def test_launch_layer_cifar(
collect_trial_profiles: Callable[[int], None]) -> None:
config = conf.load_config(
conf.cv_examples_path("cifar10_pytorch/const.yaml"))
config = conf.set_max_length(config, {"batches": 200})
config = conf.set_slots_per_trial(config, 1)
config = conf.set_profiling_enabled(config)
config = conf.set_entrypoint(
config,
"python3 -m determined.launch.horovod --autohorovod --trial model_def:CIFARTrial"
)
experiment_id = exp.run_basic_test_with_temp_config(
config, conf.cv_examples_path("cifar10_pytorch"), 1)
trials = exp.experiment_trials(experiment_id)
(Determined(conf.make_master_url()).get_trial(
trials[0].trial.id).select_checkpoint(latest=True).load(
map_location="cpu"))
collect_trial_profiles(trials[0].trial.id)
assert exp.check_if_string_present_in_trial_logs(
trials[0].trial.id,
"allocation stopped after resources exited successfully with a zero exit code",
)
def test_pytorch_parallel() -> None:
config = conf.load_config(conf.tutorials_path("mnist_pytorch/const.yaml"))
config = conf.set_slots_per_trial(config, 8)
config = conf.set_max_length(config, {"batches": 200})
config = conf.set_tensor_auto_tuning(config, True)
config = conf.set_perform_initial_validation(config, True)
exp_id = exp.run_basic_test_with_temp_config(
config, conf.tutorials_path("mnist_pytorch"), 1)
exp.assert_performed_initial_validation(exp_id)
# Check on record/batch counts we emitted in logs.
validation_size = 10000
global_batch_size = config["hyperparameters"]["global_batch_size"]
num_workers = config.get("resources", {}).get("slots_per_trial", 1)
global_batch_size = config["hyperparameters"]["global_batch_size"]
scheduling_unit = config.get("scheduling_unit", 100)
per_slot_batch_size = global_batch_size // num_workers
exp_val_batches = (validation_size +
(per_slot_batch_size - 1)) // per_slot_batch_size
patterns = [
# Expect two copies of matching training reports.
f"trained: {scheduling_unit * global_batch_size} records.*in {scheduling_unit} batches",
f"trained: {scheduling_unit * global_batch_size} records.*in {scheduling_unit} batches",
f"validated: {validation_size} records.*in {exp_val_batches} batches",
]
trial_id = exp.experiment_trials(exp_id)[0].trial.id
exp.assert_patterns_in_trial_logs(trial_id, patterns)
def test_epoch_sync(num_workers: int, global_batch_size: int,
dataset_len: int) -> None:
"""
Test that epoch_idx is synchronized across all workers regardless of whether the
number of batches is evenly divisible by the number of workers.
"""
config = conf.load_config(conf.fixtures_path("pytorch_no_op/const.yaml"))
config = conf.set_slots_per_trial(config, num_workers)
max_len_batches = 10
config = conf.set_max_length(config, {"batches": max_len_batches})
config = conf.set_hparam(config, "dataset_len", dataset_len)
config = conf.set_global_batch_size(config, global_batch_size)
e_id = exp.run_basic_test_with_temp_config(
config, conf.fixtures_path("pytorch_no_op"), 1)
t_id = exp.experiment_trials(e_id)[0].trial.id
batches_per_epoch = (dataset_len + global_batch_size -
1) // global_batch_size # ceil
for batch_idx in range(max_len_batches):
epoch_idx = batch_idx // batches_per_epoch
for rank in range(config["resources"]["slots_per_trial"]):
assert exp.check_if_string_present_in_trial_logs(
t_id,
f"rank {rank} finished batch {batch_idx} in epoch {epoch_idx}")
def test_detr_coco_pytorch_distributed() -> None:
config = conf.load_config(
conf.cv_examples_path("detr_coco_pytorch/const_fake.yaml"))
config = conf.set_max_length(config, {"batches": 200})
config = conf.set_slots_per_trial(config, 2)
exp.run_basic_test_with_temp_config(
config, conf.cv_examples_path("detr_coco_pytorch"), 1)
def test_faster_rcnn() -> None:
config = conf.load_config(conf.experimental_path("trial/FasterRCNN_tp/16-gpus.yaml"))
config = conf.set_max_length(config, {"batches": 128})
config = conf.set_slots_per_trial(config, 1)
exp.run_basic_test_with_temp_config(
