def main(unused_argv):
# Necessary to use the tfds loader.
tf.enable_v2_behavior()
if jax.process_count() > 1:
# TODO(ankugarg): Add support for multihost inference.
raise NotImplementedError(
'BLEU eval does not support multihost inference.')
rng = jax.random.PRNGKey(FLAGS.seed)
mt_eval_config = json.loads(FLAGS.mt_eval_config)
if FLAGS.experiment_config_filename:
with tf.io.gfile.GFile(FLAGS.experiment_config_filename) as f:
experiment_config = json.load(f)
if jax.process_index() == 0:
logging.info('experiment_config: %r', experiment_config)
dataset_name = experiment_config['dataset']
model_name = experiment_config['model']
else:
assert FLAGS.dataset and FLAGS.model
dataset_name = FLAGS.dataset
model_name = FLAGS.model
if jax.process_index() == 0:
logging.info('argv:\n%s', ' '.join(sys.argv))
logging.info('device_count: %d', jax.device_count())
logging.info('num_hosts : %d', jax.host_count())
logging.info('host_id : %d', jax.host_id())
model_class = models.get_model(model_name)
dataset_builder = datasets.get_dataset(dataset_name)
dataset_meta_data = datasets.get_dataset_meta_data(dataset_name)
hparam_overrides = None
if FLAGS.hparam_overrides:
if isinstance(FLAGS.hparam_overrides, str):
hparam_overrides = json.loads(FLAGS.hparam_overrides)
merged_hps = hyperparameters.build_hparams(
model_name=model_name,
initializer_name=experiment_config['initializer'],
dataset_name=dataset_name,
hparam_file=FLAGS.trial_hparams_filename,
hparam_overrides=hparam_overrides)
if jax.process_index() == 0:
logging.info('Merged hps are: %s', json.dumps(merged_hps.to_json()))
evaluator = bleu_evaluator.BLEUEvaluator(FLAGS.checkpoint_dir, merged_hps,
rng, model_class, dataset_builder,
dataset_meta_data, mt_eval_config)
evaluator.translate_and_calculate_bleu()
def main(unused_argv):
# Necessary to use the tfds imagenet loader.
tf.enable_v2_behavior()
rng = jax.random.PRNGKey(FLAGS.seed)
if FLAGS.hessian_eval_config:
hessian_eval_config = json.loads(FLAGS.hessian_eval_config)
else:
hessian_eval_config = hessian_eval.DEFAULT_EVAL_CONFIG
if FLAGS.experiment_config_filename:
with tf.io.gfile.GFile(FLAGS.experiment_config_filename, 'r') as f:
experiment_config = json.load(f)
if jax.process_index() == 0:
logging.info('experiment_config: %r', experiment_config)
dataset_name = experiment_config['dataset']
model_name = experiment_config['model']
else:
assert FLAGS.dataset and FLAGS.model
dataset_name = FLAGS.dataset
model_name = FLAGS.model
if jax.process_index() == 0:
logging.info('argv:\n%s', ' '.join(sys.argv))
logging.info('device_count: %d', jax.device_count())
logging.info('num_hosts : %d', jax.process_count())
logging.info('host_id : %d', jax.process_index())
model = models.get_model(model_name)
dataset_builder = datasets.get_dataset(dataset_name)
dataset_meta_data = datasets.get_dataset_meta_data(dataset_name)
with tf.io.gfile.GFile(FLAGS.trial_hparams_filename, 'r') as f:
hps = config_dict.ConfigDict(json.load(f))
if FLAGS.hparam_overrides:
if isinstance(FLAGS.hparam_overrides, str):
hparam_overrides = json.loads(FLAGS.hparam_overrides)
hps.update_from_flattened_dict(hparam_overrides)
run_lanczos.eval_checkpoints(
FLAGS.checkpoint_dir,
hps,
rng,
FLAGS.eval_num_batches,
model,
dataset_builder,
dataset_meta_data,
hessian_eval_config,
FLAGS.min_global_step,
FLAGS.max_global_step)
def _run(train_fn, dataset_name, eval_batch_size, eval_num_batches,
eval_train_num_batches, eval_frequency, checkpoint_steps,
num_tf_data_prefetches, num_device_prefetches,
num_tf_data_map_parallel_calls, early_stopping_target_name,
early_stopping_target_value, early_stopping_mode, eval_steps,
hparam_file, hparam_overrides, initializer_name, model_name,
loss_name, metrics_name, num_train_steps, experiment_dir, worker_id,
training_metrics_config, callback_configs, external_checkpoint_path):
"""Function that runs a Jax experiment. See flag definitions for args."""
