def _grads_and_vars_barrier(
grads_and_vars
):
"""Barrier that forces all grads to be computed before any are used."""
current_grads, current_vars = list(zip(*grads_and_vars))
current_grads = layers.with_data_dependencies(current_grads, current_grads)
return list(zip(current_grads, current_vars))
def model_fn(features, labels, mode, params):
"""Construct a TPUEstimatorSpec for a model."""
if mode != tf.estimator.ModeKeys.TRAIN:
raise NotImplementedError(
'Expected that mode == TRAIN, but got {:!r}'.format(mode))
# Data was transposed from NHWC to HWCN on the host side. Transpose it back.
# This transposition will be optimized away by the XLA compiler. It serves
# as a hint to the compiler that it should expect the input data to come
# in HWCN format rather than NHWC.
train_features = tf.transpose(features['train'], [3, 0, 1, 2])
validation_features = tf.transpose(features['validation'], [3, 0, 1, 2])
if params['use_bfloat16'] == 'ontpu':
train_features = tf.cast(train_features, tf.bfloat16)
validation_features = tf.cast(validation_features, tf.bfloat16)
global_step = tf.train.get_global_step()
# Randomly sample a network architecture.
with tf.variable_scope('rl_controller') as rl_scope:
pass
model_spec = mobile_classifier_factory.get_model_spec(params['ssd'])
tf.io.gfile.makedirs(params['checkpoint_dir'])
model_spec_filename = os.path.join(params['checkpoint_dir'],
'model_spec.json')
with tf.io.gfile.GFile(model_spec_filename, 'w') as handle:
handle.write(schema_io.serialize(model_spec))
increase_ops_prob = custom_layers.linear_decay(
global_step, params['increase_ops_warmup_steps'])
increase_filters_prob = custom_layers.linear_decay(
global_step, params['increase_filters_warmup_steps'])
model_spec, dist_info = controller.independent_sample(
model_spec,
increase_ops_probability=increase_ops_prob,
increase_filters_probability=increase_filters_prob,
name=rl_scope)
if params['enable_cost_model']:
cost_model_features = mobile_cost_model.coupled_tf_features(model_spec)
estimated_cost = cost_model_lib.estimate_cost(cost_model_features,
params['ssd'])
# We divide the regularization strength by 2 for backwards compatibility with
# the deprecated tf.contrib.layers.l2_regularizer() function, which was used
# in our published experiments.
kernel_regularizer = tf.keras.regularizers.l2(
params['model_weight_decay'] / 2)
# Set up the basic TensorFlow training/inference graph.
model = mobile_classifier_factory.get_model_for_search(
model_spec, kernel_regularizer=kernel_regularizer)
model.build(train_features.shape)
with tf.name_scope('training'):
model_logits, _ = model.apply(train_features, training=True)
# Cast back to float32 (effectively only when using use_bfloat16 is true).
model_logits = tf.cast(model_logits, tf.float32)
model_empirical_loss = tf.losses.softmax_cross_entropy(
onehot_labels=labels['train'],
logits=model_logits,
label_smoothing=0.1)
model_regularization_loss = model.regularization_loss()
model_loss = model_empirical_loss + model_regularization_loss
# Set up the model weight training logic.
model_learning_rate = custom_layers.cosine_decay_with_linear_warmup(
peak_learning_rate=params['model_learning_rate'],
global_step=global_step,
max_global_step=params['max_global_step'],
warmup_steps=params['model_warmup_steps'])
model_optimizer = tf.tpu.CrossShardOptimizer(
tf.train.RMSPropOptimizer(model_learning_rate,
decay=0.9,
momentum=params['model_momentum'],
epsilon=1.0))
model_vars = model.trainable_variables()
model_update_ops = model.updates()
with tf.control_dependencies(model_update_ops):
grads_and_vars = model_optimizer.compute_gradients(
model_loss, var_list=model_vars)
if params['use_gradient_sync_barrier']:
# Force all gradients to be computed before any are applied.
grads_and_vars = _grads_and_vars_barrier(grads_and_vars)
# NOTE: We do not pass `global_step` to apply_gradients(), so the global
# step is not incremented by `model_optimizer`. The global_step will be
# incremented later on, when we update the RL controller weights. If we
# incremented it here too, we'd end up incrementing the global_step twice
# at each training step.
model_op = model_optimizer.apply_gradients(grads_and_vars)
if params['use_gradient_sync_barrier']:
