def local_replica_groups(inner_group_size: int) -> List[List[int]]:
"""Constructs local nearest-neighbor rings given the JAX device assignment.
For inner_group_size=8, each inner group is a tray with replica order:
0/1 2/3
7/6 5/4
Args:
inner_group_size: Number of replica in each group.
Returns:
A list of replica id groups.
"""
world_size = jax.device_count()
outer_group_size, ragged = divmod(world_size, inner_group_size)
assert not ragged, 'inner group size must evenly divide global device count'
# the last device should have maximal x and y coordinate
def bounds_from_last_device(device):
x, y, z = device.coords
return (x + 1) * (device.core_on_chip + 1), (y + 1) * (z + 1)
global_x, _ = bounds_from_last_device(jax.devices()[-1])
per_host_x, per_host_y = bounds_from_last_device(jax.local_devices(0)[-1])
assert inner_group_size in [2 ** i for i in range(1, 15)], \
'inner group size must be a power of two'
if inner_group_size <= 4:
# inner group is Nx1 (core, chip, 2x1)
inner_x, inner_y = inner_group_size, 1
inner_perm = range(inner_group_size)
else:
if inner_group_size <= global_x * 2:
# inner group is Nx2 (2x2 tray, 4x2 DF pod host, row of hosts)
inner_x, inner_y = inner_group_size // 2, 2
else:
# inner group covers the full x dimension and must be >2 in y
inner_x, inner_y = global_x, inner_group_size // global_x
p = np.arange(inner_group_size)
per_group_hosts_x = 1 if inner_x < per_host_x else inner_x // per_host_x
p = p.reshape(inner_y // per_host_y, per_group_hosts_x,
per_host_y, inner_x // per_group_hosts_x)
p = p.transpose(0, 2, 1, 3)
p = p.reshape(inner_y // 2, 2, inner_x)
p[:, 1, :] = p[:, 1, ::-1]
inner_perm = p.reshape(-1)
inner_replica_groups = [[o * inner_group_size + i for i in inner_perm]
for o in range(outer_group_size)]
return inner_replica_groups
def _build_train_input(self):
"""See base class."""
num_devices = jax.device_count()
global_batch_size = self._batch_size
per_device_batch_size, ragged = divmod(global_batch_size, num_devices)
if ragged:
raise ValueError(
f'Global batch size {global_batch_size} must be divisible by '
f'num devices {num_devices}')
return dataset.load(
dataset.Split.TRAIN_AND_VALID,
preprocess_mode=dataset.PreprocessMode.LINEAR_TRAIN,
transpose=self._should_transpose_images(),
batch_dims=[jax.local_device_count(), per_device_batch_size])
def _cross_entropy_loss_fn(self, params, state, images, adv_images, labels,
target_probs, rng):
scalars = {}
images = self.normalize_fn(images)
logits, state = self.model.apply(params,
state,
rng,
images,
is_training=True)
loss = jnp.mean(utils.cross_entropy(logits, target_probs))
loss += self.config.training.weight_decay * utils.weight_decay(params)
if not self.config.training.use_cutmix:
scalars['top_1_acc'] = utils.accuracy(logits, labels)
scalars['train_loss'] = loss
scaled_loss = loss / jax.device_count()
return scaled_loss, (state, scalars)
def _replicate(x, devices=None):
x = jax.numpy.asarray(x)
if devices is None:
# match the default device assignments used in pmap:
# for single-host, that's the XLA default device assignment
# for multi-host, it's the order of jax.local_devices()
if jax.host_count() == 1:
devices = [
d for d in xb.get_backend().get_default_device_assignment(
jax.device_count()) if d.host_id == jax.host_id()
]
else:
devices = jax.local_devices()
aval = jax.ShapedArray((len(devices), ) + x.shape, x.dtype)
buffers = [jax.interpreters.xla.device_put(x, device=d) for d in devices]
return jax.pxla.ShardedDeviceArray(aval, buffers)
def lr_schedule(step: jnp.ndarray) -> jnp.ndarray:
"""Cosine learning rate schedule."""
