def sync_manual(self, computer: Computer, provider: ComputerProvider):
"""
button sync was clicked manually
"""
if not computer.meta:
return
meta = yaml_load(computer.meta)
if 'manual_sync' not in meta:
return
manual_sync = meta['manual_sync']
project_provider = ProjectProvider(self.session)
docker_provider = DockerProvider(self.session)
dockers = docker_provider.get_online()
project = project_provider.by_id(manual_sync['project'])
for docker in dockers:
if docker.computer == computer.name:
continue
source = provider.by_name(docker.computer)
ignore_folders = [
[join('models', project.name), []]
]
sync_directed(self.session, target=computer, source=source,
ignore_folders=ignore_folders)
del meta['manual_sync']
computer.meta = yaml_dump(meta)
provider.update()
def sync(project: str, computer: str, only_from: bool, only_to: bool):
_create_computer()
computer = computer or socket.gethostname()
provider = ComputerProvider(_session)
project_provider = ProjectProvider(_session)
computer = provider.by_name(computer)
computers = provider.all()
folders_excluded = []
p = project_provider.by_name(project)
assert p, f'Project={project} is not found'
ignore = yaml_load(p.ignore_folders)
excluded = []
for f in ignore:
excluded.append(str(f))
folders_excluded.append([join('data', p.name), excluded])
folders_excluded.append([join('models', p.name), []])
for c in computers:
if c.name != computer.name:
if not only_from:
sync_directed(_session, computer, c, folders_excluded)
if not only_to:
sync_directed(_session, c, computer, folders_excluded)
def sync(self):
hostname = socket.gethostname()
try:
provider = ComputerProvider(self.session)
task_synced_provider = TaskSyncedProvider(self.session)
computer = provider.by_name(hostname)
sync_start = now()
if FILE_SYNC_INTERVAL == 0:
time.sleep(1)
else:
computers = provider.all_with_last_activtiy()
computers = [
c for c in computers
if (now() - c.last_activity).total_seconds() < 10
]
computers_names = {c.name for c in computers}
for c, project, tasks in task_synced_provider.for_computer(
computer.name):
if c.name not in computers_names:
self.logger.info(
f'Computer = {c.name} '
f'is offline. Can not sync',
ComponentType.WorkerSupervisor, hostname)
continue
if c.syncing_computer:
continue
excluded = list(map(str,
yaml_load(project.ignore_folders)))
folders_excluded = [[join('data', project.name), excluded],
[join('models', project.name), []]]
computer.syncing_computer = c.name
provider.update()
sync_directed(self.session, c, computer, folders_excluded)
for t in tasks:
task_synced_provider.add(
TaskSynced(computer=computer.name, task=t.id))
time.sleep(FILE_SYNC_INTERVAL)
computer.last_synced = sync_start
computer.syncing_computer = None
provider.update()
except Exception as e:
if Session.sqlalchemy_error(e):
Session.cleanup('FileSync')
self.session = Session.create_session(key='FileSync')
self.logger = create_logger(self.session, 'FileSync')
self.logger.error(traceback.format_exc(),
ComponentType.WorkerSupervisor, hostname)
def copy_remote(session: Session, computer_from: str, path_from: str,
path_to: str):
provider = ComputerProvider(session)
src = provider.by_name(computer_from)
host = socket.gethostname()
if host != computer_from:
c = f'scp -P {src.port} {src.user}@{src.ip}:{path_from} {path_to}'
else:
f'cp {path_from} {path_to}'
subprocess.check_output(c, shell=True)
return os.path.exists(path_to)
def sync_manual(self, computer: Computer, provider: ComputerProvider):
"""
button sync was clicked manually
"""
if not computer.meta:
return
meta = yaml_load(computer.meta)
if 'manual_sync' not in meta:
return
manual_sync = meta['manual_sync']
project_provider = ProjectProvider(self.session)
docker_provider = DockerProvider(self.session)
dockers = docker_provider.get_online()
project = project_provider.by_id(manual_sync['project'])
sync_folders = manual_sync['sync_folders']
ignore_folders = manual_sync['ignore_folders']
sync_folders = correct_folders(sync_folders, project.name)
