def _process_deployment(self, task_declaration: TaskDeclaration):
assert task_declaration.state == TaskDeclaration.State.DEPLOYMENT
with async_commit():
save = False
for ta in task_declaration.get_task_assignments(states=(TaskAssignment.State.READY,)):
if self._is_task_assignment_allowed(task_declaration, ta):
ta.state = TaskAssignment.State.ACCEPTED
ta.save()
task_declaration.workers_needed -= 1
save = True
else:
ta.state = TaskAssignment.State.REJECTED
ta.save()
for va in task_declaration.get_verification_assignments(states=(VerificationAssignment.State.READY,)):
if self._is_verification_assignment_allowed(task_declaration, va):
va.state = VerificationAssignment.State.ACCEPTED
va.save()
task_declaration.verifiers_needed -= 1
save = True
else:
va.state = VerificationAssignment.State.REJECTED
va.save()
# save if were changes
if save:
task_declaration.save()
ready_to_start = task_declaration.workers_needed == 0 and task_declaration.verifiers_needed == 0
logger.info('{} ready: {} workers_needed: {} verifiers_needed: {}'.format(
task_declaration, ready_to_start, task_declaration.workers_needed, task_declaration.verifiers_needed))
if ready_to_start:
self._assign_initial_train_data(task_declaration)
return
if not save:
# recheck how many workers and verifiers really accepted
accepted_workers_count = len(task_declaration.get_task_assignments(
states=(TaskAssignment.State.ACCEPTED,)))
accepted_verifiers_count = len(task_declaration.get_verification_assignments(
states=(VerificationAssignment.State.ACCEPTED,)))
if accepted_workers_count == task_declaration.workers_requested \
and accepted_verifiers_count == task_declaration.verifiers_requested:
logger.info('All performers are accepted, start train')
task_declaration.workers_needed = 0
task_declaration.verifiers_needed = 0
self._assign_initial_train_data(task_declaration)
def _process_epoch_in_progress(self, task_declaration: TaskDeclaration):
assert task_declaration.state == TaskDeclaration.State.EPOCH_IN_PROGRESS
task_assignments = task_declaration.get_task_assignments(
states=(
TaskAssignment.State.TRAINING,
TaskAssignment.State.FINISHED,
)
)
failed = False
finished_task_assignments = []
count_timeout = 0
with async_commit():
for ta in task_assignments:
if ta.state == TaskAssignment.State.TRAINING:
if ta.iteration_is_finished:
ta.state = TaskAssignment.State.FINISHED
ta.save()
if ta.state == TaskAssignment.State.FINISHED:
if ta.train_result.error:
failed = True
else:
finished_task_assignments.append(ta)
continue
train_timeout = settings.WAIT_TRAIN_TIMEOUT
now = datetime.datetime.utcnow().replace(tzinfo=ta.train_result.modified_at.tzinfo)
if (now - ta.train_result.modified_at).total_seconds() > train_timeout:
ta.state = TaskAssignment.State.TIMEOUT
ta.save()
logger.info('Timeout of waiting for {}'.format(ta))
count_timeout += 1
if failed:
logger.info('{} is failed'.format(task_declaration))
task_declaration.state = TaskDeclaration.State.FAILED
task_declaration.save()
return
if count_timeout:
task_declaration.workers_needed += count_timeout
self._republish_for_train(task_declaration)
return
if len(finished_task_assignments) < task_declaration.workers_requested:
logger.info('Wait for finish of training for {} iteration {}'.format(
task_declaration, task_declaration.current_iteration))
return
if task_declaration.current_iteration > 1:
self._save_loss_and_accuracy(task_declaration, finished_task_assignments)
self._assign_verification_data(task_declaration, finished_task_assignments)
def _reassign_train_data(self, task_declaration: TaskDeclaration):
assert task_declaration.state == TaskDeclaration.State.DEPLOYMENT_TRAIN
task_assignments = task_declaration.get_task_assignments(
states=(
TaskAssignment.State.ACCEPTED,
TaskAssignment.State.TIMEOUT,
TaskAssignment.State.FAKE_RESULTS
)
)
# split by state
accepted_task_assignment = []
failed_task_assignments = []
for ta in task_assignments:
if ta.state == TaskAssignment.State.ACCEPTED:
accepted_task_assignment.append(ta)
continue
if ta.state == TaskAssignment.State.TIMEOUT:
failed_task_assignments.append(ta)
continue
if ta.state == TaskAssignment.State.FAKE_RESULTS:
failed_task_assignments.append(ta)
continue
assert False and 'Check query!'
assert len(failed_task_assignments) == len(accepted_task_assignment)
# assign data to new accepted task_assignments
for index, ta in enumerate(accepted_task_assignment):
failed_ta = failed_task_assignments[index]
# reassign train data
# retrieve data which producer is able to encrypt
train_data = TrainData.get_with_initial_data(
asset_id=failed_ta.train_data_id,
db=self.db,
encryption=self.encryption
)
train_data.task_assignment_id = ta.asset_id
# share data with new worker
train_data.set_encryption_key(ta.worker.enc_key)
train_data.save()
ta.train_data_id = train_data.asset_id
ta.state = TaskAssignment.State.TRAINING
ta.save()
failed_ta.state = TaskAssignment.State.FORGOTTEN
failed_ta.save()
task_declaration.state = TaskDeclaration.State.EPOCH_IN_PROGRESS
task_declaration.save()
def _reassign_verification_data(self, task_declaration: TaskDeclaration):
verification_assignments = task_declaration.get_verification_assignments(
states=(VerificationAssignment.State.ACCEPTED, VerificationAssignment.State.TIMEOUT)
)
# split accepted and overdue
accepted_verification_assignments = []
timeout_verification_assignments = []
for va in verification_assignments:
if va.state == VerificationAssignment.State.ACCEPTED:
accepted_verification_assignments.append(va)
continue
if va.state == VerificationAssignment.State.TIMEOUT:
timeout_verification_assignments.append(va)
continue
assert False and 'Check query!'