train_split = dataset.Split.from_string(FLAGS.train_split)
total_batch_size = FLAGS.train_device_batch_size * jax.device_count()
steps_per_epoch = train_split.num_examples / total_batch_size
warmup_steps = FLAGS.train_lr_warmup_epochs * steps_per_epoch
training_steps = FLAGS.train_epochs * steps_per_epoch
lr = FLAGS.train_lr_init * total_batch_size / 256
scaled_step = (jnp.maximum(step - warmup_steps, 0) /
(training_steps - warmup_steps))
lr *= 0.5 * (1.0 + jnp.cos(jnp.pi * scaled_step))
if warmup_steps:
lr *= jnp.minimum(step / warmup_steps, 1.0)
return lr
def _build_train_input(self) -> Generator[dataset.Batch, None, None]:
"""Loads the (infinitely looping) dataset iterator."""
num_devices = jax.device_count()
global_batch_size = self._batch_size
per_device_batch_size, ragged = divmod(global_batch_size, num_devices)
if ragged:
raise ValueError(
f'Global batch size {global_batch_size} must be divisible by '
f'num devices {num_devices}')
return dataset.load(
dataset.Split.TRAIN_AND_VALID,
preprocess_mode=dataset.PreprocessMode.PRETRAIN,
transpose=self._should_transpose_images(),
batch_dims=[jax.local_device_count(), per_device_batch_size])
def testPmap(self):
f = jax.pmap(lambda x: 0. / x)
with self.assertRaisesRegex(
FloatingPointError,
r"invalid value \(nan\) encountered in parallel computation"):
ans = f(jnp.array([0.]))
ans.block_until_ready()
if jax.device_count() >= 2:
with self.assertRaisesRegex(
FloatingPointError,
r"invalid value \(nan\) encountered in parallel computation"
):
ans = f(jnp.array([1., 0.]))
ans.block_until_ready()
def train_loop(experiment_class, config):
"""The main training loop.
This loop periodically saves a checkpoint to be evaluated in the eval_loop.
Args:
experiment_class: the constructor for the experiment (either byol_experiment
or eval_experiment).
config: the experiment config.
"""
experiment = experiment_class(**config)
rng = jax.random.PRNGKey(0)
step = 0
host_id = jax.host_id()
last_logging = time.time()
if config['checkpointing_config']['use_checkpointing']:
checkpoint_data = experiment.load_checkpoint()
if checkpoint_data is None:
step = 0
else:
step, rng = checkpoint_data
local_device_count = jax.local_device_count()
while step < config['max_steps']:
step_rng, rng = tuple(jax.random.split(rng))
# Broadcast the random seeds across the devices
step_rng_device = jax.random.split(step_rng, num=jax.device_count())
step_rng_device = step_rng_device[host_id *
local_device_count:(host_id + 1) *
local_device_count]
step_device = np.broadcast_to(step, [local_device_count])
# Perform a training step and get scalars to log.
scalars = experiment.step(global_step=step_device, rng=step_rng_device)
# Checkpointing and logging.
if config['checkpointing_config']['use_checkpointing']:
experiment.save_checkpoint(step, rng)
current_time = time.time()
if current_time - last_logging > FLAGS.log_tensors_interval:
logging.info('Step %d: %s', step, scalars)
last_logging = current_time
step += 1
logging.info('Saving final checkpoint')
logging.info('Step %d: %s', step, scalars)
experiment.save_checkpoint(step, rng)
def _dataset(load_fn,
is_training: bool,
total_batch_size: int,
ratio: Optional[float] = None,
one_minus_ratio: Optional[float] = None,
repeat: int = 1) -> tf.data.Dataset:
"""Creates a dataset."""
num_devices = jax.device_count()
per_device_batch_size, ragged = divmod(total_batch_size, num_devices)
if ragged:
raise ValueError(
f'Global batch size {total_batch_size} must be divisible by the '
f'total number of devices {num_devices}')
if repeat > 1:
if per_device_batch_size % repeat:
raise ValueError(
f'Per device batch size {per_device_batch_size} must be divisible '
f'by the number of repeated batches {repeat}')
per_device_batch_size //= repeat
if ratio is None and one_minus_ratio is None:
pass # Use full batch size.
