def testImageNetModel(self):
tf.random.set_seed(839382)
dataset_size = 10
batch_size = 6
input_shape = (4, 224, 224, 3)
ensemble_size = 4
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, 4)
dataset = tf.data.Dataset.from_tensor_slices((features, labels))
dataset = dataset.repeat().shuffle(dataset_size).batch(batch_size)
model = imagenet_model.resnet50(input_shape=input_shape,
width_multiplier=2,
num_classes=num_classes,
ensemble_size=ensemble_size)
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)
train_batch_size = (FLAGS.per_core_batch_size * FLAGS.num_cores
// FLAGS.batch_repetitions)
test_batch_size = FLAGS.per_core_batch_size * FLAGS.num_cores
steps_per_epoch = APPROX_IMAGENET_TRAIN_IMAGES // train_batch_size
steps_per_eval = IMAGENET_VALIDATION_IMAGES // test_batch_size
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.TPUStrategy(resolver)
train_builder = ub.datasets.ImageNetDataset(
split=tfds.Split.TRAIN,
use_bfloat16=FLAGS.use_bfloat16)
train_dataset = train_builder.load(
batch_size=train_batch_size, strategy=strategy)
test_builder = ub.datasets.ImageNetDataset(
split=tfds.Split.TEST,
use_bfloat16=FLAGS.use_bfloat16)
test_dataset = test_builder.load(
batch_size=test_batch_size, strategy=strategy)
if FLAGS.use_bfloat16:
tf.keras.mixed_precision.set_global_policy('mixed_bfloat16')
with strategy.scope():
logging.info('Building Keras ResNet-50 model')
model = imagenet_model.resnet50(
input_shape=(FLAGS.ensemble_size, 224, 224, 3),
num_classes=NUM_CLASSES,
ensemble_size=FLAGS.ensemble_size,
width_multiplier=FLAGS.width_multiplier)
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())
# Scale learning rate and decay epochs by vanilla settings.
base_lr = FLAGS.base_learning_rate * train_batch_size / 256
decay_epochs = [
(FLAGS.train_epochs * int(FLAGS.lr_decay_epochs[0])) // 90,
(FLAGS.train_epochs * int(FLAGS.lr_decay_epochs[1])) // 90,
(FLAGS.train_epochs * int(FLAGS.lr_decay_epochs[2])) // 90,
]
learning_rate = ub.schedules.WarmUpPiecewiseConstantSchedule(
steps_per_epoch=steps_per_epoch,
base_learning_rate=base_lr,
decay_ratio=0.1,
decay_epochs=decay_epochs,
warmup_epochs=FLAGS.lr_warmup_epochs)
optimizer = tf.keras.optimizers.SGD(learning_rate=learning_rate,
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': um.ExpectedCalibrationError(num_bins=FLAGS.num_bins),
'test/negative_log_likelihood': tf.keras.metrics.Mean(),
'test/accuracy': tf.keras.metrics.SparseCategoricalAccuracy(),
'test/ece': um.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())
test_diversity = {
'test/disagreement': tf.keras.metrics.Mean(),
'test/average_kl': tf.keras.metrics.Mean(),
'test/cosine_similarity': tf.keras.metrics.Mean(),
}
logging.info('Finished building Keras ResNet-50 model')
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
summary_writer = tf.summary.create_file_writer(
os.path.join(FLAGS.output_dir, 'summaries'))
@tf.function
def train_step(iterator):
"""Training StepFn."""
def step_fn(inputs):
"""Per-Replica StepFn."""
