def run_test_dataset(self, dataset: ZDataset):
test_recorder = StatRecorder(timing=True)
for ibatch in dataset.yield_batches(loop=False):
with test_recorder.go():
one_res = self.model.predict_on_batch(ibatch)
test_recorder.record(one_res)
# --
# write output
if self.is_main_process: # note: do saving only with main process
dataset.write_insts(None) # use dataset's conf
# --
# eval
x = test_recorder.summary()
zlog(f"Test-Info: {OtherHelper.printd_str(x, sep=' ')}")
aggr = ResultAggregator()
for task in self.t_center.tasks.values():
if task.name not in dataset.tasks:
continue
tn_res: ResultRecord = task.eval_insts(dataset.gold_insts,
dataset.insts,
quite=False)
if tn_res is None:
continue
aggr.add(task.name, tn_res, task.conf.eval_weight)
ret = aggr.get_res()
return ret
class TrainingRunner:
def __init__(self, conf: TRConf, model: ZModel, train_stream: Streamer,
train_batch_f: Callable, train_discard_batch_f: Callable,
dev_runners: List[TestingRunner], **kwargs):
self.conf = conf
self.kwargs = kwargs
# --
self.model = model
self.train_stream = train_stream
self.train_batch_f = train_batch_f if train_batch_f is not None else lambda x: 1 # inst -> int
self.train_discard_batch_f = train_discard_batch_f if train_discard_batch_f is not None else lambda x: False # List[inst] -> bool
self.dev_runners = dev_runners
# some records
self.tp = TrainingProgressRecord()
self.train_recorder = StatRecorder(timing=True)
# -- special!!
# store all insts for future usage (do not use if input is large)
self.stored_all_insts = None
if conf.store_all_insts:
_all_insts = []
for _one_insts in self.train_stream:
_all_insts.extend(_one_insts)
self.stored_all_insts = _all_insts
# --
# scheduled values
self.lrate = ScheduledValue("lrate", conf.lrate)
self.lrate_warmup_steps = self._determine_warmup_steps(
conf.lrate_warmup_eidx, conf.lrate_warmup_uidx, train_stream)
if self.lrate_warmup_steps > 0:
self.lrate_warmup_factor = 1. / (self.lrate_warmup_steps**
conf.lrate_decrease_alpha)
zlog(
f"For lrate-warmup, first {self.lrate_warmup_steps} steps up to {self.lrate.value}, "
f"then decrease with lrate*{self.lrate_warmup_factor}*step^{conf.lrate_decrease_alpha}",
func="plain")
else:
self.lrate_warmup_factor = 1.
self.scheduled_values = OrderedDict([("_lrate", self.lrate)])
DictHelper.update_dict(self.scheduled_values,
model.get_scheduled_values()
) # values to be scheduled as training goes
def _determine_warmup_steps(self, eidx: int, uidx: int, stream: Streamer):
# set warmup steps if using eidx
if eidx > 0:
_epoch_step_size = len(list(stream))
_warmup_steps = _epoch_step_size * eidx
zlog(
f"Determine warmup steps: {eidx} x {_epoch_step_size} = {_warmup_steps}",
func="plain")
else:
_warmup_steps = 0
_warmup_steps = max(_warmup_steps, uidx)
return _warmup_steps
def current_name(self):
return self.tp.current_suffix()
def add_scheduled_value(self, v: ScheduledValue, key_prefix=''):
key = key_prefix + v.name
assert key not in self.scheduled_values
self.scheduled_values[key] = v
def adjust_scheduled_values(self):
# adjust schedule values
ss = self.current_name()
for one_name, one_sv in self.scheduled_values.items():
if one_sv.changeable:
one_sv.adjust_at_ckp(ss, self.tp, extra_info=one_name)
# -----
# saving and loading related
def save_progress(self, file: str):
default_json_serializer.to_file(self.tp, file)
zlog(f"Save training progress to {file}", func="io")
def load_progress(self, file: str, forward_stream=False):
old_uidx = self.tp.uidx
d = default_json_serializer.from_file(file)
self.tp.from_json(d)
if forward_stream:
if old_uidx > self.tp.uidx:
zwarn(
f"Cannot go to the past: {old_uidx} -> {self.tp.uidx}, skip this!"