ignore_folders = correct_folders(ignore_folders, project.name)
if not isinstance(sync_folders, list):
sync_folders = []
if not isinstance(ignore_folders, list):
ignore_folders = []
for docker in dockers:
if docker.computer == computer.name:
continue
source = provider.by_name(docker.computer)
folders = [[s, ignore_folders] for s in sync_folders]
computer.syncing_computer = source.name
provider.update()
try:
sync_directed(
self.session,
target=computer,
source=source,
folders=folders
)
except Exception as e:
self.process_error(e)
del meta['manual_sync']
computer.meta = yaml_dump(meta)
provider.update()
def sync(project: str, computer: str, only_from: bool, only_to: bool,
online: bool):
"""
Syncs specified project on this computer with other computers
"""
check_statuses()
_create_computer()
_create_docker()
computer = computer or socket.gethostname()
provider = ComputerProvider(_session)
project_provider = ProjectProvider(_session)
computer = provider.by_name(computer)
computers = provider.all_with_last_activtiy()
p = project_provider.by_name(project)
assert p, f'Project={project} is not found'
sync_folders = yaml_load(p.sync_folders)
ignore_folders = yaml_load(p.ignore_folders)
sync_folders = correct_folders(sync_folders, p.name)
ignore_folders = correct_folders(ignore_folders, p.name)
if not isinstance(sync_folders, list):
sync_folders = []
if not isinstance(ignore_folders, list):
ignore_folders = []
folders = [[s, ignore_folders] for s in sync_folders]
for c in computers:
if c.name != computer.name:
if online and (now() - c.last_activity).total_seconds() > 100:
continue
if not only_from:
sync_directed(_session, computer, c, folders)
if not only_to:
sync_directed(_session, c, computer, folders)
class Catalyst(Executor, Callback):
def __init__(self, args: Args, report: ReportLayoutInfo, distr_info: dict,
resume: dict, grid_config: dict, trace: str, params: dict):
super().__init__(order=0)
self.resume = resume
self.distr_info = distr_info
self.args = args
self.report = report
self.experiment = None
self.runner = None
self.series_provider = ReportSeriesProvider(self.session)
self.computer_provider = ComputerProvider(self.session)
self.grid_config = grid_config
self.master = True
self.checkpoint_resume = False
self.checkpoint_stage_epoch = 0
self.trace = trace
self.params = params
def callbacks(self):
result = OrderedDict()
if self.master:
result['catalyst'] = self
return result
def on_epoch_start(self, state: RunnerState):
if self.checkpoint_resume and state.stage_epoch == 0:
state.epoch += 1
state.stage_epoch = state.stage_epoch + self.checkpoint_stage_epoch
state.checkpoint_data = {'stage_epoch': state.stage_epoch}
if self.master:
if state.stage_epoch == 0:
self.step.start(1, name=state.stage)
self.step.start(2,
name=f'epoch {state.stage_epoch}',
index=state.stage_epoch)
def on_epoch_end(self, state: RunnerState):
self.step.end(2)
for s in self.report.series:
train = state.metrics.epoch_values['train'][s.key]
val = state.metrics.epoch_values['valid'][s.key]
task_id = self.task.parent or self.task.id
train = ReportSeries(part='train',
name=s.key,
epoch=state.epoch,
task=task_id,
value=train,
time=now(),
stage=state.stage)
val = ReportSeries(part='valid',
name=s.key,
epoch=state.epoch,
task=task_id,
value=val,
time=now(),
stage=state.stage)
self.series_provider.add(train)
self.series_provider.add(val)
if s.key == self.report.metric.name:
best = False
task = self.task
if task.parent:
task = self.task_provider.by_id(task.parent)
if self.report.metric.minimize:
if task.score is None or val.value < task.score:
best = True
else:
if task.score is None or val.value > task.score:
best = True
if best:
task.score = val.value
self.task_provider.update()
def on_stage_start(self, state: RunnerState):
state.loggers = {
'console': VerboseLogger(),
'raise': RaiseExceptionLogger()
}
def on_stage_end(self, state: RunnerState):
self.checkpoint_resume = False
self.checkpoint_stage_epoch = 0
