def __init__(self, conf, tasksconf, dataconf, modelconf, evaluatorconf,
expdir, init_filename, server, task_index):
'''
NnetTrainer constructor, creates the training graph
Args:
conf: the trainer config
taskconf: the config file for each task
dataconf: the data configuration as a ConfigParser
modelconf: the neural net model configuration
evaluatorconf: the evaluator configuration for evaluating
if None no evaluation will be done
expdir: directory where the summaries will be written
init_filename: filename of the network that should be used to
initialize the model. Put to None if no network is available/wanted.
server: optional server to be used for distributed training
task_index: optional index of the worker task in the cluster
'''
self.expdir = expdir
self.server = server
self.conf = conf
self.tasksconf = tasksconf
self.task_index = task_index
self.init_filename = init_filename
self.batch_size = int(conf['batch_size'])
cluster = tf.train.ClusterSpec(server.server_def.cluster)
#create the graph
self.graph = tf.Graph()
#3 model types for multi task: single one to one; single one to many; multiple one to one
#single one to one: the whole model is shared for all tasks, only loss function can be different
#single one to many: each task has a separate output so only part of the network is shared, eg evrything but the output layer
#multiple one to one: each task has its own network. Possibly the outputs are combined in a loss function
#create the model
modelfile = os.path.join(expdir, 'model', 'model.pkl')
with open(modelfile, 'wb') as fid:
self.model = model_factory.factory(
modelconf.get('model', 'architecture'))(conf=modelconf)
pickle.dump(self.model, fid)
evaltype = evaluatorconf.get('evaluator', 'evaluator')
#get the database configurations
input_dataconfs = dict()
target_dataconfs = dict()
loss_computers = dict()
nr_input_sections = dict()
if evaltype != 'None':
evaluators = dict()
for task in self.conf['tasks'].split(' '):
taskconf = self.tasksconf[task]
#get the database configurations
input_names = modelconf.get('io', 'inputs').split(' ')
if input_names == ['']:
input_names = []
input_sections = [taskconf[i].split(' ') for i in input_names]
nr_input_sections[task] = len(input_sections)
task_input_dataconfs = []
for sectionset in input_sections:
task_input_dataconfs.append([])
for section in sectionset:
task_input_dataconfs[-1].append(
dict(dataconf.items(section)))
input_dataconfs[task] = task_input_dataconfs
output_names = taskconf['targets'].split(' ')
if output_names == ['']:
output_names = []
target_sections = [taskconf[o].split(' ') for o in output_names]
task_target_dataconfs = []
for sectionset in target_sections:
task_target_dataconfs.append([])
for section in sectionset:
task_target_dataconfs[-1].append(
dict(dataconf.items(section)))
target_dataconfs[task] = task_target_dataconfs
#create the loss computer
loss_computer = loss_computer_factory.factory(
taskconf['loss_type'])(self.batch_size)
loss_computers[task] = loss_computer
if evaltype != 'None':
evaluator = evaluator_factory.factory(evaltype)(
conf=evaluatorconf,
dataconf=dataconf,
model=self.model,
task=task)
evaluators[task] = evaluator
if 'local' in cluster.as_dict():
num_replicas = 1
device = tf.DeviceSpec(job='local')
else:
#distributed training
num_replicas = len(cluster.as_dict()['worker'])
num_servers = len(cluster.as_dict()['ps'])
ps_strategy = tf.contrib.training.GreedyLoadBalancingStrategy(
num_tasks=num_servers,
load_fn=tf.contrib.training.byte_size_load_fn)
device = tf.train.replica_device_setter(ps_tasks=num_servers,
ps_strategy=ps_strategy)
chief_ps = tf.DeviceSpec(job='ps', task=0)
self.is_chief = task_index == 0
#define the placeholders in the graph
with self.graph.as_default():
#create a local num_steps variable
self.num_steps = tf.get_variable(
name='num_steps',
shape=[],
dtype=tf.int32,
initializer=tf.constant_initializer(0),
trainable=False)
#a variable to hold the amount of steps already taken
self.global_step = tf.get_variable(
name='global_step',
shape=[],
dtype=tf.int32,