elif one_minus_ratio is None:
per_device_batch_size = max(
1,
min(round(per_device_batch_size * ratio),
per_device_batch_size - 1))
elif ratio is None:
batch_size = max(
1,
min(round(per_device_batch_size * one_minus_ratio),
per_device_batch_size - 1))
per_device_batch_size = per_device_batch_size - batch_size
else:
raise ValueError('Only one of `ratio` or `one_minus_ratio` must be '
'specified')
if repeat > 1:
per_device_batch_size *= repeat
# When testing, we need to batch data across all devices (not just local
# devices).
num_local_devices = jax.local_device_count()
if is_training:
batch_sizes = [num_local_devices, per_device_batch_size]
else:
num_hosts = jax.host_count()
assert num_hosts * num_local_devices == num_devices
batch_sizes = [num_hosts, num_local_devices, per_device_batch_size]
return load_fn(batch_sizes, is_training=is_training)
def setUp(self):
super(DistributedTest, self).setUp()
if JAX_MODE:
os.environ['XLA_FLAGS'] = (
'--xla_force_host_platform_device_count={}'.format(NUM_DEVICES))
assert jax.device_count() == NUM_DEVICES
self.key = jax.random.PRNGKey(0)
else:
physical_devices = tf.config.experimental.list_physical_devices()
tf.config.experimental.set_virtual_device_configuration(
physical_devices[0],
[tf.config.experimental.VirtualDeviceConfiguration()] * NUM_DEVICES)
self.strategy = tf.distribute.MirroredStrategy(
devices=tf.config.list_logical_devices())
self.key = [0, 0]
self.axis_name = 'i'
def test_fake_data(self):
model_dir = self.get_tmp_model_dir()
FLAGS.batch_size = 256 * jax.device_count()
FLAGS.half_precision = True
FLAGS.num_epochs = 5
FLAGS.model_dir = model_dir
start_time = time.time()
train.main([])
benchmark_time = time.time() - start_time
self.report_wall_time(benchmark_time)
self.report_extras({
'description': 'ImageNet ResNet50 with fake data',
'model_name': 'resnet50',
'parameters': f'hp=true,bs={FLAGS.batch_size}',
})
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.host_id() == 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.host_id() == 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 = 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,
FLAGS.use_deprecated_checkpointing)
def testBasic(self):
if jax.device_count() < 2:
raise SkipTest
@partial(sharded_jit, in_parts=(P(2, 1), P(2, 1)), out_parts=None)
def f(x, y):
return x + y
shape = (8, 8)
x = np.arange(prod(shape), dtype=np.float32).reshape(shape)
actual = f(x, x + 1)
expected = x + (x + 1)
self.assertAllClose(actual, expected, check_dtypes=False)
self.assertIsInstance(actual, pxla.ShardedDeviceArray)
self.assertLen(actual.device_buffers, 2)
self.assertAllClose(actual.device_buffers[0].to_py(), expected,
check_dtypes=False)
def _build_train_input(self) -> Generator[dataset.Batch, None, None]:
"""See base class."""
num_devices = jax.device_count()
global_batch_size = self.config.training.batch_size
per_device_batch_size, ragged = divmod(global_batch_size, num_devices)
if ragged:
raise ValueError(
f'Global batch size {global_batch_size} must be divisible by '
f'num devices {num_devices}')
split = dataset.Split.TRAIN_AND_VALID
return self._load_data(
split=split,
is_training=True,
batch_dims=[jax.local_device_count(), per_device_batch_size])
def lr_schedule(step: jnp.ndarray) -> jnp.ndarray:
"""Linear scaling rule optimized for 90 epochs."""
train_split = dataset.Split.from_string(FLAGS.train_split)
# See Section 5.1 of https://arxiv.org/pdf/1706.02677.pdf.
total_batch_size = FLAGS.train_device_batch_size * jax.device_count()
steps_per_epoch = train_split.num_examples / total_batch_size
current_epoch = step / steps_per_epoch # type: float
lr = (0.1 * total_batch_size) / 256
lr_linear_till = 5
boundaries = jnp.array((30, 60, 80)) * steps_per_epoch
values = jnp.array([1., 0.1, 0.01, 0.001]) * lr
index = jnp.sum(boundaries < step)
lr = jnp.take(values, index)
return lr * jnp.minimum(1., current_epoch / lr_linear_till)
def build_ensemble_optimizer(ensemble_size, shared_params, ensemble_params,
optimizer, optimizer_kwargs):
num_devices = jax.device_count()
assert ensemble_size % num_devices == 0
num_models_per_device = ensemble_size // num_devices
optim = optimizer(**optimizer_kwargs)
shared_params_optim_state = optim.init(shared_params)
ensemble_params_optim_init = jax.pmap(jax.vmap(optim.init,
in_axes=0,
out_axes=0),
in_axes=0,
out_axes=0)
ensemble_params_optim_state = ensemble_params_optim_init(ensemble_params)
return optim, shared_params_optim_state, ensemble_params_optim_state
def main(unused_argv):
