # type: ignore
# the following ssw_* methods are for the screening_fn
# so don't need to worry about the unused warning
ssw_energy_max = s_ssw_energy.max()
ssw_energy_min = s_ssw_energy.min()
def ssw_energy_quantile(q):
# the quantile will be evaluated every time, which is not efficient
# so here we cache the result, carefully
return lru_cache()(s_ssw_energy.quantile)(q)
ssw_enenrgy_quantile = ssw_energy_quantile
# return the df row whose atoms pass the screening_fn
_screening_fn = eval(screening_fn, locals()) # str to function
df = df[[ _screening_fn(atoms) for atoms in atoms_list ]]
# evaluate new found structures by their model_devi score in 3 levels: good, decent, poor
good_df = df[df[force_col] < f_trust_lo]
decent_df = df[(df[force_col] >= f_trust_lo) & (df[force_col] < f_trust_hi)]
poor_df = df[df[force_col] >= f_trust_hi]
# select the last frame from df whose model_devi score is less than the quantile
# as the initial structure for next round of exploration to replace the original one
# NOTE: equal is essential to ensure the existence of next structure
# NOTE: select the last frame can increase the diversity of structures
next_df = df[df[force_col] <= df[force_col].quantile(new_explore_system_q)].tail(1)
stats = {
'src': model_devi_file,
'total': len(df),
'good': len(good_df),
'decent': len(decent_df),
'poor': len(poor_df),
}
result: Dict[str, ArtifactDict] = {}
# TODO: refactor the following repeating code
# TODO: dump good may lead to storage issue, so disable it for now
if len(good_df) > 0:
good_file = os.path.join(work_dir, 'good.xyz')
# ase.io.write(good_file, [atoms_list[_i] for _i in good_df.index], format='extxyz')
result['good'] = {'url': good_file, 'format': DataFormat.EXTXYZ, # type: ignore
'attrs': {**model_devi_output['attrs']}}
if len(poor_df) > 0:
poor_file = os.path.join(work_dir, 'poor.xyz')
# ase.io.write(poor_file, [atoms_list[_i] for _i in poor_df.index], format='extxyz')
result['poor'] = {'url': poor_file, 'format': DataFormat.EXTXYZ, # type: ignore
'attrs': {**model_devi_output['attrs']}}
if len(decent_df) > 0:
decent_file = os.path.join(work_dir, 'decent.xyz')
ase.io.write(decent_file,
list(limit((atoms_list[_i] for _i in decent_df.index), max_decent_per_traj)),
format='extxyz')
result['decent'] = {'url': decent_file, 'format': DataFormat.EXTXYZ, # type: ignore
'attrs': {**model_devi_output['attrs']}}
if len(next_df) > 0:
next_file = os.path.join(work_dir, 'next.xyz')
ase.io.write(next_file, [atoms_list[_i] for _i in next_df.index], format='extxyz')
result['next'] = {'url': next_file, 'format': DataFormat.EXTXYZ, # type: ignore
'attrs': {**model_devi_output['attrs']}}
dump_json([result, stats, list(decent_df.index), list(next_df.index)], os.path.join(work_dir, 'result.debug.json'))
return result, stats
def bulk_select_distinct_structures(candidates: List[ArtifactDict],
descriptor_opt: dict,
dim_reducer_opt: dict,
cluster_opt: dict,
type_map: List[str],
work_dir: str,
limit_per_cluster: int = -1,
sort_by_energy: bool = False,
workers: int = 4,
) -> List[ArtifactDict]:
inputs = []
for i, (ancestor_key, candidate_group) in enumerate(groupby(candidates, key=lambda c: c['attrs']['ancestor'])):
candidate_group = list(candidate_group)
inputs.append((candidate_group, candidate_group[0]['attrs']))
import joblib
return joblib.Parallel(n_jobs=workers)(
joblib.delayed(select_distinct_structures)(
candidates=group,
attrs=attrs,
descriptor_opt=descriptor_opt,
dim_reducer_opt=dim_reducer_opt,
cluster_opt=cluster_opt,
type_map=type_map,
work_dir=os.path.join(work_dir, 'asap', f'{i:06}'),
limit_per_cluster=limit_per_cluster,
sort_by_energy=sort_by_energy,
) for i, (group, attrs) in enumerate(inputs)
) # type: ignore
def select_distinct_structures(candidates: List[ArtifactDict],
attrs: dict,
descriptor_opt: dict,
dim_reducer_opt: dict,
cluster_opt: dict,
type_map: List[str],
work_dir: str,
limit_per_cluster: int = -1,
sort_by_energy: bool = False,
):
os.makedirs(work_dir, exist_ok=True)
dump_json(attrs, os.path.join(work_dir, 'attrs.debug.json'))
# load structures and save it to a tmp file for ASAP to load
atoms_list = [atoms for _, atoms in artifacts_to_ase_atoms(candidates, type_map=type_map)]
if len(atoms_list) < 10:
# FIXME: there are a lot of potential issue when the number of atoms is small
# the root cause is in asaplib, which I guess has not been tested with small dataset
selected_atoms_list = atoms_list
else:
if sort_by_energy and 'ssw_energy' in atoms_list[0].info:
atoms_list = sorted(atoms_list, key=lambda atoms: atoms.info['ssw_energy'])
# use asaplib to group structures
# load structures to ASAP
tmp_structures_file = os.path.join(work_dir, '.tmp-structures.xyz')
ase.io.write(tmp_structures_file, atoms_list, format='extxyz')
asapxyz = ASAPXYZ(tmp_structures_file)
# group structures
asap_path_prefix = os.path.join(work_dir, 'asap')
descriptors, _ = get_descriptor(asapxyz, descriptor_opt, path_prefix=asap_path_prefix)
reduced_descriptors = reduce_dimension(descriptors, dim_reducer_opt)
trainer = get_trainer(reduced_descriptors, cluster_opt)
cluster_labels = get_cluster(asapxyz, reduced_descriptors, trainer, path_prefix=asap_path_prefix)
# dump_json(cluster_labels, os.path.join(work_dir, 'cluster.debug.json'))
# select frame from each cluster
if limit_per_cluster <= 0: # unlimit
selected_atoms_list = atoms_list # do nothing
else:
selected_frames = []
for frames in cluster_labels.values():
if len(frames) < limit_per_cluster:
selected_frames += list(frames)
else:
selected_frames += list(frames[:limit_per_cluster])
selected_atoms_list = [atoms_list[i] for i in selected_frames]
# write selected structures to file
distinct_structures_file = os.path.join(work_dir, 'distinct_structures.xyz')
ase.io.write(distinct_structures_file, selected_atoms_list, format='extxyz')
output = {
'url': distinct_structures_file,
'format': DataFormat.EXTXYZ,
'attrs': attrs,
}
dump_json(output, os.path.join(work_dir, 'output.debug.json'))
flush_stdio() # flush joblib stdio buffer
return output
return (
bulk_select_structures_by_model_devi,
bulk_select_distinct_structures,
)
(
bulk_select_structures_by_model_devi,
bulk_select_distinct_structures,
) = __export_remote_functions()