"""Simplify dataset (minimize the dataset size).
Init:
pick up init data from dataset randomly
Iter:
00: train models (same as generator)
01: calculate model deviations of the rest dataset, pick up data with proper model deviaiton
02: fp (optional, if the original dataset do not have fp data, same as generator)
"""
import glob
import logging
import os
import queue
from collections import defaultdict
from typing import Union
import dpdata
import numpy as np
from packaging.version import Version
from dpgen import dlog
from dpgen.dispatcher.Dispatcher import make_submission
# TODO: maybe the following functions can be moved to dpgen.util
from dpgen.generator.lib.utils import (
create_path,
log_iter,
make_iter_name,
record_iter,
)
from dpgen.generator.run import (
data_system_fmt,
fp_name,
fp_task_fmt,
make_fp_calculation,
make_train,
model_devi_name,
post_fp,
post_train,
run_fp,
run_train,
train_name,
train_task_fmt,
)
from dpgen.remote.decide_machine import convert_mdata
from dpgen.util import expand_sys_str, load_file, normalize, sepline, setup_ele_temp
from .arginfo import simplify_jdata_arginfo
picked_data_name = "data.picked"
rest_data_name = "data.rest"
accurate_data_name = "data.accurate"
detail_file_name_prefix = "details"
true_error_file_name = "true_error"
sys_name_fmt = "sys." + data_system_fmt
sys_name_pattern = "sys.[0-9]*[0-9]"
[docs]
def get_system_cls(jdata):
if jdata.get("labeled", False):
return dpdata.LabeledSystem
return dpdata.System
[docs]
def get_multi_system(path: Union[str, list[str]], jdata: dict) -> dpdata.MultiSystems:
"""Get MultiSystems from a path or list of paths.
Both NumPy and HDF5 formats are supported. For details
of two formats, refer to DeePMD-kit documentation.
If `labeled` in jdata is True, returns MultiSystems with LabeledSystem.
Otherwise, returns MultiSystems with System.
Parameters
----------
path : str or list of str
path or list of paths to the dataset
jdata : dict
parameters which may contain `labeled` key
Returns
-------
dpdata.MultiSystems
MultiSystems with LabeledSystem or System
"""
system = get_system_cls(jdata)
if not isinstance(path, (list, tuple)):
path = [path]
system_paths = []
for pp in path:
system_paths.extend(expand_sys_str(pp))
systems = dpdata.MultiSystems(
*[
system(s, fmt=("deepmd/npy" if "#" not in s else "deepmd/hdf5"))
for s in system_paths
],
type_map=jdata["type_map"],
)
return systems
[docs]
def init_model(iter_index, jdata, mdata):
training_init_model = jdata.get("training_init_model", False)
if not training_init_model:
return
iter0_models = []
training_iter0_model = jdata.get("training_iter0_model_path", [])
if isinstance(training_iter0_model, str):
training_iter0_model = [training_iter0_model]
for ii in training_iter0_model:
model_is = glob.glob(ii)
model_is.sort()
iter0_models += [os.path.abspath(ii) for ii in model_is]
numb_models = jdata["numb_models"]
assert numb_models == len(iter0_models), (
"training_iter0_model_path should be provided, and the number of models should be equal to %d"
% numb_models
)
work_path = os.path.join(make_iter_name(iter_index), train_name)
create_path(work_path)
cwd = os.getcwd()
for ii in range(len(iter0_models)):
train_path = os.path.join(work_path, train_task_fmt % ii)
create_path(train_path)
os.chdir(train_path)
ckpt_files = glob.glob(os.path.join(iter0_models[ii], "model.ckpt*"))
for jj in ckpt_files:
os.symlink(jj, os.path.basename(jj))
os.chdir(cwd)
[docs]
def init_pick(iter_index, jdata, mdata):
"""Pick up init data from dataset randomly."""