)
else:
_s = self.train_stream
for _ in range(self.tp.uidx - old_uidx):
_, _eos = _s.next_and_check()
if _eos: # restart and get one
_s.restart()
_s.next()
zlog(f"Forward to the future: {old_uidx} -> {self.tp.uidx}!",
func="io")
zlog(f"Load training progress from {file}", func="io")
self.adjust_scheduled_values() # also adjust values!
def save(self, prefix: str):
self.save_progress(prefix + ".tp.json")
self.model.save(prefix + ".m")
def load(self,
prefix: str,
load_progress=False,
forward_stream=False,
load_strict=None):
if prefix.endswith(".m"):
prefix = prefix[:-2]
if load_progress:
self.load_progress(prefix + ".tp.json", forward_stream)
self.model.load(prefix + ".m", strict=load_strict)
def get_train_stream(self):
return self.train_stream
# =====
# run until the end of training
def run(self):
conf = self.conf
last_report_uidx, last_dev_uidx = 0, 0
# --
if conf.valid_first: # valid before training
self.validate()
# --
_lrate_warmup_factor, _lrate_warmup_steps = self.lrate_warmup_factor, self.lrate_warmup_steps
_skip_batch = conf.skip_batch
_gen0 = Random.get_generator("train")
_gen = Random.stream(_gen0.random_sample)
# --
_accu_checker = 0
_accu_batch = conf.accu_batch
# --
# start before loop
self.adjust_scheduled_values()
# loop
act_lrate = None
while True: # loop over and over
_train_stream = self.get_train_stream(
) # we may change train_stream!!
# --
if _train_stream.is_inactive(
): # check to avoid restart after load_progress
_train_stream.restart()
insts, _eos = _train_stream.next_and_check()
if _eos: # end of epoch
zlog(
f"End of epoch at {self.tp.current_suffix(False)}: Current act_lrate is {act_lrate}.",
func="plain",
timed=True)
if conf.valid_epoch:
last_dev_uidx = self.tp.uidx
self.validate()
# todo(+N): do we need to adjust sv at a finer grained?
self.adjust_scheduled_values() # adjust after validation
if self._finished():
break
self.tp.update_eidx(1)
continue
# skip batch?
if _skip_batch > 0 and next(_gen) < _skip_batch:
continue
if self.train_discard_batch_f(insts):
continue # discard this batch due to some specific reasons (like noevt)
# run fb (possibly split batch)
self.fb_batch(insts, 1. / _accu_batch)
self.tp.update_iidx(len(insts))
# ==
# only update for certain accu fb runs
_accu_checker += 1
if _accu_checker % _accu_batch == 0:
self.tp.update_uidx(1)
cur_uidx = self.tp.uidx
# get the effective lrate and update
act_lrate = float(
self.lrate.value) # start with the lrate.value
if cur_uidx < _lrate_warmup_steps: # linear increase
act_lrate *= (cur_uidx / _lrate_warmup_steps)
else: # decrease
act_lrate *= _lrate_warmup_factor * (
cur_uidx**conf.lrate_decrease_alpha)
self._run_update(act_lrate, 1.)
# --
# report on training process
if conf.flag_verbose and (
cur_uidx - last_report_uidx) >= conf.report_ufreq:
zlog(
f"Report at {self.tp.current_suffix(False)}: Current act_lrate is {act_lrate}.",
func="plain",
timed=True)
self._run_train_report()
last_report_uidx = cur_uidx
# valid?
if (cur_uidx - last_dev_uidx) >= conf.valid_ufreq:
last_dev_uidx = self.tp.uidx
self.validate()