config, conf.experimental_path("trial/FasterRCNN_tp"), 1, max_wait_secs=4800
)
def test_pytorch_const_native_parallel() -> None:
config = conf.load_config(conf.tutorials_path("mnist_pytorch/const.yaml"))
config = conf.set_slots_per_trial(config, 8)
config = conf.set_native_parallel(config, True)
config = conf.set_max_length(config, {"batches": 200})
exp.run_basic_test_with_temp_config(config,
conf.tutorials_path("mnist_pytorch"),
1)
def test_language_modeling_mlm() -> None:
example_path = conf.model_hub_examples_path("huggingface/language-modeling")
config = conf.load_config(os.path.join(example_path, "mlm_config.yaml"))
config = conf.set_slots_per_trial(config, 8)
config = conf.set_global_batch_size(config, 16)
config = conf.set_max_length(config, {"batches": 200})
config = set_docker_image(config)
exp.run_basic_test_with_temp_config(config, example_path, 1)
def test_multiple_choice_swag() -> None:
example_path = conf.model_hub_examples_path("huggingface/multiple-choice")
config = conf.load_config(os.path.join(example_path, "swag_config.yaml"))
config = conf.set_slots_per_trial(config, 8)
config = conf.set_global_batch_size(config, 64)
config = conf.set_max_length(config, {"batches": 200})
config = set_docker_image(config)
exp.run_basic_test_with_temp_config(config, example_path, 1)
def test_pytorch_const_native_parallel() -> None:
config = conf.load_config(
conf.official_examples_path("mnist_pytorch/const.yaml"))
config = conf.set_slots_per_trial(config, 8)
config = conf.set_native_parallel(config, True)
config = conf.set_max_steps(config, 2)
exp.run_basic_test_with_temp_config(
config, conf.official_examples_path("mnist_pytorch"), 1)
def test_squad_v2_with_beam_search() -> None:
example_path = conf.model_hub_examples_path("huggingface/question-answering")
config = conf.load_config(os.path.join(example_path, "squad_v2_beam_search.yaml"))
config = conf.set_slots_per_trial(config, 8)
config = conf.set_global_batch_size(config, 16)
config = conf.set_max_length(config, {"batches": 200})
config = set_docker_image(config)
exp.run_basic_test_with_temp_config(config, example_path, 1)
def test_token_classification_ner() -> None:
example_path = conf.model_hub_examples_path("huggingface/token-classification")
config = conf.load_config(os.path.join(example_path, "ner_config.yaml"))
config = conf.set_slots_per_trial(config, 8)
config = conf.set_global_batch_size(config, 32)
config = conf.set_max_length(config, {"batches": 200})
config = set_docker_image(config)
exp.run_basic_test_with_temp_config(config, example_path, 1)
def run_tf_keras_dcgan_example() -> None:
config = conf.load_config(
conf.gan_examples_path("dcgan_tf_keras/const.yaml"))
config = conf.set_max_length(config, {"batches": 200})
config = conf.set_min_validation_period(config, {"batches": 200})
config = conf.set_slots_per_trial(config, 8)
config = conf.set_tf2_image(config)
exp.run_basic_test_with_temp_config(
config, conf.gan_examples_path("dcgan_tf_keras"), 1)
def test_tf_keras_mnist_parallel() -> None:
config = conf.load_config(
conf.tutorials_path("fashion_mnist_tf_keras/const.yaml"))
config = conf.set_slots_per_trial(config, 8)
config = conf.set_max_length(config, {"batches": 200})
experiment_id = exp.run_basic_test_with_temp_config(
config, conf.tutorials_path("fashion_mnist_tf_keras"), 1)
trials = exp.experiment_trials(experiment_id)
assert len(trials) == 1
def test_squad_amp() -> None:
example_path = conf.model_hub_examples_path("huggingface/question-answering")
config = conf.load_config(os.path.join(example_path, "squad.yaml"))
config = conf.set_slots_per_trial(config, 8)
config = conf.set_global_batch_size(config, 64)
config = conf.set_max_length(config, {"batches": 200})
config = conf.set_hparam(config, "use_apex_amp", True)
config = set_docker_image(config)
exp.run_basic_test_with_temp_config(config, example_path, 1)
def test_text_classification_xnli_amp() -> None:
example_path = conf.model_hub_examples_path("huggingface/text-classification")