pick_data = jdata["pick_data"]
init_pick_number = jdata["init_pick_number"]
# use MultiSystems with System
# TODO: support System and LabeledSystem
# TODO: support other format
systems = get_multi_system(pick_data, jdata)
# label the system
labels = []
items = systems.systems.items()
for key, system in items:
labels.extend([(key, j) for j in range(len(system))])
# random pick
iter_name = make_iter_name(iter_index)
work_path = os.path.join(iter_name, model_devi_name)
create_path(work_path)
idx = np.arange(len(labels))
np.random.shuffle(idx)
pick_idx = idx[:init_pick_number]
rest_idx = idx[init_pick_number:]
# dump the init data
sys_data_path = os.path.join(work_path, picked_data_name)
_init_dump_selected_frames(systems, labels, pick_idx, sys_data_path, jdata)
# dump the rest data
sys_data_path = os.path.join(work_path, rest_data_name)
_init_dump_selected_frames(systems, labels, rest_idx, sys_data_path, jdata)
def _init_dump_selected_frames(systems, labels, selc_idx, sys_data_path, jdata):
selc_systems = dpdata.MultiSystems(type_map=jdata["type_map"])
sys_id_dict = defaultdict(list)
for j in selc_idx:
sys_name, sys_id = labels[j]
sys_id_dict[sys_name].append(sys_id)
for sys_name, sys_id in sys_id_dict.items():
# sys_id: list[int]
# System.append is slow; thus, we combine the idx of the same system
selc_systems.append(systems[sys_name][sys_id])
selc_systems.to_deepmd_raw(sys_data_path)
selc_systems.to_deepmd_npy(sys_data_path, set_size=selc_idx.size)
[docs]
def make_model_devi(iter_index, jdata, mdata):
"""Calculate the model deviation of the rest idx."""
pick_data = jdata["pick_data"]
iter_name = make_iter_name(iter_index)
work_path = os.path.join(iter_name, model_devi_name)
create_path(work_path)
# link the model
train_path = os.path.join(iter_name, train_name)
train_path = os.path.abspath(train_path)
models = glob.glob(os.path.join(train_path, "graph*pb"))
for mm in models:
model_name = os.path.basename(mm)
os.symlink(mm, os.path.join(work_path, model_name))
# link the last rest data
last_iter_name = make_iter_name(iter_index - 1)
rest_data_path = os.path.join(last_iter_name, model_devi_name, rest_data_name)
if not os.path.exists(rest_data_path):
return False
os.symlink(
os.path.abspath(rest_data_path),
os.path.join(work_path, rest_data_name + ".old"),
)
return True
[docs]
def run_model_devi(iter_index, jdata, mdata):
"""Submit dp test tasks."""
iter_name = make_iter_name(iter_index)
work_path = os.path.join(iter_name, model_devi_name)
# generate command
commands = []
run_tasks = ["."]
# get models
models = glob.glob(os.path.join(work_path, "graph*pb"))
model_names = [os.path.basename(ii) for ii in models]
task_model_list = []
for ii in model_names:
task_model_list.append(os.path.join(".", ii))
# models
commands = []
detail_file_name = detail_file_name_prefix
system_file_name = rest_data_name + ".old"
if jdata.get("one_h5", False):
# convert system to one h5 file
system_path = os.path.join(work_path, system_file_name)
dpdata.MultiSystems(type_map=jdata["type_map"]).from_deepmd_npy(
system_path,
labeled=False,
).to_deepmd_hdf5(system_path + ".hdf5")
system_file_name += ".hdf5"
command = "rm -f {detail_file} && {dp} model-devi -m {model} -s {system} -o {detail_file}".format(
dp=mdata.get("model_devi_command", "dp"),
model=" ".join(task_model_list),
system=system_file_name,
detail_file=detail_file_name,
)
commands = [command]
# submit
model_devi_group_size = mdata.get("model_devi_group_size", 1)
forward_files = [system_file_name]
backward_files = [detail_file_name]
f_trust_lo_err = jdata.get("true_error_f_trust_lo", float("inf"))
e_trust_lo_err = jdata.get("true_error_e_trust_lo", float("inf"))
if f_trust_lo_err < float("inf") or e_trust_lo_err < float("inf"):
command_true_error = "rm -f {detail_file} && {dp} model-devi -m {model} -s {system} -o {detail_file} --real_error".format(
dp=mdata.get("model_devi_command", "dp"),
model=" ".join(task_model_list),
system=system_file_name,
detail_file=true_error_file_name,
)
commands.append(command_true_error)
backward_files.append(true_error_file_name)
api_version = mdata.get("api_version", "1.0")
if Version(api_version) < Version("1.0"):
raise RuntimeError(
"API version %s has been removed. Please upgrade to 1.0." % api_version
)
elif Version(api_version) >= Version("1.0"):
submission = make_submission(
mdata["model_devi_machine"],
mdata["model_devi_resources"],
commands=commands,
work_path=work_path,
run_tasks=run_tasks,
group_size=model_devi_group_size,
forward_common_files=model_names,
forward_files=forward_files,
backward_files=backward_files,
outlog="model_devi.log",
errlog="model_devi.log",
)
submission.run_submission()
[docs]
def post_model_devi(iter_index, jdata, mdata):
"""Calculate the model deviation."""