self.step.end(1)
@classmethod
def _from_config(cls, executor: dict, config: Config,
additional_info: dict):
args = Args()
for k, v in executor['args'].items():
v = str(v)
if v in ['False', 'True']:
v = v == 'True'
elif v.isnumeric():
v = int(v)
setattr(args, k, v)
assert 'report_config' in additional_info, 'layout was not filled'
report_config = additional_info['report_config']
grid_cell = additional_info.get('grid_cell')
report = ReportLayoutInfo(report_config)
if len(args.configs) == 0:
args.configs = [args.config]
grid_config = {}
if grid_cell is not None:
grid_config = grid_cells(executor['grid'])[grid_cell][0]
distr_info = additional_info.get('distr_info', {})
resume = additional_info.get('resume')
params = executor.get('params', {})
return cls(args=args,
report=report,
grid_config=grid_config,
distr_info=distr_info,
resume=resume,
trace=executor.get('trace'),
params=params)
def set_dist_env(self, config):
info = self.distr_info
os.environ['MASTER_ADDR'] = info['master_addr']
os.environ['MASTER_PORT'] = str(info['master_port'])
os.environ['WORLD_SIZE'] = str(info['world_size'])
os.environ['RANK'] = str(info['rank'])
distributed_params = config.get('distributed_params', {})
distributed_params['rank'] = 0
config['distributed_params'] = distributed_params
if info['rank'] > 0:
self.master = False
def parse_args_uargs(self):
args, config = parse_args_uargs(self.args, [])
config = merge_dicts_smart(config, self.grid_config)
config = merge_dicts_smart(config, self.params)
if self.distr_info:
self.set_dist_env(config)
return args, config
def _checkpoint_fix_config(self, experiment):
resume = self.resume
if not resume:
return
checkpoint_dir = join(experiment.logdir, 'checkpoints')
os.makedirs(checkpoint_dir, exist_ok=True)
file = 'last_full.pth' if resume.get('load_last') else 'best_full.pth'
path = join(checkpoint_dir, file)
computer = socket.gethostname()
if computer != resume['master_computer']:
master_computer = self.computer_provider.by_name(
resume['master_computer'])
path_from = join(master_computer.root_folder,
str(resume['master_task_id']), 'log',
'checkpoints', file)
self.info(f'copying checkpoint from: computer = '
f'{resume["master_computer"]} path_from={path_from} '
f'path_to={path}')
success = copy_remote(session=self.session,
computer_from=resume['master_computer'],
path_from=path_from,
path_to=path)
if not success:
self.error(f'copying from '
f'{resume["master_computer"]}/'
f'{path_from} failed')
else:
self.info('checkpoint copied successfully')
elif self.task.id != resume['master_task_id']:
path = join(TASK_FOLDER, str(resume['master_task_id']), 'log',
'checkpoints', file)
self.info(f'master_task_id!=task.id, using checkpoint'
f' from task_id = {resume["master_task_id"]}')
if not os.path.exists(path):
self.info(f'no checkpoint at {path}')
return
ckpt = load_checkpoint(path)
stages_config = experiment.stages_config
for k, v in list(stages_config.items()):
if k == ckpt['stage']:
stage_epoch = ckpt['checkpoint_data']['stage_epoch'] + 1
# if it is the last epoch in the stage
if stage_epoch == v['state_params']['num_epochs'] \
or resume.get('load_best'):
del stages_config[k]
break
self.checkpoint_stage_epoch = stage_epoch
v['state_params']['num_epochs'] -= stage_epoch
break
del stages_config[k]
stage = experiment.stages_config[experiment.stages[0]]
for k, v in stage['callbacks_params'].items():
if v.get('callback') == 'CheckpointCallback':
v['resume'] = path
self.info(f'found checkpoint at {path}')
def _checkpoint_fix_callback(self, callbacks: dict):
def mock(state):
pass
for k, c in callbacks.items():
if not isinstance(c, CheckpointCallback):
continue
if c.resume:
self.checkpoint_resume = True
if not self.master:
c.on_epoch_end = mock
c.on_stage_end = mock
def work(self):
args, config = self.parse_args_uargs()
set_global_seed(args.seed)
Experiment, R = import_experiment_and_runner(Path(args.expdir))
runner_params = config.pop('runner_params', {})
experiment = Experiment(config)