model_cls = models.get_model(model_name)
initializer = initializers.get_initializer(initializer_name)
dataset_builder = datasets.get_dataset(dataset_name)
dataset_meta_data = datasets.get_dataset_meta_data(dataset_name)
input_pipeline_hps = config_dict.ConfigDict(
dict(
num_tf_data_prefetches=num_tf_data_prefetches,
num_device_prefetches=num_device_prefetches,
num_tf_data_map_parallel_calls=num_tf_data_map_parallel_calls,
))
merged_hps = hyperparameters.build_hparams(
model_name=model_name,
initializer_name=initializer_name,
dataset_name=dataset_name,
hparam_file=hparam_file,
hparam_overrides=hparam_overrides,
input_pipeline_hps=input_pipeline_hps)
# Note that one should never tune an RNG seed!!! The seed is only included in
# the hparams for convenience of running hparam trials with multiple seeds per
# point.
rng_seed = merged_hps.rng_seed
if merged_hps.rng_seed < 0:
rng_seed = _create_synchronized_rng_seed()
xm_experiment = None
xm_work_unit = None
if jax.process_index() == 0:
logging.info('Running with seed %d', rng_seed)
rng = jax.random.PRNGKey(rng_seed)
# Build the loss_fn, metrics_bundle, and flax_module.
model = model_cls(merged_hps, dataset_meta_data, loss_name, metrics_name)
trial_dir = os.path.join(experiment_dir, str(worker_id))
meta_data_path = os.path.join(trial_dir, 'meta_data.json')
meta_data = {'worker_id': worker_id, 'status': 'incomplete'}
if jax.process_index() == 0:
logging.info('rng: %s', rng)
gfile.makedirs(trial_dir)
# Set up the metric loggers for host 0.
metrics_logger, init_logger = utils.set_up_loggers(
trial_dir, xm_work_unit)
hparams_fname = os.path.join(trial_dir, 'hparams.json')
logging.info('saving hparams to %s', hparams_fname)
with gfile.GFile(hparams_fname, 'w') as f:
f.write(merged_hps.to_json())
_write_trial_meta_data(meta_data_path, meta_data)
else:
metrics_logger = None
init_logger = None
try:
epoch_reports = list(
train_fn(trial_dir,
model,
dataset_builder,
initializer,
num_train_steps,
merged_hps,
rng,
eval_batch_size,
eval_num_batches,
eval_train_num_batches,
eval_frequency,
checkpoint_steps,
early_stopping_target_name,
early_stopping_target_value,
early_stopping_mode,
eval_steps,
metrics_logger,
init_logger,
training_metrics_config=training_metrics_config,
callback_configs=callback_configs,
external_checkpoint_path=external_checkpoint_path))
logging.info(epoch_reports)
meta_data['status'] = 'done'
except utils.TrainingDivergedError as err:
meta_data['status'] = 'diverged'
raise err
finally:
if jax.process_index() == 0:
_write_trial_meta_data(meta_data_path, meta_data)
def test_early_stopping(self):
"""Test training early stopping on MNIST with a small model."""