initializer=tf.constant_initializer(0),
trainable=False)
should_terminate = tf.get_variable(
name='should_terminate',
shape=[],
dtype=tf.bool,
initializer=tf.constant_initializer(False),
trainable=False)
self.terminate = should_terminate.assign(True).op
#create a check if training should continue
self.should_stop = tf.logical_or(
tf.greater_equal(self.global_step, self.num_steps),
should_terminate)
with tf.device(device):
data_queues = dict()
num_steps = []
done_ops = []
for task in self.conf['tasks'].split(' '):
#check if running in distributed model
if 'local' in cluster.as_dict():
#get the filenames
data_queue_elements, _ = input_pipeline.get_filenames(
input_dataconfs[task] + target_dataconfs[task])
#create the data queue and queue runners (inputs get shuffled! I already did this so set to False)
data_queue = tf.train.string_input_producer(
string_tensor=data_queue_elements,
shuffle=False,
seed=None,
capacity=self.batch_size * 2,
shared_name='data_queue_' + task)
data_queues[task] = data_queue
#compute the number of steps
if int(conf['numbatches_to_aggregate']) == 0:
task_num_steps = (int(conf['num_epochs']) *
len(data_queue_elements) /
self.batch_size)
else:
task_num_steps = (
int(conf['num_epochs']) *
len(data_queue_elements) /
(self.batch_size *
int(conf['numbatches_to_aggregate'])))
#set the number of steps
num_steps.append(task_num_steps)
done_ops.append(tf.no_op())
else:
with tf.device(chief_ps):
#get the data queue
data_queue = tf.FIFOQueue(
capacity=self.batch_size * (num_replicas + 1),
shared_name='data_queue_' + task,
name='data_queue_' + task,
dtypes=[tf.string],
shapes=[[]])
data_queues[task] = data_queue
#get the number of steps from the parameter server
num_steps_queue = tf.FIFOQueue(
capacity=num_replicas,
dtypes=[tf.int32],
shared_name='num_steps_queue',
name='num_steps_queue',
shapes=[[]])
#set the number of steps
task_num_steps = num_steps_queue.dequeue()
#get the done queues
for i in range(num_servers):
with tf.device('job:ps/task:%d' % i):
done_queue = tf.FIFOQueue(
capacity=num_replicas,
dtypes=[tf.bool],
shapes=[[]],
shared_name='done_queue%d' % i,
name='done_queue%d' % i)
done_ops.append(done_queue.enqueue(True))
self.set_num_steps = self.num_steps.assign(min(num_steps)).op
self.done = tf.group(*done_ops)
#training part
with tf.variable_scope('train'):
#a variable to scale the learning rate (used to reduce the
#learning rate in case validation performance drops)
learning_rate_fact = tf.get_variable(
name='learning_rate_fact',
shape=[],
initializer=tf.constant_initializer(1.0),
trainable=False)
#compute the learning rate with exponential decay and scale
#with the learning rate factor
self.learning_rate = (tf.train.exponential_decay(
learning_rate=float(conf['initial_learning_rate']),
global_step=self.global_step,
decay_steps=self.num_steps,
decay_rate=float(conf['learning_rate_decay'])) *
learning_rate_fact)
#create the optimizer
optimizer = tf.train.AdamOptimizer(self.learning_rate)
self.total_loss = tf.get_variable(
name='total_loss',
shape=[],
dtype=tf.float32,
initializer=tf.constant_initializer(0),
trainable=False)
self.reset_loss = self.total_loss.assign(0.0)
loss = []
for task in self.conf['tasks'].split(' '):
with tf.variable_scope(task):
#create the input pipeline
data, seq_length = input_pipeline.input_pipeline(
data_queue=data_queues[task],
batch_size=self.batch_size,
numbuckets=int(conf['numbuckets']),
dataconfs=input_dataconfs[task] +
target_dataconfs[task])
inputs = {
input_names[i]: d
for i, d in enumerate(
data[:nr_input_sections[task]])
}
seq_length = {
input_names[i]: d
for i, d in enumerate(
seq_length[:nr_input_sections[task]])
}
targets = {
output_names[i]: d
for i, d in enumerate(
data[nr_input_sections[task]:])
}
#target_seq_length = {
#output_names[i]: d
#for i, d in enumerate(seq_length[nr_input_sections[task]:])}
#compute the training outputs of the model
logits = self.model(inputs=inputs,
input_seq_length=seq_length,
is_training=True)
#TODO: The proper way to exploit data paralellism is via the