images, labels = inputs
batch_size = tf.shape(images)[0]
main_shuffle = tf.random.shuffle(tf.tile(
tf.range(batch_size), [FLAGS.batch_repetitions]))
to_shuffle = tf.cast(tf.cast(tf.shape(main_shuffle)[0], tf.float32)
* (1. - FLAGS.input_repetition_probability),
tf.int32)
shuffle_indices = [
tf.concat([tf.random.shuffle(main_shuffle[:to_shuffle]),
main_shuffle[to_shuffle:]], axis=0)
for _ in range(FLAGS.ensemble_size)]
images = tf.stack([tf.gather(images, indices, axis=0)
for indices in shuffle_indices], axis=1)
labels = tf.stack([tf.gather(labels, indices, axis=0)
for indices in shuffle_indices], axis=1)
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.reduce_sum(
tf.keras.losses.sparse_categorical_crossentropy(labels,
logits,
from_logits=True),
axis=1))
filtered_variables = []
for var in model.trainable_variables:
# Apply l2 on the weights. This excludes BN parameters and biases, but
# pay caution to their naming scheme.
if 'kernel' in var.name or 'bias' in var.name:
filtered_variables.append(tf.reshape(var, (-1,)))
l2_loss = FLAGS.l2 * 2 * tf.nn.l2_loss(
tf.concat(filtered_variables, axis=0))
# Scale the loss given the TPUStrategy will reduce sum all gradients.
loss = negative_log_likelihood + l2_loss
scaled_loss = loss / strategy.num_replicas_in_sync
grads = tape.gradient(scaled_loss, model.trainable_variables)
optimizer.apply_gradients(zip(grads, model.trainable_variables))
probs = tf.nn.softmax(tf.reshape(logits, [-1, NUM_CLASSES]))
flat_labels = tf.reshape(labels, [-1])
metrics['train/ece'].update_state(flat_labels, probs)
metrics['train/loss'].update_state(loss)
metrics['train/negative_log_likelihood'].update_state(
negative_log_likelihood)
metrics['train/accuracy'].update_state(flat_labels, probs)
strategy.run(step_fn, args=(next(iterator),))
@tf.function
def test_step(iterator):
"""Evaluation StepFn."""
def step_fn(inputs):
"""Per-Replica StepFn."""
images, labels = inputs
images = tf.tile(
tf.expand_dims(images, 1), [1, 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.transpose(probs, perm=[1, 0, 2])
diversity_results = um.average_pairwise_diversity(
per_probs, FLAGS.ensemble_size)
for k, v in diversity_results.items():
test_diversity['test/' + k].update_state(v)
for i in range(FLAGS.ensemble_size):
member_probs = 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)
# Negative log marginal likelihood computed in a numerically-stable way.
labels_tiled = tf.tile(
tf.expand_dims(labels, 1), [1, FLAGS.ensemble_size])
log_likelihoods = -tf.keras.losses.sparse_categorical_crossentropy(
labels_tiled, logits, from_logits=True)
negative_log_likelihood = tf.reduce_mean(
-tf.reduce_logsumexp(log_likelihoods, axis=[1]) +
tf.math.log(float(FLAGS.ensemble_size)))
probs = tf.math.reduce_mean(probs, axis=1) # marginalize
metrics['test/negative_log_likelihood'].update_state(
negative_log_likelihood)
metrics['test/accuracy'].update_state(labels, probs)
metrics['test/ece'].update_state(labels, probs)
strategy.run(step_fn, args=(next(iterator),))
metrics.update({'test/ms_per_example': tf.keras.metrics.Mean()})
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)
test_iterator = iter(test_dataset)
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_start_time = time.time()
test_step(test_iterator)
ms_per_example = (time.time() - test_start_time) * 1e6 / test_batch_size
metrics['test/ms_per_example'].update_state(ms_per_example)
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_metrics = metrics.copy()
total_metrics.update(test_diversity)
total_results = {name: metric.result()
for name, metric in total_metrics.items()}
with summary_writer.as_default():
for name, result in total_results.items():
tf.summary.scalar(name, result, step=epoch + 1)
for _, metric in total_metrics.items():
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_save_name = os.path.join(FLAGS.output_dir, 'model')
model.save(final_save_name)
logging.info('Saved model to %s', final_save_name)