config = conf.load_config(os.path.join(example_path, "xnli_config.yaml"))
config = conf.set_slots_per_trial(config, 8)
config = conf.set_global_batch_size(config, 128)
config = conf.set_max_length(config, {"batches": 200})
config = conf.set_hparam(config, "use_apex_amp", True)
config = set_docker_image(config)
exp.run_basic_test_with_temp_config(config, example_path, 1)
def test_pytorch_parallel() -> None:
config = conf.load_config(conf.tutorials_path("mnist_pytorch/const.yaml"))
config = conf.set_slots_per_trial(config, 8)
config = conf.set_max_length(config, {"batches": 200})
config = conf.set_tensor_auto_tuning(config, True)
config = conf.set_perform_initial_validation(config, True)
exp_id = exp.run_basic_test_with_temp_config(
config, conf.tutorials_path("mnist_pytorch"), 1, has_zeroth_step=True)
exp.assert_performed_initial_validation(exp_id)
def test_pytorch_parallel() -> None:
config = conf.load_config(conf.official_examples_path("trial/mnist_pytorch/const.yaml"))
config = conf.set_slots_per_trial(config, 8)
config = conf.set_native_parallel(config, False)
config = conf.set_max_length(config, {"batches": 200})
config = conf.set_tensor_auto_tuning(config, True)
exp.run_basic_test_with_temp_config(
config, conf.official_examples_path("trial/mnist_pytorch"), 1
)
def test_tensorpack_native_parallel() -> None:
config = conf.load_config(
conf.official_examples_path("trial/mnist_tp/const.yaml"))
config = conf.set_slots_per_trial(config, 8)
config = conf.set_native_parallel(config, True)
config = conf.set_max_length(config, {"batches": 32})
experiment_id = exp.run_basic_test_with_temp_config(
config, conf.official_examples_path("trial/mnist_tp"), 1)
trials = exp.experiment_trials(experiment_id)
assert len(trials) == 1
def test_estimator_when_detecting_gpus() -> None:
config = conf.load_config(
conf.fixtures_path("estimator_gpu_detection/const.yaml"))
config = conf.set_slots_per_trial(config, 8)
config = conf.set_max_length(config, {"batches": 200})
exp.run_basic_test_with_temp_config(
config,
conf.fixtures_path("estimator_gpu_detection/"),
1,
has_zeroth_step=False)
def test_tf_keras_mnist_parallel() -> None:
config = conf.load_config(conf.official_examples_path("fashion_mnist_tf_keras/const.yaml"))
config = conf.set_slots_per_trial(config, 8)
config = conf.set_native_parallel(config, False)
config = conf.set_max_steps(config, 2)
experiment_id = exp.run_basic_test_with_temp_config(
config, conf.official_examples_path("fashion_mnist_tf_keras"), 1
)
trials = exp.experiment_trials(experiment_id)
assert len(trials) == 1
def test_tf_keras_single_gpu(tf2: bool) -> None:
config = conf.load_config(conf.official_examples_path("cifar10_cnn_tf_keras/const.yaml"))
config = conf.set_slots_per_trial(config, 1)
config = conf.set_max_steps(config, 2)
config = conf.set_tf2_image(config) if tf2 else conf.set_tf1_image(config)
experiment_id = exp.run_basic_test_with_temp_config(
config, conf.official_examples_path("cifar10_cnn_tf_keras"), 1
)
trials = exp.experiment_trials(experiment_id)
assert len(trials) == 1
def test_on_trial_close_callback() -> None:
config = conf.load_config(conf.fixtures_path("estimator_no_op/const.yaml"))
config = conf.set_slots_per_trial(config, 8)
config = conf.set_max_length(config, {"batches": 3})
exp_id = exp.run_basic_test_with_temp_config(
config, conf.fixtures_path("estimator_no_op"), 1)
assert exp.check_if_string_present_in_trial_logs(
exp.experiment_trials(exp_id)[0].trial.id,
"rank 0 has completed on_trial_close")
def test_mnist_estimator_const_parallel(tf2: bool) -> None:
config = conf.load_config(
conf.fixtures_path("mnist_estimator/single-multi-slot.yaml"))
config = conf.set_slots_per_trial(config, 8)
config = conf.set_max_length(config, {"batches": 200})
config = conf.set_tf2_image(config) if tf2 else conf.set_tf1_image(config)
config = conf.set_perform_initial_validation(config, True)
