def build_pipeline(node_id):
with Node('trainer_%d' % node_id):
with Job.current().init_group, Task():
data_arr = Struct(('val', np.array(list(range(10)))))
data = ConstRecord(ops, data_arr)
ds = Dataset(data, name='dataset:%d' % node_id)
full_reader = ds.reader(ops)
total = ops.Const([100])
def inc_total(rec):
ops.Add([total, rec.val()], [total])
epoch_reader = ReaderWithLimit(full_reader, num_iter=3)
pipe(epoch_reader, processor=inc_total)
Job.current().add_stop_signal(epoch_reader.data_finished())
return [total]
def test_upload_checkpoint(self):
try:
tmpdir = tempfile.mkdtemp()
upload_dir = os.path.join(tmpdir, "upload")
os.mkdir(upload_dir)
num_nodes = 3
# The uploaded files do not exist yet.
for node_id in range(num_nodes):
node_name = 'trainer_%d' % node_id
upload_path = os.path.join(upload_dir, node_name)
self.assertFalse(os.path.exists(upload_path))
# Create and run the job runner.
for node_id in range(3):
ws = workspace.C.Workspace()
session = LocalSession(ws)
checkpoint = MultiNodeCheckpointManager(tmpdir, 'minidb')
with Cluster():
with Job() as job:
build_pipeline(node_id)
job.compile(LocalSession)
local_upload_builder = UploadToLocalFile(upload_dir)
job_runner = JobRunner(
job,
checkpoint,
upload_task_group_builder=local_upload_builder)
num_epochs = job_runner.train(session)
self.assertEquals(num_epochs, len(EXPECTED_TOTALS))
# The uploaded files should exist now.
for node_id in range(num_nodes):
node_name = 'trainer_%d' % node_id
upload_path = os.path.join(upload_dir, node_name)
self.assertTrue(os.path.exists(upload_path))
finally:
shutil.rmtree(tmpdir)
def test_ckpt_save_failure(self):
num_nodes = 3
# The goal of this test is to ensure that the job runs
# successfully even if saving a checkpoint fails.
# Hence tmpdir is a non existent directory to emulate a failure
# while saving checkpoints
tmpdir = "/tmp/path_does_not_exist/"
# Check the saving checkpoint failure does not cause job failure
workspace.ResetWorkspace()
for node_id in range(num_nodes):
ws = workspace.C.Workspace()
session = LocalSession(ws)
checkpoint = MultiNodeCheckpointManager(tmpdir, 'minidb')
with Cluster():
with Job() as job:
build_pipeline(node_id)
compiled_job = job.compile(LocalSession)
job_runner = JobRunner(compiled_job, checkpoint)
num_epochs = job_runner(session)
# make sure all epochs are executed even though saving the checkpoint failed
# Saving checkpoint failure should not cause job failure
self.assertEquals(num_epochs, len(EXPECTED_TOTALS))
def build_job():
with Node('reader'):
with Job() as job:
with job.init_group:
init_net = core.Net('init_net')
data_arr = Struct(('val', np.array(range(10))))
data = ConstRecord(init_net, data_arr)
ds = Dataset(data)
full_reader = ds.reader(init_net)
total = init_net.Const([100])
Task(step=init_net)
def inc_total(rec):
net = core.Net('inc_total')
net.Add([total, rec.val()], [total])
return [net]
epoch_reader = ReaderWithLimit(full_reader, num_iter=3)
pipe(epoch_reader, processor=inc_total)
job.add_stop_signal(epoch_reader.data_finished())
total_fetcher = Task(step=core.Net('empty'), outputs=[total])
return job, total_fetcher
def test_ckpt_name_and_load_model_from_ckpts(self):
try:
num_nodes = 3
tmpdir = tempfile.mkdtemp()
# First, check if the checkpoint name generation mechanism is
# correct.
checkpoint = MultiNodeCheckpointManager(tmpdir, 'minidb')
with Cluster():
with Job() as job:
for node_id in range(num_nodes):
build_pipeline(node_id)
compiled_job = job.compile(LocalSession)
checkpoint.init(compiled_job.nodes_to_checkpoint())
for node_id in range(num_nodes):
epoch = 5
node_name = 'trainer_%d' % node_id
expected_db_name = tmpdir + '/' + node_name + '.5'
self.assertEquals(
checkpoint.get_ckpt_db_name(node_name, epoch),
expected_db_name)
shutil.rmtree(tmpdir)
# Next, check mechanism to load model from checkpoints.
tmpdir = tempfile.mkdtemp()
workspace.ResetWorkspace()
for node_id in range(num_nodes):
ws = workspace.C.Workspace()
session = LocalSession(ws)
checkpoint = MultiNodeCheckpointManager(tmpdir, 'minidb')
with Cluster():
with Job() as job:
build_pipeline(node_id)
compiled_job = job.compile(LocalSession)
job_runner = JobRunner(compiled_job, checkpoint)
num_epochs = job_runner(session)
self.assertEquals(num_epochs, len(EXPECTED_TOTALS))
# There are 12 global blobs after finishing up the job runner.
# (only blobs on init_group are checkpointed)
self.assertEquals(len(ws.blobs), 12)
ws = workspace.C.Workspace()
session = LocalSession(ws)
self.assertEquals(len(ws.blobs), 0)
model_blob_names = ['trainer_1/task_2/GivenTensorInt64Fill:0',
'trainer_2/task_2/GivenTensorInt64Fill:0']
checkpoint = MultiNodeCheckpointManager(tmpdir, 'minidb')
with Cluster():
with Job() as job:
for node_id in range(num_nodes):
build_pipeline(node_id)
compiled_job = job.compile(LocalSession)
job_runner = JobRunner(compiled_job, checkpoint)
job_runner.load_blobs_from_checkpoints(
blob_names=model_blob_names, epoch=1, session=session)
# Check that we can successfully load from checkpoints of epochs
# 1 to 4, but not epoch 5.
for epoch in range(1, 5):
self.assertTrue(
job_runner.load_blobs_from_checkpoints(
blob_names=model_blob_names, epoch=epoch,
session=session))
# Check that all the model blobs are loaded.
for blob_name in model_blob_names:
self.assertTrue(ws.has_blob(blob_name))
self.assertEquals(
ws.fetch_blob(blob_name),
np.array([EXPECTED_TOTALS[epoch - 1]]))
self.assertFalse(
job_runner.load_blobs_from_checkpoints(
blob_names=model_blob_names, epoch=5, session=session))
finally:
shutil.rmtree(tmpdir)
def example_job():
with Job() as job:
with job.init_group:
example_loop()
example_task()
return job