rng = jax.random.PRNGKey(0)
# Set the numpy seed to make the fake data deterministc. mocking.mock_data
# ultimately calls numpy.random.
np.random.seed(0)
model_name = 'fully_connected'
loss_name = 'cross_entropy'
metrics_name = 'classification_metrics'
initializer_name = 'noop'
dataset_name = 'mnist'
model_cls = models.get_model(model_name)
initializer = initializers.get_initializer(initializer_name)
dataset_builder = datasets.get_dataset(dataset_name)
hparam_overrides = {
'lr_hparams': {
'base_lr': 0.1,
'schedule': 'cosine'
},
'batch_size': 8,
'train_size': 160,
'valid_size': 96,
'test_size': 80,
}
input_pipeline_hps = config_dict.ConfigDict(
dict(
num_tf_data_prefetches=-1,
num_device_prefetches=0,
num_tf_data_map_parallel_calls=-1,
))
hps = hyperparameters.build_hparams(
model_name,
initializer_name,
dataset_name,
hparam_file=None,
hparam_overrides=hparam_overrides,
input_pipeline_hps=input_pipeline_hps)
eval_batch_size = 16
num_examples = 256
def as_dataset(self, *args, **kwargs):
del args
del kwargs
# pylint: disable=g-long-lambda,g-complex-comprehension
return tf.data.Dataset.from_generator(
lambda: ({
'image': np.ones(shape=(28, 28, 1), dtype=np.uint8),
'label': 9,
} for i in range(num_examples)),
output_types=self.info.features.dtype,
output_shapes=self.info.features.shape,
)
# This will override the tfds.load(mnist) call to return 100 fake samples.
with tfds.testing.mock_data(as_dataset_fn=as_dataset,
num_examples=num_examples):
dataset = dataset_builder(shuffle_rng=jax.random.PRNGKey(0),
batch_size=hps.batch_size,
eval_batch_size=eval_batch_size,
hps=hps)
model = model_cls(hps, datasets.get_dataset_meta_data(dataset_name),
loss_name, metrics_name)
num_train_steps = 40
early_stopping_target_name = 'test/ce_loss'
early_stopping_target_value = 0.005
early_stopping_mode = 'less'
eval_num_batches = 5
eval_every = 10
checkpoint_steps = [1, 3, 15]
metrics_logger, init_logger = utils.set_up_loggers(self.test_dir)
epoch_reports = list(
trainer.train(
train_dir=self.test_dir,
model=model,
dataset_builder=lambda *unused_args, **unused_kwargs: dataset,
initializer=initializer,
num_train_steps=num_train_steps,
hps=hps,
rng=rng,
eval_batch_size=eval_batch_size,
eval_num_batches=eval_num_batches,
eval_train_num_batches=eval_num_batches,
eval_frequency=eval_every,
checkpoint_steps=checkpoint_steps,
early_stopping_target_name=early_stopping_target_name,
early_stopping_target_value=early_stopping_target_value,
early_stopping_mode=early_stopping_mode,
metrics_logger=metrics_logger,
init_logger=init_logger))
self.assertLen(epoch_reports, 3)
self.assertGreater(epoch_reports[-2][early_stopping_target_name],
early_stopping_target_value)
self.assertLess(epoch_reports[-1][early_stopping_target_name],
early_stopping_target_value)
def test_trainer(self):
"""Test training for two epochs on MNIST with a small model."""
rng = jax.random.PRNGKey(0)
