def make_Kuf(k, X, a, b, ms):
omegas = 2. * np.pi * ms / (b - a)
if float_type is tf.float32:
omegas = omegas.astype(np.float32)
Kuf_cos = tf.transpose(tf.cos(omegas * (X - a)))
omegas_sin = omegas[omegas != 0] # don't compute zeros freq.
Kuf_sin = tf.transpose(tf.sin(omegas_sin * (X - a)))
# correct Kfu outside [a, b]
lt_a_sin = tf.tile(tf.transpose(X) < a, [len(ms) - 1, 1])
gt_b_sin = tf.tile(tf.transpose(X) > b, [len(ms) - 1, 1])
lt_a_cos = tf.tile(tf.transpose(X) < a, [len(ms), 1])
gt_b_cos = tf.tile(tf.transpose(X) > b, [len(ms), 1])
if isinstance(k, GPflow.kernels.Matern12):
# Kuf_sin[:, np.logical_or(X.flatten() < a, X.flatten() > b)] = 0
Kuf_sin = tf.where(tf.logical_or(lt_a_sin, gt_b_sin),
tf.zeros(tf.shape(Kuf_sin), float_type), Kuf_sin)
Kuf_cos = tf.where(
lt_a_cos,
tf.tile(tf.exp(-tf.abs(tf.transpose(X - a)) / k.lengthscales),
[len(ms), 1]), Kuf_cos)
Kuf_cos = tf.where(
gt_b_cos,
tf.tile(tf.exp(-tf.abs(tf.transpose(X - b)) / k.lengthscales),
[len(ms), 1]), Kuf_cos)
elif isinstance(k, GPflow.kernels.Matern32):
arg = np.sqrt(3) * tf.abs(tf.transpose(X) - a) / k.lengthscales
edge = tf.tile((1 + arg) * tf.exp(-arg), [len(ms), 1])
Kuf_cos = tf.where(lt_a_cos, edge, Kuf_cos)
arg = np.sqrt(3) * tf.abs(tf.transpose(X) - b) / k.lengthscales
edge = tf.tile((1 + arg) * tf.exp(-arg), [len(ms), 1])
Kuf_cos = tf.where(gt_b_cos, edge, Kuf_cos)
arg = np.sqrt(3) * tf.abs(tf.transpose(X) - a) / k.lengthscales
edge = (tf.transpose(X) - a) * tf.exp(-arg) * omegas_sin[:, None]
Kuf_sin = tf.where(lt_a_sin, edge, Kuf_sin)
arg = np.sqrt(3) * tf.abs(tf.transpose(X) - b) / k.lengthscales
edge = (tf.transpose(X) - b) * tf.exp(-arg) * omegas_sin[:, None]
Kuf_sin = tf.where(gt_b_sin, edge, Kuf_sin)
elif isinstance(k, GPflow.kernels.Matern52):
# edges not implemented yet
Kuf_cos = tf.with_dependencies(
[tf.assert_greater_equal(X, a)],
Kuf_cos,
message='Edges not implemented for Matern52',
name='assert_left_edge')
Kuf_sin = tf.with_dependencies(
[tf.assert_less_equal(X, b)],
Kuf_sin,
message='Edges not implemented for Matern52',
name='assert_right_edge')
else:
raise NotImplementedError
return tf.concat([Kuf_cos, Kuf_sin], axis=0)
def diagonal_lstm(inputs, conf, scope='diagonal_lstm'):
with tf.variable_scope(scope):
skewed_inputs = skew(inputs, scope="skewed_i")