exp_id = exp.run_basic_test_with_temp_config(
config, conf.cv_examples_path("mnist_estimator"), 1)
exp.assert_performed_initial_validation(exp_id)
def test_pytorch_gan_parallel() -> None:
config = conf.load_config(
conf.gan_examples_path("gan_mnist_pytorch/const.yaml"))
config = conf.set_max_length(config, {"batches": 200})
config = conf.set_slots_per_trial(config, 8)
experiment_id = exp.run_basic_test_with_temp_config(
config, conf.gan_examples_path("gan_mnist_pytorch"), 1)
trials = exp.experiment_trials(experiment_id)
(Determined(conf.make_master_url()).get_trial(
trials[0]["id"]).select_checkpoint(latest=True).load(
map_location="cpu"))
def test_tensorpack_parallel(aggregation_frequency: int) -> None:
config = conf.load_config(
conf.official_examples_path("mnist_tp/const.yaml"))
config = conf.set_slots_per_trial(config, 8)
config = conf.set_native_parallel(config, False)
config = conf.set_max_steps(config, 2)
config = conf.set_aggregation_frequency(config, aggregation_frequency)
experiment_id = exp.run_basic_test_with_temp_config(
config, conf.official_examples_path("mnist_tp"), 1)
trials = exp.experiment_trials(experiment_id)
assert len(trials) == 1
def test_pytorch_const_parallel(aggregation_frequency: int, use_amp: bool) -> None:
if use_amp and aggregation_frequency > 1:
pytest.skip("Mixed precision is not support with aggregation frequency > 1.")
config = conf.load_config(conf.tutorials_path("mnist_pytorch/const.yaml"))
config = conf.set_slots_per_trial(config, 8)
config = conf.set_max_length(config, {"batches": 200})
config = conf.set_aggregation_frequency(config, aggregation_frequency)
if use_amp:
config = conf.set_amp_level(config, "O1")
exp.run_basic_test_with_temp_config(config, conf.tutorials_path("mnist_pytorch"), 1)
def test_tf_keras_parallel(aggregation_frequency: int, tf2: bool) -> None:
config = conf.load_config(
conf.cv_examples_path("cifar10_tf_keras/const.yaml"))
config = conf.set_slots_per_trial(config, 8)
config = conf.set_max_length(config, {"batches": 200})
config = conf.set_aggregation_frequency(config, aggregation_frequency)
config = conf.set_tf2_image(config) if tf2 else conf.set_tf1_image(config)
experiment_id = exp.run_basic_test_with_temp_config(
config, conf.cv_examples_path("cifar10_tf_keras"), 1)
trials = exp.experiment_trials(experiment_id)
assert len(trials) == 1
def test_pytorch_cifar10_parallel() -> None:
config = conf.load_config(
conf.official_examples_path("trial/cifar10_cnn_pytorch/const.yaml"))
config = conf.set_max_steps(config, 2)
config = conf.set_slots_per_trial(config, 8)
experiment_id = exp.run_basic_test_with_temp_config(
config, conf.official_examples_path("trial/cifar10_cnn_pytorch"), 1)
trials = exp.experiment_trials(experiment_id)
nn = (Determined(conf.make_master_url()).get_trial(
trials[0]["id"]).select_checkpoint(latest=True).load())
assert isinstance(nn, torch.nn.Module)
def test_distributed_logging() -> None:
config = conf.load_config(conf.fixtures_path("pytorch_no_op/const.yaml"))
config = conf.set_slots_per_trial(config, 8)
config = conf.set_max_length(config, {"batches": 1})
e_id = exp.run_basic_test_with_temp_config(
config, conf.fixtures_path("pytorch_no_op"), 1)
t_id = exp.experiment_trials(e_id)[0]["id"]
for i in range(config["resources"]["slots_per_trial"]):
assert exp.check_if_string_present_in_trial_logs(
t_id, "finished train_batch for rank {}".format(i))
def test_tf_keras_native_parallel(tf2: bool) -> None:
config = conf.load_config(
conf.official_examples_path("trial/cifar10_cnn_tf_keras/const.yaml"))
config = conf.set_slots_per_trial(config, 8)
config = conf.set_native_parallel(config, True)
config = conf.set_max_length(config, {"batches": 200})
config = conf.set_tf2_image(config) if tf2 else conf.set_tf1_image(config)
experiment_id = exp.run_basic_test_with_temp_config(
config, conf.official_examples_path("trial/cifar10_cnn_tf_keras"), 1)
trials = exp.experiment_trials(experiment_id)
assert len(trials) == 1