#SyncReplicasOptimizer defined below. However for some reason it hangs
#and I have not yet found a solution for it. For the moment the gradients
#are accumulated in a way that does not allow data paralellism and there
# is no advantage on having multiple workers. (We also accumulate the loss)
#create an optimizer that aggregates gradients
#if int(conf['numbatches_to_aggregate']) > 0:
#optimizer = tf.train.SyncReplicasOptimizer(
#opt=optimizer,
#replicas_to_aggregate=int(
#conf['numbatches_to_aggregate'])#,
##total_num_replicas=num_replicas
#)
#compute the loss
task_loss = loss_computers[task](targets, logits,
seq_length)
#append the task loss to the global loss
loss.append(task_loss)
#accumulate losses from tasks
with tf.variable_scope('accumulate_loss_from_tasks'):
loss = tf.reduce_mean(loss)
#accumulate losses from batches
self.acc_loss = self.total_loss.assign_add(loss)
##compute the gradients
#grads_and_vars = optimizer.compute_gradients(self.loss)
#with tf.variable_scope('clip'):
#clip_value = float(conf['clip_grad_value'])
##clip the gradients
#grads_and_vars = [(tf.clip_by_value(grad, -clip_value, clip_value), var)
#for grad, var in grads_and_vars]
self.params = tf.trainable_variables()
grads = [
tf.get_variable(param.op.name,
param.get_shape().as_list(),
initializer=tf.constant_initializer(0),
trainable=False)
for param in self.params
]
self.reset_grad = tf.variables_initializer(grads)
#compute the gradients
minibatch_grads_and_vars = optimizer.compute_gradients(
loss)
with tf.variable_scope('clip'):
clip_value = float(conf['clip_grad_value'])
#clip the gradients
minibatch_grads_and_vars = [
(tf.clip_by_value(grad, -clip_value,
clip_value), var)
for grad, var in minibatch_grads_and_vars
]
(minibatchgrads,
minibatchvars) = zip(*minibatch_grads_and_vars)
#update gradients by accumulating them
self.update_gradients = [
grad.assign_add(batchgrad)
for batchgrad, grad in zip(minibatchgrads, grads)
]
#opperation to apply the gradients
grads_and_vars = list(zip(grads, minibatchvars))
apply_gradients_op = optimizer.apply_gradients(
grads_and_vars=grads_and_vars,
global_step=self.global_step,
name='apply_gradients')
#all remaining operations with the UPDATE_OPS GraphKeys
update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
#create an operation to update the gradients, the batch_loss
#and do all other update ops
self.update_op = tf.group(*([apply_gradients_op] +
update_ops),
name='update')
if evaltype != 'None':
#validation part
with tf.variable_scope('validate'):
#create a variable to hold the validation loss
self.validation_loss = tf.get_variable(
name='validation_loss',
shape=[],
dtype=tf.float32,
initializer=tf.constant_initializer(0),
trainable=False)
#create a variable to save the last step where the model
#was validated
validated_step = tf.get_variable(
name='validated_step',
shape=[],
dtype=tf.int32,
initializer=tf.constant_initializer(
-int(conf['valid_frequency'])),
trainable=False)
#a check if validation is due
self.should_validate = tf.greater_equal(
self.global_step - validated_step,
int(conf['valid_frequency']))
val_batch_loss = []
valbatches = []
for task in self.conf['tasks'].split(' '):
with tf.variable_scope(task):
task_val_batch_loss, task_valbatches, _, _ = evaluators[
task].evaluate()
val_batch_loss.append(task_val_batch_loss)
valbatches.append(task_valbatches)
val_batch_loss = tf.reduce_mean(val_batch_loss)
self.valbatches = min(valbatches)
self.update_loss = self.validation_loss.assign(
self.validation_loss +
val_batch_loss #/self.valbatches
).op
#update the learning rate factor
self.half_lr = learning_rate_fact.assign(
learning_rate_fact / 2).op
#create an operation to updated the validated step
self.update_validated_step = validated_step.assign(
self.global_step).op
#variable to hold the best validation loss so far
self.best_validation = tf.get_variable(
name='best_validation',
shape=[],
dtype=tf.float32,
initializer=tf.constant_initializer(1.79e+308),
trainable=False)
#op to update the best velidation loss
self.update_best = self.best_validation.assign(