# input-to-state (K_is * x_i) : 1x1 convolution. generate 4h x n x n tensor.
input_to_state = conv2d(skewed_inputs, conf.hidden_dims * 4, [1, 1], "B", scope="i_to_s")
column_wise_inputs = tf.transpose(
input_to_state, [2, 0, 1, 3]) # [width, batch, height, hidden_dims * 4]
if conf.log_level == 'DEBUG':
logger.warning("[assert] check equal of skew and unskew")
unskewed_inputs = unskew(skewed_inputs, scope="skewed_i")
skew_assert_op = tf.Assert(tf.equal(inputs, unskewed_inputs, 'skew_check'), [unskewed_inputs])
input_to_state = tf.with_dependencies([skew_assert_op], input_to_state)
width, batch, height, channel = get_shape(column_wise_inputs)
rnn_inputs = tf.reshape(column_wise_inputs,
[-1, width, height * channel]) # [batch, max_time, height * hidden_dims * 4]
rnn_input_list = [tf.squeeze(rnn_input, squeeze_dims=[1])
for rnn_input in tf.split(split_dim=1, num_split=width, value=rnn_inputs)]
cell = DiagonalLSTMCell(conf.hidden_dims, height, channel)
if conf.use_dynamic_rnn:
outputs, states = tf.nn.dynamic_rnn(cell,
inputs=rnn_inputs, dtype=tf.float32) # [batch, width, height * hidden_dims]
else:
output_list, state_list = tf.nn.rnn(cell,
inputs=rnn_input_list, dtype=tf.float32) # width * [batch, height * hidden_dims]
packed_outputs = tf.pack(output_list, 1) # [batch, width, height * hidden_dims]
width_first_outputs = tf.reshape(packed_outputs,
[-1, width, height, conf.hidden_dims]) # [batch, width, height, hidden_dims]
skewed_outputs = tf.transpose(width_first_outputs, [0, 2, 1, 3])
outputs = unskew(skewed_outputs)
return outputs
def create_train_op(self,
total_loss,
optimizer,
global_step=None,
update_ops=None,
variables_to_train=None,
clip_by_global_norm=False,
gradient_noise_scale=None,
gradient_multipliers=None,
gate_gradients=tf.train.Optimizer.GATE_OP,
aggregation_method=None,
colocate_gradients_with_ops=False):
"""Creates an `Operation` that evaluates the gradients and returns the loss.
Args:
total_loss: A `Tensor` representing the total loss.
optimizer: A tf.Optimizer to use for computing the gradients.
global_step: A `Tensor` representing the global step variable. If left as
`_USE_GLOBAL_STEP`, then tf.contrib.framework.global_step() is used.
update_ops: An optional list of updates to execute. If `update_ops` is
`None`, then the update ops are set to the contents of the
`tf.GraphKeys.UPDATE_OPS` collection. If `update_ops` is not `None`, but
it doesn't contain all of the update ops in `tf.GraphKeys.UPDATE_OPS`,
a warning will be displayed.
variables_to_train: an optional list of variables to train. If None, it will
default to all tf.trainable_variables().
clip_grad_global_norm: A bool, performs gradient clipping using global norm if True
else performs gradient clipping using local norm.
gradient_noise_scale: if not None, add noises to the gradients
gradient_multipliers: if not None, a dict, multiples gradient with given args
gate_gradients: How to gate the computation of gradients. See tf.Optimizer.
aggregation_method: Specifies the method used to combine gradient terms.
Valid values are defined in the class `AggregationMethod`.
colocate_gradients_with_ops: Whether or not to try colocating the gradients
with the ops that generated them.
Returns:
A `Tensor` that when evaluated, computes the gradients and returns the total
loss value.
"""
if global_step is None:
global_step = tf.get_variable('global_step',
shape=[],
dtype=tf.int64,
initializer=tf.zeros_initializer,
trainable=False)
# Update ops use GraphKeys.UPDATE_OPS collection if update_ops is None.
global_update_ops = set(tf.get_collection(tf.GraphKeys.UPDATE_OPS))
if update_ops is None:
update_ops = global_update_ops
else:
update_ops = set(update_ops)
if not global_update_ops.issubset(update_ops):
log.warn(
'update_ops in create_train_op does not contain all the update_ops in GraphKeys.UPDATE_OPS'
)