# todo(+N): do we need to adjust sv at a finer grained?
self.adjust_scheduled_values() # adjust after validation
if self._finished():
break
# =====
zlog(f"Finish training because of: {self._reach_ends()}", func="plain")
zlog(
f"zzzzzfinal: After training, the best point is: {self.tp.info_best()}.",
func="report")
# only finish when reaching any of the endings
def _reach_ends(self):
conf = self.conf
cur_eidx, cur_uidx, cur_aidx = self.tp.eidx, self.tp.uidx, self.tp.aidx
return cur_eidx >= conf.max_eidx, cur_uidx >= conf.max_uidx, cur_aidx >= conf.anneal_times
def _finished(self):
conf = self.conf
cur_eidx, cur_uidx, cur_aidx = self.tp.eidx, self.tp.uidx, self.tp.aidx
return (cur_eidx >= conf.min_eidx) and (
cur_uidx >= conf.min_uidx) and any(self._reach_ends())
# training for one batch
def fb_batch(self, insts: List, loss_factor: float):
with self.train_recorder.go():
res = self._run_fb(insts, loss_factor)
self.train_recorder.record(res)
# --
# do validation and record checkpoints
def validate(self):
conf = self.conf
# report & reset training stat
if self.tp.uidx > 0:
train_result = self._run_train_report(
) # first report training stat
self.train_recorder.reset() # reset training stat
else: # for validate_first
train_result = None
# dev
ss, cur_cidx = self.current_name(), self.tp.cidx
zlog("", func="plain") # empty line
with Timer(info=f"Valid {ss}",
print_date=True), self.model.ema_wrap_dev():
# no validation if specified
if (self.tp.eidx < conf.valid_start_eidx) or (
self.tp.uidx < conf.valid_start_uidx):
zlog("No validation since not the time yet!\n", func="plain")
return
# validate
if len(self.dev_runners
) == 0: # simply use train if there are no dev
zlog(
"Use training results for dev since there are no dev set provided!",
func="warn")
dev_result = train_result
else:
dev_result = self._run_validate(self.dev_runners)
# record
cur_no_bad = (self.tp.eidx < conf.bad_start_eidx) or (
self.tp.uidx < conf.bad_start_uidx)
cur_record_best = (self.tp.cidx >= conf.record_best_cidx)
if_overall_best, if_best, if_anneal = self.tp.update_checkpoint(
train_result, dev_result, cur_no_bad, cur_record_best,
conf.anneal_patience)
# save curr & best
self.save(conf.model_prefix + conf.model_suffix_curr)
if if_overall_best:
zlog("Curr is overall best " +
str(self.tp.info_overall_best()),
func="result")
else:
zlog("Curr not overall best, the overall best is " +
str(self.tp.info_overall_best()),
func="result")
if if_best:
self.save(conf.model_prefix + conf.model_suffix_best)
zlog("Curr is best: " + str(self.tp.info_best()),
func="result")
else:
zlog("Curr not best, the best is " + str(self.tp.info_best()),
func="result")
if cur_cidx >= conf.save_start_cidx and cur_cidx % conf.save_cfreq == 0:
self.save(conf.model_prefix + ss) # speical save
if if_anneal and conf.anneal_restore:
zlog("Restore from previous best model!!", func="plain")
self.load(conf.model_prefix + conf.model_suffix_best, False)
zlog("", func="plain") # empty line
# =====
# template methods which can be overridden
# return one train result
def _run_fb(self, insts: List, loss_factor: float):
res = self.model.loss_on_batch(insts, loss_factor=loss_factor)
return res
# return None
def _run_update(self, lrate: float, grad_factor: float):
self.model.update(lrate, grad_factor)
# print and return train summary
def _run_train_report(self) -> ResultRecord:
x = self.train_recorder.summary()
zlog(f"Train-Info: {OtherHelper.printd_str(x, sep=' ')}")
return ResultRecord(results=x, description=None)
# run and return dev results
def _run_validate(self, dev_runners: List[TestingRunner]) -> ResultRecord:
if len(dev_runners) == 0: return ResultRecord.get_nil()
all_records: List[ResultRecord] = [r.run() for r in dev_runners]