iter_name = make_iter_name(iter_index)
work_path = os.path.join(iter_name, model_devi_name)
f_trust_lo = jdata["model_devi_f_trust_lo"]
f_trust_hi = jdata["model_devi_f_trust_hi"]
e_trust_lo = jdata["model_devi_e_trust_lo"]
e_trust_hi = jdata["model_devi_e_trust_hi"]
f_trust_lo_err = jdata.get("true_error_f_trust_lo", float("inf"))
f_trust_hi_err = jdata.get("true_error_f_trust_hi", float("inf"))
e_trust_lo_err = jdata.get("true_error_e_trust_lo", float("inf"))
e_trust_hi_err = jdata.get("true_error_e_trust_hi", float("inf"))
use_true_error = f_trust_lo_err < float("inf") or e_trust_lo_err < float("inf")
type_map = jdata.get("type_map", [])
sys_accurate = dpdata.MultiSystems(type_map=type_map)
sys_candinate = dpdata.MultiSystems(type_map=type_map)
sys_failed = dpdata.MultiSystems(type_map=type_map)
labeled = jdata.get("labeled", False)
sys_entire = dpdata.MultiSystems(type_map=type_map).from_deepmd_npy(
os.path.join(work_path, rest_data_name + ".old"), labeled=labeled
)
detail_file_name = detail_file_name_prefix
if not use_true_error:
with open(os.path.join(work_path, detail_file_name)) as f:
for line in f:
if line.startswith("# data.rest.old"):
name = (line.split()[1]).split("/")[-1]
elif line.startswith("#"):
columns = line.split()[1:]
cidx_step = columns.index("step")
cidx_max_devi_f = columns.index("max_devi_f")
try:
cidx_devi_e = columns.index("devi_e")
except ValueError:
# DeePMD-kit < 2.2.2
cidx_devi_e = None
else:
idx = int(line.split()[cidx_step])
f_devi = float(line.split()[cidx_max_devi_f])
if cidx_devi_e is not None:
e_devi = float(line.split()[cidx_devi_e])
else:
e_devi = 0.0
subsys = sys_entire[name][idx]
if f_devi >= f_trust_hi or e_devi >= e_trust_hi:
sys_failed.append(subsys)
elif (
f_trust_lo <= f_devi < f_trust_hi
or e_trust_lo <= e_devi < e_trust_hi
):
sys_candinate.append(subsys)
elif f_devi < f_trust_lo and e_devi < e_trust_lo:
sys_accurate.append(subsys)
else:
raise RuntimeError(
"reach a place that should NOT be reached..."
)
else:
with (
open(os.path.join(work_path, detail_file_name)) as f,
open(os.path.join(work_path, true_error_file_name)) as f_err,
):
for line, line_err in zip(f, f_err):
if line.startswith("# data.rest.old"):
name = (line.split()[1]).split("/")[-1]
elif line.startswith("#"):
columns = line.split()[1:]
cidx_step = columns.index("step")
cidx_max_devi_f = columns.index("max_devi_f")
cidx_devi_e = columns.index("devi_e")
else:
idx = int(line.split()[cidx_step])
f_devi = float(line.split()[cidx_max_devi_f])
f_err = float(line_err.split()[cidx_max_devi_f])
e_devi = float(line.split()[cidx_devi_e])
e_err = float(line_err.split()[cidx_devi_e])
subsys = sys_entire[name][idx]
if (
f_devi >= f_trust_hi
or e_devi >= e_trust_hi
or f_err >= f_trust_hi_err
or e_err >= e_trust_hi_err
):
sys_failed.append(subsys)
elif (
f_trust_lo <= f_devi < f_trust_hi
or e_trust_lo <= e_devi < e_trust_hi
or f_trust_lo_err <= f_err < f_trust_hi_err
or e_trust_lo_err <= e_err < e_trust_hi_err
):
sys_candinate.append(subsys)
elif (
f_devi < f_trust_lo
and e_devi < e_trust_lo
and f_err < f_trust_lo_err
and e_err < e_trust_lo_err
):
sys_accurate.append(subsys)
else:
raise RuntimeError(
"reach a place that should NOT be reached..."