self.validation_loss).op
#a variable that holds the amount of workers at the
#validation point
waiting_workers = tf.get_variable(
name='waiting_workers',
shape=[],
dtype=tf.int32,
initializer=tf.constant_initializer(0),
trainable=False)
#an operation to signal a waiting worker
self.waiting = waiting_workers.assign_add(1).op
#an operation to set the waiting workers to zero
self.reset_waiting = waiting_workers.initializer
#an operation to check if all workers are waiting
self.all_waiting = tf.equal(waiting_workers,
num_replicas - 1)
tf.summary.scalar('validation loss',
self.validation_loss)
else:
self.update_loss = None
tf.summary.scalar('learning rate', self.learning_rate)
#create a histogram for all trainable parameters
for param in tf.trainable_variables():
tf.summary.histogram(param.name, param)
#create the scaffold
self.scaffold = tf.train.Scaffold()
def __init__(self, conf, tasksconf, dataconf, modelconf, evaluatorconf,
expdir, init_filename, server, task_index):
'''
MultiTaskTrainer constructor, creates the training graph
Args:
conf: the trainer config
taskconf: the config file for each task
dataconf: the data configuration as a ConfigParser
modelconf: the neural net model configuration
evaluatorconf: the evaluator configuration for evaluating
if None no evaluation will be done
expdir: directory where the summaries will be written
init_filename: filename of the network that should be used to
initialize the model. Put to None if no network is available/wanted.
server: optional server to be used for distributed training
task_index: optional index of the worker task in the cluster
'''
self.expdir = expdir
self.server = server
self.conf = conf
self.tasksconf = tasksconf
self.task_index = task_index
self.init_filename = init_filename
self.batch_size = int(conf['batch_size'])
cluster = tf.train.ClusterSpec(server.server_def.cluster)
#create the graph
self.graph = tf.Graph()
#create the model
modelfile = os.path.join(expdir, 'model', 'model.pkl')
model_names = modelconf.get('hyper', 'model_names').split(' ')
self.models = dict()
with open(modelfile, 'wb') as fid:
for model_name in model_names:
self.models[model_name] = model_factory.factory(
modelconf.get(model_name, 'architecture'))(conf=dict(
modelconf.items(model_name)),
name=model_name)
pickle.dump(self.models, fid)
evaltype = evaluatorconf.get('evaluator', 'evaluator')
#define a trainer per traintask
self.task_trainers = []
for task in self.conf['tasks'].split(' '):
taskconf = self.tasksconf[task]
task_trainer = task_trainer_script.TaskTrainer(
task, conf, taskconf, self.models, modelconf, dataconf,
evaluatorconf, self.batch_size)
self.task_trainers.append(task_trainer)
if 'local' in cluster.as_dict():
num_replicas = 1
device = tf.DeviceSpec(job='local')
else:
#distributed training
num_replicas = len(cluster.as_dict()['worker'])
num_servers = len(cluster.as_dict()['ps'])
ps_strategy = tf.contrib.training.GreedyLoadBalancingStrategy(
num_tasks=num_servers,
load_fn=tf.contrib.training.byte_size_load_fn)
device = tf.train.replica_device_setter(ps_tasks=num_servers,
ps_strategy=ps_strategy)
chief_ps = tf.DeviceSpec(job='ps', task=0)
self.is_chief = task_index == 0
#define the placeholders in the graph
with self.graph.as_default():
#create a local num_steps variable
self.num_steps = tf.get_variable(
name='num_steps',
shape=[],
dtype=tf.int32,
initializer=tf.constant_initializer(0),
trainable=False)
#a variable to hold the amount of steps already taken
self.global_step = tf.get_variable(
name='global_step',
shape=[],
dtype=tf.int32,
initializer=tf.constant_initializer(0),
trainable=False)
should_terminate = tf.get_variable(
name='should_terminate',
shape=[],
dtype=tf.bool,
initializer=tf.constant_initializer(False),
trainable=False)
self.terminate = should_terminate.assign(True).op
#create a check if training should continue
self.should_stop = tf.logical_or(
tf.greater_equal(self.global_step, self.num_steps),
should_terminate)
with tf.device(device):
num_steps = []
done_ops = []
#set the dataqueues for each trainer