# note: use devs[0] as the criterion, assuming that is the dev itself!!
r = ResultRecord(results={
f"res{ii}": v.results
for ii, v in enumerate(all_records)
},
description={
f"res{ii}": str(v)
for ii, v in enumerate(all_records)
},
score=all_records[0].score)
return r
class RunCenter:
def __init__(self, conf: RunCenterConf, model, t_center: TaskCenter,
d_center: DataCenter):
self.conf = conf
self.model = model
self.t_center = t_center
self.d_center = d_center
# ==
# for train
self.tp = TrainingProgressRecord()
self.train_recorder = StatRecorder(timing=True)
# --
self.lrate = ScheduledValue("lrate", conf.lrate)
self.scheduled_values = OrderedDict([("_lrate", self.lrate)
]) # add all scheduled values
DictHelper.update_dict(self.scheduled_values,
model.get_scheduled_values())
DictHelper.update_dict(self.scheduled_values,
d_center.get_scheduled_values())
# --
self.lrate_warmup_steps = conf.lrate_warmup_uidx
if self.lrate_warmup_steps > 0:
self.lrate_warmup_factor = 1. / (self.lrate_warmup_steps**
conf.lrate_decrease_alpha)
zlog(
f"For lrate-warmup, first {self.lrate_warmup_steps} steps up to {self.lrate.value}, "
f"then decrease with lrate*{self.lrate_warmup_factor}*step^{conf.lrate_decrease_alpha}",
func="plain")
else:
self.lrate_warmup_factor = 1.
# ==
# for ddp
self.ddp_world_size = BK.ddp_world_size()
self.ddp_rank = BK.ddp_rank()
self.is_main_process = BK.is_main_process() # handle the savings!
if self.ddp_world_size > 1:
assert conf.accu_batch == 1, "accu_batch and ddp may conflict!!"
# --
# --
# helpers
def current_name(self):
return self.tp.current_suffix()
def adjust_scheduled_values(self):
# adjust schedule values
ss = self.current_name()
for one_name, one_sv in self.scheduled_values.items():
if one_sv.changeable:
one_sv.adjust_at_ckp(ss, self.tp, extra_info=one_name)
# saving and loading related
def save_progress(self, file: str):
default_json_serializer.to_file(self.tp, file)
zlog(f"Save training progress to {file}", func="io")
def load_progress(self, file: str, forward_stream=False):
d = default_json_serializer.from_file(file)
self.tp.from_json(d)
assert not forward_stream, "Error: 'forward_stream' not supported in this mode!!"
zlog(f"Load training progress from {file}", func="io")
self.adjust_scheduled_values() # also adjust values!
def save(self, prefix: str):
if self.is_main_process: # note: do saving only with main process
self.save_progress(prefix + ".tp.json")
self.model.save(prefix + ".m")
def load(self, prefix: str, load_progress=False, forward_stream=False):
if prefix.endswith(".m"):
prefix = prefix[:-2]
if load_progress:
self.load_progress(prefix + ".tp.json", forward_stream)
self.model.load(prefix + ".m")
# go
# training for one batch
def fb_batch(self, ibatch, loss_factor: float):
with self.train_recorder.go():
res = self.model.loss_on_batch(ibatch, loss_factor)
self.train_recorder.record(res)
# --
def train_finished(self):
conf = self.conf
return self.tp.uidx >= conf.max_uidx
# print and return train summary
def run_train_report(self) -> ResultRecord:
x = self.train_recorder.summary()
zlog(f"Train-Info: {OtherHelper.printd_str(x, sep=' ')}")
# also report uidx_counter/iidx_counter
zlog(f"UidxCounter: {self.tp.uidx_counter}")
zlog(f"IidxCounter: {self.tp.iidx_counter}")
return ResultRecord(results=x, description=None)
# run test on one dataset
def run_test_dataset(self, dataset: ZDataset):
test_recorder = StatRecorder(timing=True)
for ibatch in dataset.yield_batches(loop=False):
with test_recorder.go():
one_res = self.model.predict_on_batch(ibatch)
test_recorder.record(one_res)
# --
# write output
if self.is_main_process: # note: do saving only with main process
dataset.write_insts(None) # use dataset's conf
# --
# eval
x = test_recorder.summary()
zlog(f"Test-Info: {OtherHelper.printd_str(x, sep=' ')}")
aggr = ResultAggregator()
for task in self.t_center.tasks.values():
if task.name not in dataset.tasks:
continue
tn_res: ResultRecord = task.eval_insts(dataset.gold_insts,
dataset.insts,
quite=False)
if tn_res is None:
continue
aggr.add(task.name, tn_res, task.conf.eval_weight)
ret = aggr.get_res()
return ret
# --
# main runs
def do_train(self):
model, t_center, d_center = self.model, self.t_center, self.d_center
conf = self.conf
last_dev_uidx = 0
# --
if conf.valid_first: # valid before training
self.do_dev()
# --
_lrate_warmup_factor, _lrate_warmup_steps = self.lrate_warmup_factor, self.lrate_warmup_steps
_accu_batch = conf.accu_batch
self.adjust_scheduled_values() # once before train
train_yielder = d_center.yield_train_yielder()
while not self.train_finished(): # loop over and over
# sample batch
cur_yielder = next(train_yielder)
# fb for accu_batch/ddp steps (use the same dataset/yielder)
cur_dname = None
for _i0 in range(self.ddp_world_size):
# if 1:
for _i1 in range(_accu_batch):
cur_ibatch = next(cur_yielder)
cur_dname = cur_ibatch.dataset.name
if _i0 == self.ddp_rank: # only for current rank!!