)
counter = {
"candidate": sys_candinate.get_nframes(),
"accurate": sys_accurate.get_nframes(),
"failed": sys_failed.get_nframes(),
}
fp_sum = sum(counter.values())
for cc_key, cc_value in counter.items():
dlog.info(
f"{cc_key:9s} : {cc_value:6d} in {fp_sum:6d} {cc_value / fp_sum * 100:6.2f} %"
)
if counter["candidate"] == 0 and counter["failed"] > 0:
raise RuntimeError(
"no candidate but still have failed cases, stop. You may want to refine the training or to increase the trust level hi"
)
# label the candidate system
labels = []
items = sys_candinate.systems.items()
for key, system in items:
labels.extend([(key, j) for j in range(len(system))])
# candinate: pick up randomly
iter_pick_number = jdata["iter_pick_number"]
idx = np.arange(counter["candidate"])
assert len(idx) == len(labels)
np.random.shuffle(idx)
pick_idx = idx[:iter_pick_number]
rest_idx = idx[iter_pick_number:]
if counter["candidate"] == 0:
dlog.info("no candidate")
else:
dlog.info(
"total candidate {:6d} picked {:6d} ({:6.2f} %) rest {:6d} ({:6.2f} % )".format(
counter["candidate"],
len(pick_idx),
float(len(pick_idx)) / counter["candidate"] * 100.0,
len(rest_idx),
float(len(rest_idx)) / counter["candidate"] * 100.0,
)
)
# dump the picked candinate data
picked_systems = dpdata.MultiSystems(type_map=type_map)
for j in pick_idx:
sys_name, sys_id = labels[j]
picked_systems.append(sys_candinate[sys_name][sys_id])
sys_data_path = os.path.join(work_path, picked_data_name)
picked_systems.to_deepmd_raw(sys_data_path)
picked_systems.to_deepmd_npy(sys_data_path, set_size=iter_pick_number)
# dump the rest data (not picked candinate data and failed data)
rest_systems = dpdata.MultiSystems(type_map=type_map)
for j in rest_idx:
sys_name, sys_id = labels[j]
rest_systems.append(sys_candinate[sys_name][sys_id])
rest_systems += sys_failed
sys_data_path = os.path.join(work_path, rest_data_name)
rest_systems.to_deepmd_raw(sys_data_path)
rest_systems.to_deepmd_npy(sys_data_path, set_size=rest_systems.get_nframes())
# dump the accurate data -- to another directory
sys_data_path = os.path.join(work_path, accurate_data_name)
sys_accurate.to_deepmd_raw(sys_data_path)
sys_accurate.to_deepmd_npy(sys_data_path, set_size=sys_accurate.get_nframes())
[docs]
def make_fp_labeled(iter_index, jdata):
dlog.info("already labeled, skip make_fp and link data directly")
pick_data = jdata["pick_data"]
iter_name = make_iter_name(iter_index)
work_path = os.path.join(iter_name, fp_name)
create_path(work_path)
picked_data_path = os.path.join(iter_name, model_devi_name, picked_data_name)
if os.path.exists(os.path.abspath(picked_data_path)):
os.symlink(
os.path.abspath(picked_data_path),
os.path.abspath(os.path.join(work_path, "task." + fp_task_fmt % (0, 0))),
)
os.symlink(
os.path.abspath(picked_data_path),
os.path.abspath(os.path.join(work_path, "data." + data_system_fmt % 0)),
)
[docs]
def make_fp_configs(iter_index, jdata):
pick_data = jdata["pick_data"]
iter_name = make_iter_name(iter_index)
work_path = os.path.join(iter_name, fp_name)
create_path(work_path)
picked_data_path = os.path.join(iter_name, model_devi_name, picked_data_name)
if not os.path.exists(os.path.abspath(picked_data_path)):
return
systems = get_multi_system(picked_data_path, jdata)
ii = 0
jj = 0
for system in systems:
for subsys in system:
task_name = "task." + fp_task_fmt % (ii, jj)
task_path = os.path.join(work_path, task_name)
create_path(task_path)
subsys.to("vasp/poscar", os.path.join(task_path, "POSCAR"))
jj += 1
ii += 1
[docs]
def make_fp(iter_index, jdata, mdata):
labeled = jdata.get("labeled", False)
if labeled:
make_fp_labeled(iter_index, jdata)
else:
make_fp_configs(iter_index, jdata)
jdata["model_devi_engine"] = "lammps"
make_fp_calculation(iter_index, jdata, mdata)
[docs]
def run_iter(param_file, machine_file):
"""Init (iter 0): init_pick.