# Finish computing gradients for the shared model weights before we
# start on the RL update step.
#
# NOTE: The barrier above forces TensorFlow to finish computing grads
# for all of the trainable variables before any of the grads can be
# consumed. So while the call to with_data_dependencies() here only
# explicitly depends on grads_and_vars[0][0], the call implicitly forces
# TensorFlow to finish computing the gradients for *all* trainable
# variables before computing the validation features.
validation_features = layers.with_data_dependencies(
[grads_and_vars[0][0]], [validation_features])[0]
with tf.name_scope('validation'):
# Estimate the model accuracy on a batch of examples from the validation
# set. Force this logic to run after the model optimization step.
with tf.control_dependencies([model_op]):
validation_logits, _ = model.apply(validation_features,
training=False)
# NOTE(b/130311965): An earlier version of this code cast validation_logits
# from bfloat16 to float32 before applying an argmax when the --use_bfloat16
# flag was true. As of cl/240923609, this caused XLA to compute incorrect
# model accuracies. Please avoid casting from bfloat16 to bfloat32 before
# taking the argmax.
is_prediction_correct = tf.equal(
tf.argmax(validation_logits, axis=1),
tf.argmax(labels['validation'], axis=1))
validation_accuracy = tf.reduce_mean(
tf.cast(is_prediction_correct, tf.float32))
# Estimate the reward for the current network architecture and update the
# reward to incorporate the cost of the network architecture.
if params['enable_cost_model']:
rl_stats = search_space_utils.reward_for_single_cost_model(
validation_accuracy,
rl_reward_function=params['rl_reward_function'],
estimated_cost=estimated_cost,
rl_cost_model_target=params['rl_cost_model_target'],
rl_cost_model_exponent=params['rl_cost_model_exponent'])
rl_cost_ratio = rl_stats['rl_cost_ratio']
rl_reward = rl_stats['rl_reward']
rl_cost_adjustment = rl_stats['rl_cost_adjustment']
else:
rl_reward = validation_accuracy
# Compute a baseline. We first take a cross-replica sum of the rewards
# for all the TPU shards, then incorporate the result into an exponential
# moving average. Within a single batch, each TPU shard will select a
# different set of op masks from the RL controller. Each shard will basically
# evaluate a different candidate architecture in our search space.
# Count the number of TPU shards (cores) used for training.
num_tpu_shards = tf.tpu.cross_replica_sum(
tf.ones(shape=(), dtype=rl_reward.dtype))
rl_step_baseline = tf.tpu.cross_replica_sum(rl_reward)
rl_step_baseline = rl_step_baseline / num_tpu_shards
rl_baseline = custom_layers.update_exponential_moving_average(
rl_step_baseline, momentum=params['rl_baseline_momentum'])
# Apply a REINFORCE update to the RL controller.
log_prob = dist_info['sample_log_prob']
rl_advantage = rl_reward - rl_baseline
rl_empirical_loss = -tf.stop_gradient(rl_advantage) * log_prob
# We set rl_entropy_loss proportional to (-entropy) so that minimizing the
# loss will lead to an entropy that is as large as possible.
rl_entropy = dist_info['entropy']
rl_entropy_loss = -params['rl_entropy_regularization'] * rl_entropy
# We use an RL learning rate of 0 for the first N epochs of training. See
# Appendix A of FBNet. (https://arxiv.org/pdf/1812.03443.pdf). Although they
# don't mention it explicitly, there are some indications that ProxylessNAS
# (https://openreview.net/forum?id=HylVB3AqYm) might also be doing this.
enable_rl_optimizer = tf.cast(
tf.greater_equal(global_step, params['rl_delay_steps']), tf.float32)
rl_learning_rate = params['rl_learning_rate'] * enable_rl_optimizer
if params['use_exponential_rl_learning_rate_schedule']:
# rl_learning_rate_progress will be 0 when the RL controller starts
# learning and 1 when the search ends.
rl_learning_rate_progress = tf.nn.relu(
tf.div(
tf.cast(global_step - params['rl_delay_steps'], tf.float32),
max(1, params['max_global_step'] - params['rl_delay_steps'])))
# exponentially increase the RL learning rate over time.
rl_learning_rate_multiplier = tf.pow(10.0, rl_learning_rate_progress)
rl_learning_rate = rl_learning_rate * rl_learning_rate_multiplier
rl_update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS, rl_scope.name)
with tf.control_dependencies(rl_update_ops):
