def testWideResnet(self):
tf.random.set_seed(83922)
dataset_size = 10
batch_size = 4 # must be divisible by ensemble_size
input_shape = (32, 32, 1)
num_classes = 2
features = tf.random.normal((dataset_size,) + input_shape)
coeffs = tf.random.normal([tf.reduce_prod(input_shape), num_classes])
net = tf.reshape(features, [dataset_size, -1])
logits = tf.matmul(net, coeffs)
labels = tf.random.categorical(logits, 1)
dataset = tf.data.Dataset.from_tensor_slices((features, labels))
dataset = dataset.repeat().shuffle(dataset_size).batch(batch_size)
model = batchensemble_model.wide_resnet(input_shape=input_shape,
depth=10,
width_multiplier=1,
num_classes=num_classes,
ensemble_size=2,
random_sign_init=-0.5,
l2=0.)
model.compile(
'adam',
loss=tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True))
history = model.fit(dataset,
steps_per_epoch=dataset_size // batch_size,
epochs=2)
loss_history = history.history['loss']
self.assertAllGreaterEqual(loss_history, 0.)
def main(argv):
del argv # unused arg
tf.io.gfile.makedirs(FLAGS.output_dir)
logging.info('Saving checkpoints at %s', FLAGS.output_dir)
tf.random.set_seed(FLAGS.seed)
ds_info = tfds.builder(FLAGS.dataset).info
per_core_batch_size = FLAGS.per_core_batch_size // FLAGS.ensemble_size
batch_size = per_core_batch_size * FLAGS.num_cores
# Train_proportion is a float so need to convert steps_per_epoch to int.
steps_per_epoch = int(
(ds_info.splits['train'].num_examples * FLAGS.train_proportion) //
batch_size)
steps_per_eval = ds_info.splits['test'].num_examples // batch_size
num_classes = ds_info.features['label'].num_classes
if FLAGS.use_gpu:
logging.info('Use GPU')
strategy = tf.distribute.MirroredStrategy()
else:
logging.info('Use TPU at %s',
FLAGS.tpu if FLAGS.tpu is not None else 'local')
resolver = tf.distribute.cluster_resolver.TPUClusterResolver(
tpu=FLAGS.tpu)
tf.config.experimental_connect_to_cluster(resolver)
tf.tpu.experimental.initialize_tpu_system(resolver)
strategy = tf.distribute.experimental.TPUStrategy(resolver)
train_input_fn = utils.load_input_fn(split=tfds.Split.TRAIN,
name=FLAGS.dataset,
batch_size=per_core_batch_size,
use_bfloat16=FLAGS.use_bfloat16,
proportion=FLAGS.train_proportion)
clean_test_input_fn = utils.load_input_fn(split=tfds.Split.TEST,
name=FLAGS.dataset,
batch_size=per_core_batch_size,
use_bfloat16=FLAGS.use_bfloat16)
train_dataset = strategy.experimental_distribute_datasets_from_function(
train_input_fn)
test_datasets = {
'clean':
strategy.experimental_distribute_datasets_from_function(
clean_test_input_fn),
}
if FLAGS.corruptions_interval > 0:
if FLAGS.dataset == 'cifar10':
load_c_input_fn = utils.load_cifar10_c_input_fn
else:
load_c_input_fn = functools.partial(utils.load_cifar100_c_input_fn,
path=FLAGS.cifar100_c_path)
corruption_types, max_intensity = utils.load_corrupted_test_info(
FLAGS.dataset)
for corruption in corruption_types:
for intensity in range(1, max_intensity + 1):
input_fn = load_c_input_fn(corruption_name=corruption,
corruption_intensity=intensity,
batch_size=per_core_batch_size,
use_bfloat16=FLAGS.use_bfloat16)
test_datasets['{0}_{1}'.format(corruption, intensity)] = (
strategy.experimental_distribute_datasets_from_function(
input_fn))
if FLAGS.use_bfloat16:
policy = tf.keras.mixed_precision.experimental.Policy('mixed_bfloat16')
tf.keras.mixed_precision.experimental.set_policy(policy)
summary_writer = tf.summary.create_file_writer(
os.path.join(FLAGS.output_dir, 'summaries'))
with strategy.scope():
logging.info('Building Keras model')
model = batchensemble_model.wide_resnet(
input_shape=ds_info.features['image'].shape,
depth=28,
width_multiplier=10,
num_classes=num_classes,
ensemble_size=FLAGS.ensemble_size,
random_sign_init=FLAGS.random_sign_init,
l2=FLAGS.l2)
logging.info('Model input shape: %s', model.input_shape)
logging.info('Model output shape: %s', model.output_shape)
logging.info('Model number of weights: %s', model.count_params())