# Make sure update_ops are computed before total_loss.
if update_ops:
with tf.control_dependencies(update_ops):
barrier = tf.no_op(name='update_barrier')
total_loss = tf.with_dependencies([barrier], total_loss)
if variables_to_train is None:
variables_to_train = tf.trainable_variables()
else:
for v in variables_to_train:
assert v in tf.trainable_variables()
assert variables_to_train
if clip_by_global_norm:
grads_and_vars = self._clip_grad_global_norms(
variables_to_train,
total_loss,
optimizer,
global_norm=8,
gate_gradients=gate_gradients,
gradient_noise_scale=gradient_noise_scale,
GATE_GRAPH=2,
grad_loss=None,
agre_method=aggregation_method,
col_grad_ops=colocate_gradients_with_ops)
else:
grads_and_vars = optimizer.compute_gradients(
total_loss,
variables_to_train,
gate_gradients=gate_gradients,
aggregation_method=aggregation_method,
colocate_gradients_with_ops=colocate_gradients_with_ops)
grads_and_vars = self._clip_grad_norms(grads_and_vars, max_norm=8)
if gradient_multipliers is not None:
grads_and_vars = self._multiply_gradients(grads_and_vars,
gradient_multipliers)
grad_updates = optimizer.apply_gradients(grads_and_vars,
global_step=global_step)
with tf.name_scope('train_op'):
# Make sure total_loss is valid.
total_loss = tf.check_numerics(total_loss,
'LossTensor is inf or nan')
# Ensure the train_tensor computes grad_updates.
with tf.control_dependencies([grad_updates]):
total_loss = tf.identity(total_loss)
return total_loss
def build_graph(reader,
model,
train_data_pattern,
label_loss_fn=losses.CrossEntropyLoss(),
batch_size=1000,
base_learning_rate=0.01,
optimizer_class=tf.train.AdamOptimizer,
regularization_penalty=1e-3,
num_readers=1,
num_epochs=None):
"""Creates the Tensorflow graph.
This will only be called once in the life of
a training model, because after the graph is created the model will be
restored from a meta graph file rather than being recreated.
Args:
reader: The data file reader. It should inherit from BaseReader.
model: The core model (e.g. logistic or neural net). It should inherit
from BaseModel.
train_data_pattern: glob path to the training data files.
label_loss_fn: What kind of loss to apply to the model. It should inherit
from BaseLoss.
batch_size: How many examples to process at a time.
base_learning_rate: What learning rate to initialize the optimizer with.
optimizer_class: Which optimization algorithm to use.
regularization_penalty: How much weight to give the regularization loss
compared to the label loss.
num_readers: How many threads to use for I/O operations.
num_epochs: How many passes to make over the data. 'None' means an
unlimited number of passes.
"""
with tf.device(
tf.train.replica_device_setter(FLAGS.ps_tasks,
merge_devices=True)):
global_step = tf.Variable(0, trainable=False, name="global_step")
optimizer = optimizer_class(base_learning_rate)
unused_video_id, model_input_raw, labels_batch, num_frames = (
get_input_data_tensors(reader,
train_data_pattern,
batch_size=batch_size,
num_readers=num_readers,
num_epochs=num_epochs))
tf.summary.histogram("model/input_raw", model_input_raw)
feature_dim = len(model_input_raw.get_shape()) - 1
model_input = tf.nn.l2_normalize(model_input_raw, feature_dim)
with tf.name_scope("model"):
result = model.create_model(model_input,
num_frames=num_frames,
vocab_size=reader.num_classes,
labels=labels_batch)
for variable in slim.get_model_variables():
tf.summary.histogram(variable.op.name, variable)
predictions = result["predictions"]
tf.summary.histogram("model_activations", predictions)
if "loss" in result.keys():
label_loss = result["loss"]
else:
label_loss = label_loss_fn.calculate_loss(
predictions, labels_batch)
tf.summary.scalar("label_loss", label_loss)
if "regularization_loss" in result.keys():
reg_loss = result["regularization_loss"]
else:
reg_loss = tf.constant(0.0)
if regularization_penalty != 0:
tf.summary.scalar("reg_loss", reg_loss)
# Adds update_ops (e.g., moving average updates in batch normalization) as
# a dependency to the train_op.
update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
if "update_ops" in result.keys():
update_ops += result["update_ops"]
if update_ops:
with tf.control_dependencies(update_ops):
barrier = tf.no_op(name="gradient_barrier")
label_loss = tf.with_dependencies([barrier], label_loss)