# Set the numpy seed to make the fake data deterministc. mocking.mock_data
# ultimately calls numpy.random.
np.random.seed(0)
model_name = 'fully_connected'
loss_name = 'cross_entropy'
metrics_name = 'classification_metrics'
initializer_name = 'noop'
dataset_name = 'mnist'
model_cls = models.get_model(model_name)
initializer = initializers.get_initializer(initializer_name)
dataset_builder = datasets.get_dataset(dataset_name)
hparam_overrides = {
'lr_hparams': {
'base_lr': 0.1,
'schedule': 'cosine'
},
'batch_size': 8,
'train_size': 160,
'valid_size': 96,
'test_size': 80,
}
input_pipeline_hps = config_dict.ConfigDict(
dict(
num_tf_data_prefetches=-1,
num_device_prefetches=0,
num_tf_data_map_parallel_calls=-1,
))
hps = hyperparameters.build_hparams(
model_name,
initializer_name,
dataset_name,
hparam_file=None,
hparam_overrides=hparam_overrides,
input_pipeline_hps=input_pipeline_hps)
eval_batch_size = 16
num_examples = 256
def as_dataset(self, *args, **kwargs):
del args
del kwargs
# pylint: disable=g-long-lambda,g-complex-comprehension
return tf.data.Dataset.from_generator(
lambda: ({
'image': np.ones(shape=(28, 28, 1), dtype=np.uint8),
'label': 9,
} for i in range(num_examples)),
output_types=self.info.features.dtype,
output_shapes=self.info.features.shape,
)
# This will override the tfds.load(mnist) call to return 100 fake samples.
with tfds.testing.mock_data(as_dataset_fn=as_dataset,
num_examples=num_examples):
dataset = dataset_builder(shuffle_rng=jax.random.PRNGKey(0),
batch_size=hps.batch_size,
eval_batch_size=eval_batch_size,
hps=hps)
model = model_cls(hps, datasets.get_dataset_meta_data(dataset_name),
loss_name, metrics_name)
num_train_steps = 40
eval_num_batches = 5
eval_every = 10
checkpoint_steps = [1, 3, 15]
metrics_logger, init_logger = utils.set_up_loggers(self.test_dir)
epoch_reports = list(
trainer.train(
train_dir=self.test_dir,
model=model,
dataset_builder=lambda *unused_args, **unused_kwargs: dataset,
initializer=initializer,
num_train_steps=num_train_steps,
hps=hps,
rng=rng,
eval_batch_size=eval_batch_size,
eval_num_batches=eval_num_batches,
eval_train_num_batches=eval_num_batches,
eval_frequency=eval_every,
checkpoint_steps=checkpoint_steps,
metrics_logger=metrics_logger,
init_logger=init_logger))
# check that the additional checkpoints are saved.
checkpoint_dir = os.path.join(self.test_dir, 'checkpoints')
saved_steps = []
for f in tf.io.gfile.listdir(checkpoint_dir):
if f[:5] == 'ckpt_':
saved_steps.append(int(f[5:]))
self.assertEqual(set(saved_steps), set(checkpoint_steps))
self.assertLen(epoch_reports, num_train_steps / eval_every)
with tf.io.gfile.GFile(os.path.join(self.test_dir,
'measurements.csv')) as f:
df = pandas.read_csv(f)
train_err = df['train/error_rate'].values[-1]
self.assertEqual(df['preemption_count'].values[-1], 0)
self.assertLess(train_err, 0.9)
self.assertEqual(set(df.columns.values), set(get_column_names()))
model = model_cls(hps, {'apply_one_hot_in_loss': False}, loss_name,
metrics_name)
# Test reload from the checkpoint by increasing num_train_steps.
num_train_steps_reload = 100
epoch_reports = list(
trainer.train(
train_dir=self.test_dir,
model=model,
dataset_builder=lambda *unused_args, **unused_kwargs: dataset,
initializer=initializer,
num_train_steps=num_train_steps_reload,
hps=hps,
rng=rng,
eval_batch_size=eval_batch_size,
eval_num_batches=eval_num_batches,
eval_train_num_batches=eval_num_batches,
eval_frequency=eval_every,
checkpoint_steps=checkpoint_steps,
metrics_logger=metrics_logger,
init_logger=init_logger))
self.assertLen(epoch_reports,
(num_train_steps_reload - num_train_steps) / eval_every)
with tf.io.gfile.GFile(os.path.join(self.test_dir,
'measurements.csv')) as f:
df = pandas.read_csv(f)
train_err = df['train/error_rate'].values[-1]
train_loss = df['train/ce_loss'].values[-1]
self.assertLess(train_err, 0.35)
self.assertLess(train_loss, 0.1)
self.assertEqual(df['valid/num_examples'].values[-1],
eval_num_batches * eval_batch_size)
self.assertEqual(df['preemption_count'].values[-1], 1)