# Linearly scale learning rate and the decay epochs by vanilla settings.
base_lr = FLAGS.base_learning_rate * batch_size / 128
lr_decay_epochs = [(int(start_epoch_str) * FLAGS.train_epochs) // 200
for start_epoch_str in FLAGS.lr_decay_epochs]
lr_schedule = utils.LearningRateSchedule(
steps_per_epoch,
base_lr,
decay_ratio=FLAGS.lr_decay_ratio,
decay_epochs=lr_decay_epochs,
warmup_epochs=FLAGS.lr_warmup_epochs)
optimizer = tf.keras.optimizers.SGD(lr_schedule,
momentum=0.9,
nesterov=True)
metrics = {
'train/negative_log_likelihood':
tf.keras.metrics.Mean(),
'train/accuracy':
tf.keras.metrics.SparseCategoricalAccuracy(),
'train/loss':
tf.keras.metrics.Mean(),
'train/ece':
ed.metrics.ExpectedCalibrationError(num_bins=FLAGS.num_bins),
'test/negative_log_likelihood':
tf.keras.metrics.Mean(),
'test/accuracy':
tf.keras.metrics.SparseCategoricalAccuracy(),
'test/ece':
ed.metrics.ExpectedCalibrationError(num_bins=FLAGS.num_bins),
}
for i in range(FLAGS.ensemble_size):
metrics['test/nll_member_{}'.format(i)] = tf.keras.metrics.Mean()
metrics['test/accuracy_member_{}'.format(i)] = (
tf.keras.metrics.SparseCategoricalAccuracy())
if FLAGS.corruptions_interval > 0:
corrupt_metrics = {}
for intensity in range(1, max_intensity + 1):
for corruption in corruption_types:
dataset_name = '{0}_{1}'.format(corruption, intensity)
corrupt_metrics['test/nll_{}'.format(dataset_name)] = (
tf.keras.metrics.Mean())
corrupt_metrics['test/accuracy_{}'.format(
dataset_name)] = (
tf.keras.metrics.SparseCategoricalAccuracy())
corrupt_metrics['test/ece_{}'.format(dataset_name)] = (
ed.metrics.ExpectedCalibrationError(
num_bins=FLAGS.num_bins))
checkpoint = tf.train.Checkpoint(model=model, optimizer=optimizer)
latest_checkpoint = tf.train.latest_checkpoint(FLAGS.output_dir)
initial_epoch = 0
if latest_checkpoint:
# checkpoint.restore must be within a strategy.scope() so that optimizer
# slot variables are mirrored.
checkpoint.restore(latest_checkpoint)
logging.info('Loaded checkpoint %s', latest_checkpoint)
initial_epoch = optimizer.iterations.numpy() // steps_per_epoch
@tf.function
def train_step(iterator):
"""Training StepFn."""
def step_fn(inputs):
"""Per-Replica StepFn."""
images, labels = inputs
images = tf.tile(images, [FLAGS.ensemble_size, 1, 1, 1])
labels = tf.tile(labels, [FLAGS.ensemble_size])
with tf.GradientTape() as tape:
logits = model(images, training=True)
if FLAGS.use_bfloat16:
logits = tf.cast(logits, tf.float32)
negative_log_likelihood = tf.reduce_mean(
tf.keras.losses.sparse_categorical_crossentropy(
labels, logits, from_logits=True))
l2_loss = sum(model.losses)
loss = negative_log_likelihood + l2_loss
# Scale the loss given the TPUStrategy will reduce sum all gradients.
scaled_loss = loss / strategy.num_replicas_in_sync
grads = tape.gradient(scaled_loss, model.trainable_variables)
# Separate learning rate implementation.
if FLAGS.fast_weight_lr_multiplier != 1.0:
grads_and_vars = []
for grad, var in zip(grads, model.trainable_variables):
# Apply different learning rate on the fast weight approximate
# posterior/prior parameters. This is excludes BN and slow weights,
# but pay caution to the naming scheme.
if ('batch_norm' not in var.name
and 'kernel' not in var.name):
grads_and_vars.append(
(grad * FLAGS.fast_weight_lr_multiplier, var))
else:
grads_and_vars.append((grad, var))
optimizer.apply_gradients(grads_and_vars)
else:
optimizer.apply_gradients(zip(grads,
model.trainable_variables))
probs = tf.nn.softmax(logits)
metrics['train/ece'].update_state(labels, probs)
metrics['train/loss'].update_state(loss)
metrics['train/negative_log_likelihood'].update_state(
negative_log_likelihood)
metrics['train/accuracy'].update_state(labels, logits)
strategy.run(step_fn, args=(next(iterator), ))
@tf.function
def test_step(iterator, dataset_name):
"""Evaluation StepFn."""