# if 1:
self.fb_batch(cur_ibatch, 1. / _accu_batch)
self.tp.update_iidx(len(cur_ibatch), cur_dname)
# update
self.tp.update_uidx(1, cur_dname)
cur_uidx = self.tp.uidx
# get the effective lrate and update
act_lrate = float(self.lrate.value) # start with the lrate.value
if cur_uidx < _lrate_warmup_steps: # linear increase
act_lrate *= (cur_uidx / _lrate_warmup_steps)
else: # decrease
act_lrate *= _lrate_warmup_factor * (cur_uidx**
conf.lrate_decrease_alpha)
with self.train_recorder.go('update'): # also record this!
self.model.update(act_lrate, 1.)
# valid?
if (cur_uidx - last_dev_uidx) >= conf.valid_ufreq:
last_dev_uidx = cur_uidx
self.do_dev()
# todo(+N): do we need to adjust sv at a finer grained?
self.adjust_scheduled_values() # adjust after validation
# --
zlog(
f"zzzzzfinal: After training, the best point is: {self.tp.info_best()}.",
func="report")
def do_dev(self):
conf = self.conf
# report & reset training stat
if self.tp.uidx > 0:
train_result = self.run_train_report(
) # first report training stat
self.train_recorder.reset() # reset training stat
else: # for validate_first
train_result = ResultRecord.get_nil()
# dev
ss, cur_cidx = self.current_name(), self.tp.cidx
zlog("", func="plain") # empty line
with Timer(info=f"Valid {ss}", print_date=True):
# no validation if specified
if self.tp.uidx < conf.valid_start_uidx:
zlog("No validation since not the time yet!\n", func="plain")
return
# validate
if len(self.d_center.get_datasets(
wset="dev")) == 0: # simply use train if there are no dev
zlog(
"Use training results for dev since there are no dev set provided!",
func="warn")
dev_result = train_result
else:
dev_result = self.do_test("dev")
# record
cur_record_best = (self.tp.cidx >= conf.record_best_start_cidx)
if_overall_best, if_best, if_anneal = self.tp.update_checkpoint(
train_result, dev_result, record_best=cur_record_best)
# save curr & best
self.save(conf.model_save_prefix + conf.model_save_suffix_curr)
if if_overall_best:
zlog("Curr is overall best " +
str(self.tp.info_overall_best()),
func="result")
else:
zlog("Curr not overall best, the overall best is " +
str(self.tp.info_overall_best()),
func="result")
if if_best:
self.save(conf.model_save_prefix + conf.model_save_suffix_best)
zlog("Curr is best: " + str(self.tp.info_best()),
func="result")
else:
zlog("Curr not best, the best is " + str(self.tp.info_best()),
func="result")
if cur_cidx >= conf.save_special_start_cidx and cur_cidx % conf.save_special_cfreq == 0:
self.save(conf.model_save_prefix + ss) # speical save
# --
zlog("", func="plain") # empty line
def do_test(self, wset="test"):
model, t_center, d_center = self.model, self.t_center, self.d_center
conf = self.conf
# --
to_test_datasets = d_center.get_datasets(wset=wset)
t_center.prepare_datasets(to_test_datasets) # re-prepare!!
aggr = ResultAggregator()
for one_ii, one_dataset in enumerate(to_test_datasets):
with Timer(
info=
f"Test({one_ii+1}/{len(to_test_datasets)}): {one_dataset}",
print_date=True):
one_res = self.run_test_dataset(one_dataset)
aggr.add(one_dataset.name, one_res,
one_dataset.conf.group_eval_weight)
ret = aggr.get_res()
return ret