# Check that the correct learning rate was saved in the measurements file.
final_learning_rate = df['learning_rate'].values[-1]
final_step = df['global_step'].values[-1]
self.assertEqual(num_train_steps_reload, final_step)
# final_step will be one larger than the last step used to calculate the
# lr_decay, hense we plug in (final_step - 1) to the decay formula.
# Note that there is a small numerical different here with np vs jnp.
decay_factor = (1 + np.cos(
(final_step - 1) / num_train_steps_reload * np.pi)) * 0.5
self.assertEqual(float(final_learning_rate),
hps.lr_hparams['base_lr'] * decay_factor)
self.assertEqual(set(df.columns.values), set(get_column_names()))
def test_run_lanczos(self):
"""Test training for two epochs on MNIST with a small model."""
rng = jax.random.PRNGKey(0)
# Set the numpy seed to make the fake data deterministc. mocking.mock_data
# ultimately calls numpy.random.
np.random.seed(0)
model_name = 'fully_connected'
loss_name = 'cross_entropy'
metrics_name = 'classification_metrics'
initializer_name = 'noop'
dataset_name = 'mnist'
model_cls = models.get_model(model_name)
initializer = initializers.get_initializer(initializer_name)
dataset_builder = datasets.get_dataset(dataset_name)
hparam_overrides = {
'lr_hparams': {
'base_lr': 0.1,
'schedule': 'cosine'
},
'batch_size': 8,
'train_size': 160,
'valid_size': 96,
'test_size': 80,
}
input_pipeline_hps = config_dict.ConfigDict(dict(
num_tf_data_prefetches=-1,
num_device_prefetches=0,
num_tf_data_map_parallel_calls=-1,
))
hps = hyperparameters.build_hparams(
model_name,
initializer_name,
dataset_name,
hparam_file=None,
hparam_overrides=hparam_overrides,
input_pipeline_hps=input_pipeline_hps)
model = model_cls(hps, datasets.get_dataset_meta_data(dataset_name),
loss_name, metrics_name)
eval_batch_size = 16
num_examples = 256
def as_dataset(self, *args, **kwargs):
del args
del kwargs
# pylint: disable=g-long-lambda,g-complex-comprehension
return tf.data.Dataset.from_generator(
lambda: ({
'image': np.ones(shape=(28, 28, 1), dtype=np.uint8),
'label': 9,
} for i in range(num_examples)),
output_types=self.info.features.dtype,
output_shapes=self.info.features.shape,
)
# This will override the tfds.load(mnist) call to return 100 fake samples.
with tfds.testing.mock_data(
as_dataset_fn=as_dataset, num_examples=num_examples):
dataset = dataset_builder(
shuffle_rng=jax.random.PRNGKey(0),
batch_size=hps.batch_size,
eval_batch_size=eval_batch_size,
hps=hps)
num_train_steps = 41
eval_num_batches = 5
eval_every = 10
checkpoint_steps = [10, 30, 40]
metrics_logger, init_logger = None, None
_ = list(
trainer.train(
train_dir=self.test_dir,
model=model,
dataset_builder=lambda *unused_args, **unused_kwargs: dataset,
initializer=initializer,
num_train_steps=num_train_steps,
hps=hps,
rng=rng,
eval_batch_size=eval_batch_size,
eval_num_batches=eval_num_batches,
eval_train_num_batches=eval_num_batches,
eval_frequency=eval_every,
checkpoint_steps=checkpoint_steps,
metrics_logger=metrics_logger,
init_logger=init_logger))
checkpoint_dir = os.path.join(self.test_dir, 'checkpoints')
rng = jax.random.PRNGKey(0)
run_lanczos.eval_checkpoints(
checkpoint_dir,
hps,
rng,
eval_num_batches,
model_cls=model_cls,
dataset_builder=lambda *unused_args, **unused_kwargs: dataset,
dataset_meta_data=datasets.get_dataset_meta_data(dataset_name),
hessian_eval_config=hessian_eval.DEFAULT_EVAL_CONFIG,
)