assert len(accepted_verification_assignments) == len(timeout_verification_assignments)
train_results = [
{
'worker_id': ta.worker_id,
'result': ta.train_result.weights_ipfs
}
for ta in task_declaration.get_task_assignments(states=(TaskAssignment.State.FINISHED,))
]
for index, va in enumerate(accepted_verification_assignments):
assert va.verification_data_id is None
verification_data = VerificationData.create(
verification_assignment_id=va.asset_id,
# share data with verifier
public_key=va.verifier.enc_key,
test_dir_ipfs=task_declaration.dataset.test_dir_ipfs,
model_code_ipfs=task_declaration.train_model.code_ipfs,
train_results=train_results,
db=self.db,
encryption=self.encryption
)
va.verification_data_id = verification_data.asset_id
va.state = VerificationAssignment.State.VERIFYING
va.save()
failed_va = timeout_verification_assignments[index]
failed_va.state = VerificationAssignment.State.FORGOTTEN
failed_va.save()
task_declaration.state = TaskDeclaration.State.VERIFY_IN_PROGRESS
task_declaration.save()
def _republish_for_train(self, task_declaration: TaskDeclaration):
assert task_declaration.workers_needed > 0
task_declaration.state = TaskDeclaration.State.DEPLOYMENT_TRAIN
task_declaration.current_iteration_retry += 1
task_declaration.save()
task_assignments = task_declaration.get_task_assignments(
states=(TaskAssignment.State.REJECTED,)
)
for ta in task_assignments:
ta.state = TaskAssignment.State.REASSIGN
# return back ownership
ta.save(recipients=ta.worker.address)
def _update_train_data_for_next_iteration(self, task_declaration: TaskDeclaration):
assert task_declaration.state == TaskDeclaration.State.VERIFY_IN_PROGRESS
task_declaration.current_iteration += 1
task_declaration.current_iteration_retry = 0
task_declaration.progress = (
task_declaration.current_iteration * task_declaration.epochs_in_iteration * 100
/ task_declaration.epochs)
count_ta = 0
for ta in task_declaration.get_task_assignments(states=(TaskAssignment.State.FINISHED,)):
train_data = ta.train_data
# share data to worker
train_data.set_encryption_key(ta.worker.enc_key)
train_data.save()
ta.state = TaskAssignment.State.TRAINING
ta.save()
count_ta += 1
assert task_declaration.workers_requested == count_ta
task_declaration.state = TaskDeclaration.State.EPOCH_IN_PROGRESS
task_declaration.save()
def _assign_initial_train_data(self, task_declaration: TaskDeclaration):
assert task_declaration.state == TaskDeclaration.State.DEPLOYMENT
# start of train
task_declaration.current_iteration += 1
task_declaration.current_iteration_retry = 0
accepted_task_assignment = task_declaration.get_task_assignments(states=(TaskAssignment.State.ACCEPTED,))
count_ta = 0
train_dirs_ipfs, files = Directory(multihash=task_declaration.dataset.train_dir_ipfs).ls()
test_dirs_ipfs, files = Directory(multihash=task_declaration.dataset.test_dir_ipfs).ls()
all_train_chunks_ipfs = self._chunk_it(
iterable=[x.multihash for x in train_dirs_ipfs],
count=task_declaration.workers_requested
)
assert len(all_train_chunks_ipfs) == task_declaration.workers_requested
all_test_chunks_ipfs = self._chunk_it(
iterable=[x.multihash for x in test_dirs_ipfs],
count=task_declaration.workers_requested
)
assert len(all_test_chunks_ipfs) == task_declaration.workers_requested
list_td_ta = []
with async_commit():
# create TrainData
for index, task_assignment in enumerate(accepted_task_assignment):
train_chunks_ipfs = all_train_chunks_ipfs[index]
test_chunks_ipfs = all_test_chunks_ipfs[index]
train_data = TrainData.create(
model_code_ipfs=task_declaration.train_model.code_ipfs,
train_chunks_ipfs=train_chunks_ipfs,
test_chunks_ipfs=test_chunks_ipfs,
data_index=index,
db=self.db,
encryption=self.encryption
)
list_td_ta.append((train_data, task_assignment))
logger.debug('Created {}, train chunks: {}, count:{}, test chunks: {}, count:{}'.format(
train_data, train_chunks_ipfs, len(train_chunks_ipfs), test_chunks_ipfs, len(test_chunks_ipfs)))
count_ta += 1
assert task_declaration.workers_requested == count_ta
with async_commit():
# share to worker
for train_data, task_assignment in list_td_ta:
train_data.task_assignment_id = task_assignment.asset_id
train_data.set_encryption_key(task_assignment.worker.enc_key)
train_data.save()
task_assignment.train_data_id = train_data.asset_id
task_assignment.state = TaskAssignment.State.TRAINING
task_assignment.save()
task_declaration.state = TaskDeclaration.State.EPOCH_IN_PROGRESS
task_declaration.save()