# In order to evaluate train_op, we must first evaluate validation_accuracy.
# And to evaluate validation_accuracy, we must first evaluate model_op. So
# running this op will perform a step of model training followed by
# a step of RL controller training.
if params['use_gradient_sync_barrier']:
transform_grads_fn = _grads_and_vars_barrier
else:
transform_grads_fn = None
train_op = tpu_optimizer_ops.apply_adam(
rl_empirical_loss,
regularization_loss=rl_entropy_loss,
global_step=global_step,
var_list=tf.trainable_variables(rl_scope.name),
learning_rate=rl_learning_rate,
beta1=0.0,
beta2=0.999,
epsilon=1e-8,
transform_grads_fn=transform_grads_fn)
# TensorBoard logging
tensorboard_scalars = collections.OrderedDict([
('model/loss', model_loss),
('model/empirical_loss', model_empirical_loss),
('model/regularization_loss', model_regularization_loss),
('model/learning_rate', model_learning_rate),
('rlcontroller/empirical_loss', rl_empirical_loss),
('rlcontroller/entropy_loss', rl_entropy_loss),
('rlcontroller/validation_accuracy', validation_accuracy),
('rlcontroller/reward', rl_reward),
('rlcontroller/step_baseline', rl_step_baseline),
('rlcontroller/baseline', rl_baseline),
('rlcontroller/advantage', rl_advantage),
('rlcontroller/log_prob', log_prob),
])
if params['enable_cost_model']:
tensorboard_scalars['rlcontroller/estimated_cost'] = estimated_cost
tensorboard_scalars['rlcontroller/cost_ratio'] = rl_cost_ratio
tensorboard_scalars[
'rlcontroller/cost_adjustment'] = rl_cost_adjustment
tensorboard_scalars['rlcontroller/learning_rate'] = rl_learning_rate
tensorboard_scalars['rlcontroller/increase_ops_prob'] = increase_ops_prob
tensorboard_scalars['rlcontroller/increase_filters_prob'] = (
increase_filters_prob)
# Log the values of all the choices made by the RL controller.
for name_i, logits_i in dist_info['logits_by_path'].items():
assert len(logits_i.shape) == 1, logits_i
for j in range(int(logits_i.shape[0])):
key = 'rlpathlogits/{:s}/{:d}'.format(name_i, j)
tensorboard_scalars[key] = logits_i[j]
for name_i, logits_i in dist_info['logits_by_tag'].items():
assert len(logits_i.shape) == 1, logits_i
for j in range(int(logits_i.shape[0])):
key = 'rltaglogits/{:s}/{:d}'.format(name_i, j)
tensorboard_scalars[key] = logits_i[j]
# NOTE: host_call only works on rank-1 tensors. There's also a fairly
# large performance penalty if we try to pass too many distinct tensors
# from the TPU to the host at once. We avoid these problems by (i) calling
# tf.stack to merge all of the float32 scalar values into a single rank-1
# tensor that can be sent to the host relatively cheaply and (ii) reshaping
# the remaining values from scalars to rank-1 tensors.
def host_call_fn(step, scalar_values):
values = tf.unstack(scalar_values)
with tf2.summary.create_file_writer(
params['checkpoint_dir']).as_default():
with tf2.summary.record_if(
tf.math.equal(step[0] % params['tpu_iterations_per_loop'],
0)):
for key, value in zip(list(tensorboard_scalars.keys()),
values):
tf2.summary.scalar(key, value, step=step[0])
return tf.summary.all_v2_summary_ops()
host_call_values = tf.stack(list(tensorboard_scalars.values()))
host_call = (host_call_fn,
[tf.reshape(global_step, [1]), host_call_values])
# Construct the estimator specification.
return tf.estimator.tpu.TPUEstimatorSpec(mode=mode,
loss=model_loss,
train_op=train_op,
host_call=host_call)