# Incorporate the L2 weight penalties etc.
final_loss = regularization_penalty * reg_loss + label_loss
train_op = optimizer.minimize(final_loss, global_step=global_step)
tf.add_to_collection("global_step", global_step)
tf.add_to_collection("loss", label_loss)
tf.add_to_collection("predictions", predictions)
tf.add_to_collection("input_batch_raw", model_input_raw)
tf.add_to_collection("input_batch", model_input)
tf.add_to_collection("num_frames", num_frames)
tf.add_to_collection("labels", tf.cast(labels_batch, tf.float32))
tf.add_to_collection("train_op", train_op)
def main(_):
ps_hosts = FLAGS.ps_hosts.split(",")
worker_hosts = FLAGS.worker_hosts.split(",")
# Create a cluster from the parameter server and worker hosts.
cluster = tf.train.ClusterSpec({"ps": ps_hosts, "worker": worker_hosts})
# Create and start a server for the local task.
server = tf.train.Server(cluster,
job_name=FLAGS.job_name,
task_index=FLAGS.task_index)
if FLAGS.job_name == "ps":
server.join()
elif FLAGS.job_name == "worker":
# Assigns ops to the local worker by default.
with tf.device(
tf.train.replica_device_setter(
worker_device="/job:worker/task:%d" % FLAGS.task_index,
cluster=cluster)):
# set Keras learning phase to train
keras.backend.set_learning_phase(1)
# do not initialize variables on the fly
keras.backend.manual_variable_initialization(True)
# Build Keras model
model = ...
# keras model predictions
preds = model.output
# placeholder for training targets
targets = tf.placeholder(...)
# our categorical crossentropy loss
xent_loss = tf.reduce_mean(
keras.objectives.categorical_crossentropy(targets, preds))
# we create a global_step tensor for distributed training
# (a counter of iterations)
global_step = tf.Variable(0, name='global_step', trainable=False)
# apply regularizers if any
if model.regularizers:
total_loss = xent_loss * 1. # copy tensor
for regularizer in model.regularizers:
total_loss = regularizer(total_loss)
else:
total_loss = xent_loss
# set up TF optimizer
optimizer = tf.train.RMSPropOptimizer(learning_rate,
decay=0.9,
momentum=FLAGS.momentum,
epsilon=1e-8)
# Set up model update ops (batch norm ops).
# The gradients should only be computed after updating the moving average
# of the batch normalization parameters, in order to prevent a data race
# between the parameter updates and moving average computations.
with tf.control_dependencies(model.updates):
barrier = tf.no_op(name='update_barrier')
# define gradient updates
with tf.control_dependencies([barrier]):
grads = optimizer.compute_gradients(
total_loss,
model.trainable_weights,
gate_gradients=tf.Optimizer.GATE_OP,
aggregation_method=None,
colocate_gradients_with_ops=False)
# define train tensor
train_tensor = tf.with_dependencies([grad_updates],
total_loss,
name='train')
# blah blah
saver = tf.train.Saver()
summary_op = tf.merge_all_summaries()
init_op = tf.initialize_all_variables()
# Create a "supervisor", which oversees the training process.
sv = tf.train.Supervisor(is_chief=(FLAGS.task_index == 0),
logdir="/tmp/train_logs",
init_op=init_op,
summary_op=summary_op,
saver=saver,
global_step=global_step,
save_model_secs=600)
# The supervisor takes care of session initialization, restoring from
# a checkpoint, and closing when done or an error occurs.
with sv.managed_session(server.target) as sess:
# Loop until the supervisor shuts down or 1000000 steps have completed.
step = 0
while not sv.should_stop() and step < 1000000:
# Run a training step asynchronously.
# See `tf.train.SyncReplicasOptimizer` for additional details on how to
# perform *synchronous* training.
# feed_dict must contain the model inputs (the tensors listed in model.inputs)
# and the "targets" placeholder we created ealier
# it's a dictionary mapping tensors to batches of Numpy data
# like: feed_dict={model.inputs[0]: np_train_data_batch, targets: np_train_labels_batch}
loss_value, step_value = sess.run([train_op, global_step],
feed_dict={...})
# Ask for all the services to stop.
sv.stop()
