def test_init_run_config_standalone_client(self):
# When `train_distribute` is specified, TF_CONFIG is detected and
# `experimental.remote_cluster` is set use distribute coordinator with
# STANDALONE_CLIENT mode.
config = run_config_lib.RunConfig(
train_distribute=mirrored_strategy.MirroredStrategy(),
experimental_distribute=DistributeConfig(
remote_cluster={"chief": ["fake_worker"]}))
self.assertEqual(config._distribute_coordinator_mode,
dc.CoordinatorMode.STANDALONE_CLIENT)
def test_should_run_distribute_coordinator(self):
"""Tests that should_run_distribute_coordinator return a correct value."""
# We don't use distribute coordinator for local training.
self.assertFalse(
dc_training.should_run_distribute_coordinator(
run_config_lib.RunConfig()))
# When `train_distribute` is not specified, don't use distribute
# coordinator.
with test.mock.patch.dict(
"os.environ", {"TF_CONFIG": json.dumps(TF_CONFIG_WITH_CHIEF)}):
self.assertFalse(
dc_training.should_run_distribute_coordinator(
run_config_lib.RunConfig()))
# When `train_distribute` is specified and TF_CONFIG is detected, use
# distribute coordinator.
with test.mock.patch.dict(
"os.environ", {"TF_CONFIG": json.dumps(TF_CONFIG_WITH_CHIEF)}):
config_with_train_distribute = run_config_lib.RunConfig(
experimental_distribute=DistributeConfig(
train_distribute=mirrored_strategy.MirroredStrategy(
num_gpus=2)))
config_with_eval_distribute = run_config_lib.RunConfig(
experimental_distribute=DistributeConfig(
eval_distribute=mirrored_strategy.MirroredStrategy(
num_gpus=2)))
self.assertTrue(
dc_training.should_run_distribute_coordinator(
config_with_train_distribute))
self.assertFalse(
dc_training.should_run_distribute_coordinator(
config_with_eval_distribute))
# With a master in the cluster, don't run distribute coordinator.
with test.mock.patch.dict(
"os.environ",
{"TF_CONFIG": json.dumps(TF_CONFIG_WITH_MASTER)}):
config = run_config_lib.RunConfig(
experimental_distribute=DistributeConfig(
train_distribute=mirrored_strategy.MirroredStrategy(
num_gpus=2)))
self.assertFalse(dc_training.should_run_distribute_coordinator(config))
def test_previously_unexpected_cluster_spec(self):
with test.mock.patch.dict(
"os.environ", {"TF_CONFIG": json.dumps(TF_CONFIG_WITHOUT_TASK)}):
run_config_lib.RunConfig(
experimental_distribute=DistributeConfig(
train_distribute=mirrored_strategy.MirroredStrategy(num_gpus=2)))