runner: Runner = R(**runner_params)
register()
self.experiment = experiment
self.runner = runner
stages = experiment.stages[:]
if self.master:
task = self.task if not self.task.parent \
else self.task_provider.by_id(self.task.parent)
task.steps = len(stages)
self.task_provider.commit()
self._checkpoint_fix_config(experiment)
_get_callbacks = experiment.get_callbacks
def get_callbacks(stage):
res = self.callbacks()
for k, v in _get_callbacks(stage).items():
res[k] = v
self._checkpoint_fix_callback(res)
return res
experiment.get_callbacks = get_callbacks
if experiment.logdir is not None:
dump_environment(config, experiment.logdir, args.configs)
if self.distr_info:
info = yaml_load(self.task.additional_info)
info['resume'] = {
'master_computer': self.distr_info['master_computer'],
'master_task_id': self.task.id - self.distr_info['rank'],
'load_best': True
}
self.task.additional_info = yaml_dump(info)
self.task_provider.commit()
experiment.stages_config = {
k: v
for k, v in experiment.stages_config.items()
if k == experiment.stages[0]
}
runner.run_experiment(experiment, check=args.check)
if self.master and self.trace:
traced = trace_model_from_checkpoint(self.experiment.logdir, self)
torch.jit.save(traced, self.trace)
return {'stage': experiment.stages[-1], 'stages': stages}
class Catalyst(Executor, Callback):
def __init__(self, args: Args, report: ReportLayoutInfo, distr_info: dict,
resume: dict, grid_config: dict, trace: str, params: dict,
**kwargs):
super().__init__(**kwargs)
self.series_provider = ReportSeriesProvider(self.session)
self.computer_provider = ComputerProvider(self.session)
self.memory_provider = MemoryProvider(self.session)
self.order = 0
self.resume = resume
self.distr_info = distr_info
self.args = args
self.report = report
self.experiment = None
self.runner = None
self.grid_config = grid_config
self.master = True
self.trace = trace
self.params = params
self.last_batch_logged = None
self.loader_started_time = None
self.parent = None
self.node = CallbackNode.All
def get_parent_task(self):
if self.parent:
return self.parent
return self.task
def callbacks(self):
result = OrderedDict()
if self.master:
result['catalyst'] = self
return result
def on_loader_start(self, state: State):
self.loader_started_time = now()
def on_epoch_start(self, state: State):
stage_index = self.experiment.stages.index(state.stage_name)
self.step.start(1, name=state.stage_name, index=stage_index)
self.step.start(2, name=f'epoch {state.epoch}', index=state.epoch - 1)
def on_batch_start(self, state: State):
if self.last_batch_logged and state.loader_step != state.loader_len:
if (now() - self.last_batch_logged).total_seconds() < 10:
return
task = self.get_parent_task()
task.batch_index = state.loader_step
task.batch_total = state.loader_len
task.loader_name = state.loader_name
duration = int((now() - self.loader_started_time).total_seconds())
task.epoch_duration = duration
task.epoch_time_remaining = int(
duration *
(task.batch_total / task.batch_index)) - task.epoch_duration
if state.epoch_metrics.get('train_loss') is not None:
task.loss = float(state.epoch_metrics['train_loss'])
if state.epoch_metrics.get('valid_loss') is not None:
task.loss = float(state.epoch_metrics['valid_loss'])
self.task_provider.update()
self.last_batch_logged = now()
def on_epoch_end(self, state: State):
self.step.end(2)
values = state.epoch_metrics
for k, v in values.items():
part = ''
name = k
for loader in state.loaders:
if k.startswith(loader):
part = loader
name = k.replace(loader, '')
if name.startswith('_'):
name = name[1:]
task_id = self.task.parent or self.task.id
series = ReportSeries(part=part,
name=name,
epoch=state.epoch - 1,
task=task_id,
value=v,
time=now(),
stage=state.stage_name)
self.series_provider.add(series)
if name == self.report.metric.name:
best = False
task = self.task
if task.parent:
task = self.task_provider.by_id(task.parent)
if self.report.metric.minimize:
if task.score is None or v < task.score:
best = True
else:
if task.score is None or v > task.score:
best = True
if best:
task.score = v
self.task_provider.update()
def on_stage_end(self, state: State):
self.step.end(1)
@classmethod
def _from_config(cls, executor: dict, config: Config,
additional_info: dict):
args = Args()
for k, v in executor['args'].items():
v = str(v)
if v in ['False', 'True']:
v = v == 'True'
elif v.isnumeric():
v = int(v)
setattr(args, k, v)
assert 'report_config' in additional_info, 'layout was not filled'
report_config = additional_info['report_config']
report = ReportLayoutInfo(report_config)
if len(args.configs) == 0:
args.configs = [args.config]
distr_info = additional_info.get('distr_info', {})
resume = additional_info.get('resume')
params = executor.get('params', {})
params.update(additional_info.get('params', {}))
grid_config = executor.copy()
grid_config.pop('args', '')
return cls(args=args,
report=report,
grid_config=grid_config,
distr_info=distr_info,
resume=resume,
trace=executor.get('trace'),
params=params)
def set_dist_env(self, config):
info = self.distr_info
os.environ['MASTER_ADDR'] = info['master_addr']
os.environ['MASTER_PORT'] = str(info['master_port'])
os.environ['WORLD_SIZE'] = str(info['world_size'])
os.environ['RANK'] = str(info['rank'])
os.environ['LOCAL_RANK'] = "0"
distributed_params = config.get('distributed_params', {})
distributed_params['rank'] = info['rank']
config['distributed_params'] = distributed_params
torch.cuda.set_device(0)
torch.distributed.init_process_group(backend="nccl",
init_method="env://")
if info['rank'] > 0:
self.master = False
self.node = CallbackNode.Worker
else:
self.node = CallbackNode.Master
def parse_args_uargs(self):
args, config = parse_args_uargs(self.args, [])
config = merge_dicts_smart(config, self.grid_config)
config = merge_dicts_smart(config, self.params)
if self.distr_info:
self.set_dist_env(config)
return args, config
def _fix_memory(self, experiment):
if not torch.cuda.is_available():
return
max_memory = torch.cuda.get_device_properties(0).total_memory / (2**30)
stages_config = experiment.stages_config
for k, v in list(stages_config.items()):
query = {}
# noinspection PyProtectedMember
for kk, vv in experiment._config['model_params'].items():
query[kk] = vv
for kk, vv in v['data_params'].items():
query[kk] = vv
variants = self.memory_provider.find(query)
variants = [v for v in variants if v.memory < max_memory]
if len(variants) == 0:
continue
variant = max(variants, key=lambda x: x.memory)
v['data_params']['batch_size'] = variant.batch_size
def _checkpoint_fix_config(self, experiment):
resume = self.resume
if not resume:
return
if experiment.logdir is None:
return
checkpoint_dir = join(experiment.logdir, 'checkpoints')
os.makedirs(checkpoint_dir, exist_ok=True)
file = 'last_full.pth' if resume.get('load_last') else 'best_full.pth'
path = join(checkpoint_dir, file)
computer = socket.gethostname()
if computer != resume['master_computer']:
master_computer = self.computer_provider.by_name(
resume['master_computer'])
path_from = join(master_computer.root_folder,
str(resume['master_task_id']), experiment.logdir,
'checkpoints', file)
self.info(f'copying checkpoint from: computer = '
f'{resume["master_computer"]} path_from={path_from} '
f'path_to={path}')
success = copy_remote(session=self.session,
computer_from=resume['master_computer'],
path_from=path_from,
path_to=path)
if not success:
self.error(f'copying from '
f'{resume["master_computer"]}/'
f'{path_from} failed')
else:
self.info('checkpoint copied successfully')
elif self.task.id != resume['master_task_id']:
path = join(TASK_FOLDER, str(resume['master_task_id']),
experiment.logdir, 'checkpoints', file)
self.info(f'master_task_id!=task.id, using checkpoint'
f' from task_id = {resume["master_task_id"]}')
if not os.path.exists(path):
self.info(f'no checkpoint at {path}')
return
ckpt = load_checkpoint(path)
stages_config = experiment.stages_config
for k, v in list(stages_config.items()):