def step_fn(inputs):
"""Per-Replica StepFn."""
images, labels = inputs
images = tf.tile(images, [FLAGS.ensemble_size, 1, 1, 1])
logits = model(images, training=False)
if FLAGS.use_bfloat16:
logits = tf.cast(logits, tf.float32)
probs = tf.nn.softmax(logits)
per_probs = tf.split(probs,
num_or_size_splits=FLAGS.ensemble_size,
axis=0)
for i in range(FLAGS.ensemble_size):
member_probs = per_probs[i]
member_loss = tf.keras.losses.sparse_categorical_crossentropy(
labels, member_probs)
metrics['test/nll_member_{}'.format(i)].update_state(
member_loss)
metrics['test/accuracy_member_{}'.format(i)].update_state(
labels, member_probs)
probs = tf.reduce_mean(per_probs, axis=0)
negative_log_likelihood = tf.reduce_mean(
tf.keras.losses.sparse_categorical_crossentropy(labels, probs))
if dataset_name == 'clean':
metrics['test/negative_log_likelihood'].update_state(
negative_log_likelihood)
metrics['test/accuracy'].update_state(labels, probs)
metrics['test/ece'].update_state(labels, probs)
else:
corrupt_metrics['test/nll_{}'.format(
dataset_name)].update_state(negative_log_likelihood)
corrupt_metrics['test/accuracy_{}'.format(
dataset_name)].update_state(labels, probs)
corrupt_metrics['test/ece_{}'.format(
dataset_name)].update_state(labels, probs)
strategy.run(step_fn, args=(next(iterator), ))
train_iterator = iter(train_dataset)
start_time = time.time()
for epoch in range(initial_epoch, FLAGS.train_epochs):
logging.info('Starting to run epoch: %s', epoch)
for step in range(steps_per_epoch):
train_step(train_iterator)
current_step = epoch * steps_per_epoch + (step + 1)
max_steps = steps_per_epoch * FLAGS.train_epochs
time_elapsed = time.time() - start_time
steps_per_sec = float(current_step) / time_elapsed
eta_seconds = (max_steps - current_step) / steps_per_sec
message = ('{:.1%} completion: epoch {:d}/{:d}. {:.1f} steps/s. '
'ETA: {:.0f} min. Time elapsed: {:.0f} min'.format(
current_step / max_steps, epoch + 1,
FLAGS.train_epochs, steps_per_sec, eta_seconds / 60,
time_elapsed / 60))
if step % 20 == 0:
logging.info(message)
datasets_to_evaluate = {'clean': test_datasets['clean']}
if (FLAGS.corruptions_interval > 0
and (epoch + 1) % FLAGS.corruptions_interval == 0):
datasets_to_evaluate = test_datasets
for dataset_name, test_dataset in datasets_to_evaluate.items():
test_iterator = iter(test_dataset)
logging.info('Testing on dataset %s', dataset_name)
for step in range(steps_per_eval):
if step % 20 == 0:
logging.info('Starting to run eval step %s of epoch: %s',
step, epoch)
test_step(test_iterator, dataset_name)
logging.info('Done with testing on %s', dataset_name)
corrupt_results = {}
if (FLAGS.corruptions_interval > 0
and (epoch + 1) % FLAGS.corruptions_interval == 0):
corrupt_results = utils.aggregate_corrupt_metrics(
corrupt_metrics, corruption_types, max_intensity)
logging.info('Train Loss: %.4f, Accuracy: %.2f%%',
metrics['train/loss'].result(),
metrics['train/accuracy'].result() * 100)
logging.info('Test NLL: %.4f, Accuracy: %.2f%%',
metrics['test/negative_log_likelihood'].result(),
metrics['test/accuracy'].result() * 100)
for i in range(FLAGS.ensemble_size):
logging.info(
'Member %d Test Loss: %.4f, Accuracy: %.2f%%', i,
metrics['test/nll_member_{}'.format(i)].result(),
metrics['test/accuracy_member_{}'.format(i)].result() * 100)
total_results = {
name: metric.result()
for name, metric in metrics.items()
}
total_results.update(corrupt_results)
with summary_writer.as_default():
for name, result in total_results.items():
tf.summary.scalar(name, result, step=epoch + 1)
for metric in metrics.values():
metric.reset_states()
if (FLAGS.checkpoint_interval > 0
and (epoch + 1) % FLAGS.checkpoint_interval == 0):
checkpoint_name = checkpoint.save(
os.path.join(FLAGS.output_dir, 'checkpoint'))
logging.info('Saved checkpoint to %s', checkpoint_name)
final_checkpoint_name = checkpoint.save(
os.path.join(FLAGS.output_dir, 'checkpoint'))
logging.info('Saved last checkpoint to %s', final_checkpoint_name)