tasks (iter > 0):
00 make_train (same as generator)
01 run_train (same as generator)
02 post_train (same as generator)
03 make_model_devi
04 run_model_devi
05 post_model_devi
06 make_fp
07 run_fp (same as generator)
08 post_fp (same as generator)
"""
jdata = load_file(param_file)
mdata = load_file(machine_file)
jdata_arginfo = simplify_jdata_arginfo()
jdata = normalize(jdata_arginfo, jdata)
# set up electron temperature
use_ele_temp = jdata.get("use_ele_temp", 0)
if use_ele_temp == 1:
setup_ele_temp(False)
elif use_ele_temp == 2:
setup_ele_temp(True)
if mdata.get("handlers", None):
if mdata["handlers"].get("smtp", None):
que = queue.Queue(-1)
queue_handler = logging.handlers.QueueHandler(que)
smtp_handler = logging.handlers.SMTPHandler(**mdata["handlers"]["smtp"])
listener = logging.handlers.QueueListener(que, smtp_handler)
dlog.addHandler(queue_handler)
listener.start()
mdata = convert_mdata(mdata)
max_tasks = 10000
numb_task = 9
record = "record.dpgen"
iter_rec = [0, -1]
if os.path.isfile(record):
with open(record) as frec:
for line in frec:
iter_rec = [int(x) for x in line.split()]
dlog.info("continue from iter %03d task %02d" % (iter_rec[0], iter_rec[1]))
cont = True
ii = -1
while cont:
ii += 1
if ii < iter_rec[0]:
continue
iter_name = make_iter_name(ii)
sepline(iter_name, "=")
for jj in range(numb_task):
if ii * max_tasks + jj <= iter_rec[0] * max_tasks + iter_rec[1]:
continue
task_name = "task %02d" % jj
sepline(f"{iter_name} {task_name}", "-")
jdata["model_devi_jobs"] = [{} for _ in range(ii + 1)]
if ii == 0 and jj < 6 and (jj >= 3 or not jdata.get("init_data_sys", [])):
if jj == 0:
log_iter("init_train", ii, jj)
init_model(ii, jdata, mdata)
elif jj == 3:
log_iter("init_pick", ii, jj)
init_pick(ii, jdata, mdata)
else:
dlog.info("first iter, skip step 1-5")
elif jj == 0:
log_iter("make_train", ii, jj)
make_train(ii, jdata, mdata)
elif jj == 1:
log_iter("run_train", ii, jj)
# disp = make_dispatcher(mdata['train_machine'])
run_train(ii, jdata, mdata)
elif jj == 2:
log_iter("post_train", ii, jj)
post_train(ii, jdata, mdata)
elif jj == 3:
log_iter("make_model_devi", ii, jj)
cont = make_model_devi(ii, jdata, mdata)
if not cont or ii >= jdata.get("stop_iter", ii + 1):
break
elif jj == 4:
log_iter("run_model_devi", ii, jj)
# disp = make_dispatcher(mdata['model_devi_machine'])
run_model_devi(ii, jdata, mdata)
elif jj == 5:
log_iter("post_model_devi", ii, jj)
post_model_devi(ii, jdata, mdata)
elif jj == 6:
log_iter("make_fp", ii, jj)
make_fp(ii, jdata, mdata)
elif jj == 7:
log_iter("run_fp", ii, jj)
if jdata.get("labeled", False):
dlog.info("already have labeled data, skip run_fp")
else:
# disp = make_dispatcher(mdata['fp_machine'])
run_fp(ii, jdata, mdata)
elif jj == 8:
log_iter("post_fp", ii, jj)
if jdata.get("labeled", False):
dlog.info("already have labeled data, skip post_fp")
else:
post_fp(ii, jdata)
else:
raise RuntimeError("unknown task %d, something wrong" % jj)
record_iter(record, ii, jj)
[docs]
def gen_simplify(args):
if args.PARAM and args.MACHINE:
if args.debug:
dlog.setLevel(logging.DEBUG)
dlog.info("start simplifying")
run_iter(args.PARAM, args.MACHINE)
dlog.info("finished")