# TODO(gdahl) Figure out a better way to handle passing more complicated
# flags to the binary.
training_metrics_config = None
if FLAGS.training_metrics_config:
training_metrics_config = json.loads(FLAGS.training_metrics_config)
checkpoint_steps = [int(s.strip()) for s in FLAGS.checkpoint_steps]
eval_steps = [int(s.strip()) for s in FLAGS.eval_steps]
if jax.host_id() == 0:
tf.io.gfile.makedirs(FLAGS.experiment_dir)
log_dir = os.path.join(FLAGS.experiment_dir, 'r=3/')
tf.io.gfile.makedirs(log_dir)
log_path = os.path.join(
log_dir, 'worker{}_{}.log'.format(FLAGS.worker_id, jax.host_id()))
with tf.io.gfile.GFile(log_path, 'a') as logfile:
utils.add_log_file(logfile)
if jax.host_id() == 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())
logging.info('checkpoint_steps: %r', checkpoint_steps)
logging.info('eval_steps: %r', eval_steps)
trainer.run(
dataset_name=FLAGS.dataset,
eval_batch_size=FLAGS.eval_batch_size,
eval_num_batches=FLAGS.eval_num_batches,
eval_train_num_batches=FLAGS.eval_train_num_batches,
eval_frequency=FLAGS.eval_frequency,
checkpoint_steps=checkpoint_steps,
eval_steps=eval_steps,
hparam_file=FLAGS.hparam_file,
hparam_overrides=FLAGS.hparam_overrides,
initializer_name=FLAGS.initializer,
model_name=FLAGS.model,
loss_name=FLAGS.loss,
metrics_name=FLAGS.metrics,
num_train_steps=FLAGS.num_train_steps,
experiment_dir=FLAGS.experiment_dir,
worker_id=FLAGS.worker_id,
training_metrics_config=training_metrics_config,
use_deprecated_checkpointing=FLAGS.use_deprecated_checkpointing)
def test_distributed_init(self):
n_devices = jax.device_count()
batch_size = 5 * n_devices
x = np.random.uniform(size=(batch_size, 1))
y = 1.4 * x + 0.1 * np.random.uniform(size=(batch_size, 2))
model = eg.Model(
eg.Linear(2),
loss=[eg.losses.MeanSquaredError()],
)
model = model.distributed()
model.init_on_batch(x)
assert model.module.kernel.shape == (n_devices, 1, 2)
assert model.module.bias.shape == (n_devices, 2)
def from_local(self, model: M) -> M:
# device_idxs used to inform pmap about the number of devices
device_idxs = jnp.arange(jax.device_count())
def device_fn(idx: int, model: M) -> M:
# fold rng state so its unique for each device
return model.map(
lambda key: jax.random.fold_in(key, idx),
tx.Rng,
inplace=True,
)
model = jax.pmap(
device_fn,
in_axes=(0, None),
)(device_idxs, model)
return model
def build_train_input(data_dir, batch_size, img_size, augmentation):
num_devices = jax.device_count()
bs_per_device, ragged = divmod(batch_size, num_devices)
if ragged:
raise ValueError(
f'Batch size {batch_size} must be divisible by num devices {num_devices}'
)
return input_pipeline.load(
input_pipeline.Split.TRAIN_AND_VALID,
data_dir=data_dir,
is_training=True,
batch_dims=[jax.local_device_count(), bs_per_device],
transpose=True,
image_size=(img_size, img_size),
augment_name=augmentation,
augment_before_mix=True,
name='imagenet',
fake_data=False)
def get_gpu_count():
"""
Return the number of available gpus (regardless of whether torch, tf or jax is used)
"""
if is_torch_available():
import torch
return torch.cuda.device_count()
elif is_tf_available():
import tensorflow as tf
return len(tf.config.list_physical_devices("GPU"))
elif is_flax_available():
import jax
return jax.device_count()
else:
return 0
def test_counters(self):
res = unittest.TestResult()
ts = unittest.makeSuite(self.InnerTest) # pytype: disable=module-attr
ts.run(res)
active_pmap = int(jax.device_count() > 1)
self.assertEqual(self.InnerTest.test_1_count, 0)
self.assertEqual(self.InnerTest.test_2_count, 0)
self.assertEqual(self.InnerTest.test_3_count, 4 + active_pmap)
self.assertEqual(self.InnerTest.test_4_count, 4)
self.assertEqual(self.InnerTest.test_5_count, 1)
self.assertEqual(self.InnerTest.test_6_count, 2)