if k == ckpt['stage']:
stage_epoch = ckpt['checkpoint_data']['epoch'] + 1
# if it is the last epoch in the stage
if stage_epoch >= v['state_params']['num_epochs'] \
or resume.get('load_best'):
del stages_config[k]
break
self.checkpoint_stage_epoch = stage_epoch
v['state_params']['num_epochs'] -= stage_epoch
break
del stages_config[k]
stage = experiment.stages_config[experiment.stages[0]]
for k, v in stage['callbacks_params'].items():
if v.get('callback') == 'CheckpointCallback':
v['resume'] = path
self.info(f'found checkpoint at {path}')
def _checkpoint_fix_callback(self, callbacks: dict):
def mock(state):
pass
for k, c in callbacks.items():
if not isinstance(c, CheckpointCallback):
continue
if c.resume:
self.checkpoint_resume = True
if not self.master:
c.on_epoch_end = mock
c.on_stage_end = mock
c.on_batch_start = mock
def work(self):
args, config = self.parse_args_uargs()
set_global_seed(args.seed)
Experiment, R = import_experiment_and_runner(Path(args.expdir))
runner_params = config.pop('runner_params', {})
experiment = Experiment(config)
runner: Runner = R(**runner_params)
self.experiment = experiment
self.runner = runner
stages = experiment.stages[:]
if self.task.parent:
self.parent = self.task_provider.by_id(self.task.parent)
if self.master:
task = self.get_parent_task()
task.steps = len(stages)
self.task_provider.commit()
self._checkpoint_fix_config(experiment)
self._fix_memory(experiment)
_get_callbacks = experiment.get_callbacks
def get_callbacks(stage):
res = self.callbacks()
for k, v in _get_callbacks(stage).items():
res[k] = v
self._checkpoint_fix_callback(res)
return res
experiment.get_callbacks = get_callbacks
if experiment.logdir is not None:
dump_environment(config, experiment.logdir, args.configs)
if self.distr_info:
info = yaml_load(self.task.additional_info)
info['resume'] = {
'master_computer': self.distr_info['master_computer'],
'master_task_id': self.task.id - self.distr_info['rank'],
'load_best': True
}
self.task.additional_info = yaml_dump(info)
self.task_provider.commit()
experiment.stages_config = {
k: v
for k, v in experiment.stages_config.items()
if k == experiment.stages[0]
}
runner.run_experiment(experiment)
if runner.state.exception:
raise runner.state.exception
if self.master and self.trace:
traced = trace_model_from_checkpoint(self.experiment.logdir, self)
torch.jit.save(traced, self.trace)
return {'stage': experiment.stages[-1], 'stages': stages}
def sync(self):
hostname = socket.gethostname()
try:
provider = ComputerProvider(self.session)
task_synced_provider = TaskSyncedProvider(self.session)
computer = provider.by_name(hostname)
sync_start = now()
if FILE_SYNC_INTERVAL == 0:
time.sleep(1)
else:
self.sync_manual(computer, provider)
computers = provider.all_with_last_activtiy()
computers = [
c for c in computers
if (now() - c.last_activity).total_seconds() < 10
]
computers_names = {c.name for c in computers}
for c, project, tasks in task_synced_provider.for_computer(
computer.name):
if c.sync_with_this_computer:
if c.name not in computers_names:
self.logger.info(f'Computer = {c.name} '
f'is offline. Can not sync',
ComponentType.WorkerSupervisor,
hostname)
continue
if c.syncing_computer:
continue
sync_folders = yaml_load(project.sync_folders)
ignore_folders = yaml_load(project.ignore_folders)
sync_folders = correct_folders(sync_folders,
project.name)
ignore_folders = correct_folders(ignore_folders,
project.name)
if not isinstance(sync_folders, list):
sync_folders = []
if not isinstance(ignore_folders, list):
ignore_folders = []
folders = [[s, ignore_folders] for s in sync_folders]
computer.syncing_computer = c.name
provider.update()
sync_directed(self.session, c, computer, folders)
for t in tasks:
task_synced_provider.add(
TaskSynced(computer=computer.name, task=t.id)
)
time.sleep(FILE_SYNC_INTERVAL)
computer.last_synced = sync_start
computer.syncing_computer = None
provider.update()
except Exception as e:
self.process_error(e)