# Load the saved file.
hessian_dir = os.path.join(checkpoint_dir, 'hessian', 'training_metrics')
pytree_list = checkpoint.load_pytree(hessian_dir)
# Convert to a regular list (checkpointer will have converted the saved
# list to a dict of keys '0', '1', ...
pytree_list = [pytree_list[str(i)] for i in range(len(pytree_list))]
# Test that the logged steps are correct.
saved_steps = [row['step'] for row in pytree_list]
self.assertEqual(saved_steps, checkpoint_steps)
def test_shampoo_wrn(self):
"""Test distributed shampoo on fake dataset."""
model_name = 'simple_cnn'
model_cls = models.get_model(model_name)
hparam_overrides = {
'optimizer': 'distributed_shampoo',
'batch_size': 1,
'train_size': 10,
'valid_size': 10,
'input_shape': (32, 32, 3),
'output_shape': (10,),
'opt_hparams': {
'block_size': 32,
'beta1': 0.9,
'beta2': 0.999,
'diagonal_epsilon': 1e-10,
'matrix_epsilon': 1e-6,
'weight_decay': 0.0,
'start_preconditioning_step': 5,
'preconditioning_compute_steps': 1,
'statistics_compute_steps': 1,
'best_effort_shape_interpretation': True,
'graft_type': distributed_shampoo.GraftingType.SGD,
'nesterov': True,
'exponent_override': 0,
'batch_axis_name': 'batch',
'num_devices_for_pjit': None,
'shard_optimizer_states': False,
'inverse_failure_threshold': 0.1,
'clip_by_scaled_gradient_norm': None,
'precision': lax.Precision.HIGHEST,
'moving_average_for_momentum': False,
'skip_preconditioning_dim_size_gt': 4096,
'best_effort_memory_usage_reduction': False,
},
}
input_pipeline_hps = config_dict.ConfigDict(dict(
num_tf_data_prefetches=-1,
num_device_prefetches=0,
num_tf_data_map_parallel_calls=-1,
))
hps = hyperparameters.build_hparams(
model_name,
initializer_name='noop',
dataset_name='fake',
hparam_file=None,
hparam_overrides=hparam_overrides,
input_pipeline_hps=input_pipeline_hps)
initializer = initializers.get_initializer('noop')
dataset_builder = datasets.get_dataset('fake')
dataset = dataset_builder(
shuffle_rng=jax.random.PRNGKey(0),
batch_size=hps.batch_size,
eval_batch_size=hps.batch_size,
hps=hps)
loss_name = 'cross_entropy'
metrics_name = 'classification_metrics'
dataset_meta_data = datasets.get_dataset_meta_data('fake')
model = model_cls(hps, dataset_meta_data, loss_name, metrics_name)
metrics_logger, init_logger = utils.set_up_loggers(self.test_dir)
_ = list(
trainer.train(
train_dir=self.test_dir,
model=model,
dataset_builder=lambda *unused_args, **unused_kwargs: dataset,
initializer=initializer,
num_train_steps=1,
hps=hps,
rng=jax.random.PRNGKey(42),
eval_batch_size=hps.batch_size,
eval_num_batches=None,
eval_train_num_batches=None,
eval_frequency=10,
checkpoint_steps=[],
metrics_logger=metrics_logger,
init_logger=init_logger))
with tf.io.gfile.GFile(os.path.join(self.test_dir,
'measurements.csv')) as f:
df = pandas.read_csv(f)
valid_ce_loss = df['valid/ce_loss'].values[-1]
self.assertLess(valid_ce_loss, 1e-3)