# Test methods do not use `self.variant`.
self.assertLen(res.errors, 1 + 2 + 4 + 4 + active_pmap)
for _, msg in res.errors:
self.assertRegex(
msg, 'RuntimeError: Test is wrapped .+ but never calls self.variant')
def _build_dataset(self, data_rng: spec.RandomState, split: str,
data_dir: str, batch_size):
if batch_size % jax.device_count() > 0:
raise ValueError('Batch size must be divisible by the number of devices')
ds_builder = tfds.builder('imagenet2012:5.*.*')
ds_builder.download_and_prepare()
ds = input_pipeline.create_input_iter(
ds_builder,
batch_size,
self.train_mean,
self.train_stddev,
self.center_crop_size,
self.resize_size,
self.aspect_ratio_range,
self.scale_ratio_range,
train=True,
cache=False)
return ds
def _build_train_input(self):
num_devices = jax.device_count()
global_batch_size = self.config.train_batch_size
bs_per_device, ragged = divmod(global_batch_size, num_devices)
if ragged:
raise ValueError(
f'Global batch size {global_batch_size} must be divisible by '
f'num devices {num_devices}')
return input_pipeline.load(
input_pipeline.Split.TRAIN_AND_VALID,
is_training=True,
batch_dims=[jax.local_device_count(), bs_per_device],
transpose=self.config.get('transpose', False),
image_size=(self.train_imsize, self.train_imsize),
augment_name=self.config.augment_name,
augment_before_mix=self.config.get('augment_before_mix', True),
name=self.config.which_dataset,
fake_data=False)
def main(unused_argv):
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())
if FLAGS.batch_size is None or FLAGS.batch_size <= 0:
raise ValueError("""FLAGS.batch_size value is invalid,
expected a positive non-zero integer.""")
if FLAGS.dataset is None:
raise ValueError("""FLAGS.dataset value is invalid,
expected a non-empty string describing dataset name.""")
batch_size = FLAGS.batch_size
num_batches = FLAGS.num_batches
dataset_name = FLAGS.dataset
model_name = FLAGS.model
initializer_name = 'noop'
hparam_overrides = {
'batch_size': batch_size,
}
hps = hyperparameters.build_hparams(model_name=model_name,
initializer_name=initializer_name,
dataset_name=dataset_name,
hparam_file=None,
hparam_overrides=hparam_overrides)
rng = jax.random.PRNGKey(0)
rng, data_rng = jax.random.split(rng)
dataset = datasets.get_dataset(FLAGS.dataset)(data_rng, batch_size,
batch_size, hps)
train_iter = dataset.train_iterator_fn()
for i in range(num_batches):
batch = next(train_iter)
logging.info('train batch_num = %d, batch = %r', i, batch)
for batch in dataset.valid_epoch(num_batches):
logging.info('validation batch = %r', batch)
def testPyTreeOutputs(self):
if jax.device_count() < 2:
raise SkipTest
def f(x):
return x + 1, ((x + 2, x + 3), x + 4)
shape = (4, 4)
x = np.arange(prod(shape)).reshape(shape)
in_parts = (P(2, 1), )
out_parts = (P(2, 1), ((P(1, 2), None), P(2, 1)))
result = sharded_jit(f, in_parts, out_parts)(x)
expected = f(x)
self.assertAllClose(result, expected, check_dtypes=False)
out_parts = None
result = sharded_jit(f, in_parts, out_parts)(x)
self.assertAllClose(result, expected, check_dtypes=False)
def make_dataset_iterator(
self, replay_client: reverb.Client) -> Iterator[reverb.ReplaySample]:
"""Creates a dataset."""
batch_size_per_learner = self._config.batch_size // jax.process_count()
batch_size_per_device, ragged = divmod(self._config.batch_size,
jax.device_count())
if ragged:
raise ValueError(
'Learner batch size must be divisible by total number of devices!')
dataset = datasets.make_reverb_dataset(
table=self._config.replay_table_name,
server_address=replay_client.server_address,
batch_size=batch_size_per_device,
num_parallel_calls=None,
max_in_flight_samples_per_worker=2 * batch_size_per_learner)
return utils.multi_device_put(dataset.as_numpy_iterator(),
jax.local_devices())
def test_ordered_effect_remains_ordered_across_multiple_devices(self):
# TODO(sharadmv): remove jaxlib check when minimum version is bumped
# TODO(sharadmv): enable this test on GPU and TPU when backends are
# supported
if jaxlib.version < (0, 3, 8):
raise unittest.SkipTest(
"`emit_python_callback` only supported in jaxlib >= 0.3.8")
if jax.device_count() < 2:
raise unittest.SkipTest("Test requires >= 2 devices.")
log = []
def log_value(x):
log.append(x)
return ()
@functools.partial(jax.jit, device=jax.devices()[0])
def f(x):
# Expensive computation
x = x.dot(x)
x = jnp.log(x.sum())
return callback_p.bind(x,
callback=log_value,
effect='log',
out_avals=[])
@functools.partial(jax.jit, device=jax.devices()[1])
def g(x):
return callback_p.bind(x,
callback=log_value,
effect='log',
out_avals=[])
f(jnp.ones((500, 500)))
g(3.)
f(jnp.ones((500, 500)))
g(3.)
f(jnp.ones((500, 500)))
g(3.)
dispatch.runtime_tokens.block_until_ready()
x_, y_ = float(jnp.log(1.25e8)), 3.
expected_log = [x_, y_, x_, y_, x_, y_]
self.assertListEqual(log, expected_log)
def main(argv: Sequence[str]) -> None:
if len(argv) > 1:
raise app.UsageError('Too many command-line arguments.')
# Hide any GPUs form TensorFlow. Otherwise TF might reserve memory and make
# it unavailable to JAX.
tf.config.experimental.set_visible_devices([], 'GPU')
logging.info('JAX process: %d / %d', jax.process_index(),
jax.process_count())
logging.info('JAX local devices: %r', jax.local_devices())
logging.info('JAX total devices: %r', jax.device_count())
# Add a note so that we can tell which task is which JAX host.
# (Depending on the platform task 0 is not guaranteed to be host 0)
platform.work_unit().set_task_status(
f'process_index: {jax.process_index()}, '
f'process_count: {jax.process_count()}')
platform.work_unit().create_artifact(platform.ArtifactType.DIRECTORY,
FLAGS.output_dir, 'output_dir')
tf.io.gfile.makedirs(FLAGS.output_dir)
# Process config here
if FLAGS.config_file:
with tf.io.gfile.GFile(FLAGS.config_file, 'r') as reader:
config = json.load(reader)
else:
config = json.loads(FLAGS.config)
# # Save config to workdir if it's not yet exists
if jax.process_index() == 0:
config_file = os.path.join(FLAGS.output_dir, 'config.json')
with tf.io.gfile.GFile(config_file, 'w') as writer:
writer.write(json.dumps(config, indent=4))
config['output_dir'] = FLAGS.output_dir
if 'num_total_memories' not in config:
config['num_total_memories'] = get_num_total_memories(
ml_collections.ConfigDict(config))
generate(ml_collections.ConfigDict(config))
def _update_func(
self,
base_rng,
state,
inputs,
):
"""Computes loss and updates model parameters."""
step = state.step
rng = jax.random.fold_in(base_rng, jax.lax.axis_index('batch'))
rng = jax.random.fold_in(rng, step)
grad_loss_fn = jax.value_and_grad(self._loss_fn, has_aux=True)
(_, loss_dict), scaled_grads = grad_loss_fn(state.params, inputs, rng)
grads = jax.lax.psum(scaled_grads, axis_name='batch')
grad_norm = optax.global_norm(grads)
loss_dict['scalars']['grad_norm'] = grad_norm
# Compute and apply updates via our optimizer.
learning_rate = self.learning_rate(state.step)
loss_dict['scalars']['learning_rate'] = learning_rate
_, opt_apply = self.optimizer(learning_rate)
updates, new_opt_state = opt_apply(grads, state.opt_state,
state.params)
new_params = optax.apply_updates(state.params, updates)
# Update ema params
ema_rate = self.config.evaluation.ema_rate
new_ema_params = jax.tree_multimap(
lambda x, y: x + (1 - ema_rate) * (y - x),
state.ema_params,
new_params,
)
new_state = state.replace(step=step + 1,
params=new_params,
ema_params=new_ema_params,
opt_state=new_opt_state)
# Rescale loss dict and return
loss_dict['scalars'] = jax.tree_map(
lambda x: jax.lax.psum(x, axis_name='batch') / jax.device_count(),
loss_dict['scalars'],
)
return new_state, loss_dict