for task_trainer in self.task_trainers:
task_num_steps, task_done_ops = task_trainer.set_dataqueues(
cluster)
num_steps.append(task_num_steps)
done_ops += task_done_ops
self.set_num_steps = self.num_steps.assign(min(num_steps)).op
self.done = tf.group(*done_ops)
#training part
with tf.variable_scope('train'):
#a variable to scale the learning rate (used to reduce the
#learning rate in case validation performance drops)
learning_rate_fact = tf.get_variable(
name='learning_rate_fact',
shape=[],
initializer=tf.constant_initializer(1.0),
trainable=False)
#compute the learning rate with exponential decay and scale
#with the learning rate factor
self.learning_rate = (tf.train.exponential_decay(
learning_rate=float(conf['initial_learning_rate']),
global_step=self.global_step,
decay_steps=self.num_steps,
decay_rate=float(conf['learning_rate_decay'])) *
learning_rate_fact)
#For each task, set the task specific training ops
for task_trainer in self.task_trainers:
task_trainer.train(self.learning_rate)
#Group ops over tasks
self.process_minibatch = tf.group(
*([
task_trainer.process_minibatch
for task_trainer in self.task_trainers
]),
name='process_minibatch_all_tasks')
self.reset_grad_loss_norm = tf.group(
*([
task_trainer.reset_grad_loss_norm
for task_trainer in self.task_trainers
]),
name='reset_grad_loss_norm_all_tasks')
tmp = []
# should'nt this be task_trainer instead of task?
# for task in self.task_trainers:
for task_trainer in self.task_trainers:
tmp += task_trainer.normalize_gradients
self.normalize_gradients = tf.group(
*(tmp), name='normalize_gradients_all_tasks')
#accumulate losses from tasks
with tf.variable_scope('accumulate_losses_from_tasks'):
self.loss_all_tasks = [
task_trainer.normalized_loss
for task_trainer in self.task_trainers
]
self.total_loss = tf.reduce_mean(self.loss_all_tasks,
name='acc_loss')
tmp = []
for task_trainer in self.task_trainers:
tmp.append(task_trainer.apply_gradients)
#all remaining operations with the UPDATE_OPS GraphKeys
update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
#an op to increment the global step
global_step_inc = self.global_step.assign_add(1)
#create an operation to update the gradients, the batch_loss
#and do all other update ops
#self.update_op = tf.group(
#*(tmp + update_ops + [global_step_inc]),
#name='update')
self.other_update_op = tf.group(*(update_ops +
[global_step_inc]),
name='other_update')
if evaltype != 'None':
#validation part
with tf.variable_scope('validate'):
#create a variable to save the last step where the model
#was validated
validated_step = tf.get_variable(
name='validated_step',
shape=[],
dtype=tf.int32,
initializer=tf.constant_initializer(
-int(conf['valid_frequency'])),
trainable=False)
#a check if validation is due
self.should_validate = tf.greater_equal(
self.global_step - validated_step,
int(conf['valid_frequency']))
#For each task, if requested, set the task specific validation ops
#The number of validation batches is the minimum number of validation
#batches over all tasks.
tasks_excluded_for_val = ['None']
if evaluatorconf.has_option('evaluator',
'tasks_excluded_for_val'):
tasks_excluded_for_val = evaluatorconf.get(
'evaluator',
'tasks_excluded_for_val').split(' ')
self.val_task_trainers = [
task_trainer for task_trainer in self.task_trainers
if task_trainer.task_name not in
tasks_excluded_for_val
]
valbatches = []
for task_trainer in self.val_task_trainers:
valbatches.append(
task_trainer.evaluate_evaluator())
self.valbatches = min(valbatches)
#Group ops over tasks
self.process_val_batch = tf.group(*([
task_trainer.process_val_batch
for task_trainer in self.val_task_trainers
]))
self.reset_val_loss_norm = tf.group(*([
task_trainer.reset_val_loss_norm
for task_trainer in self.val_task_trainers
]))
self.val_loss_all_tasks = []
for task_trainer in self.val_task_trainers:
self.val_loss_all_tasks.append(
task_trainer.val_loss_normalized)
self.validation_loss = tf.reduce_mean(
self.val_loss_all_tasks)
#update the learning rate factor
self.half_lr = learning_rate_fact.assign(
learning_rate_fact / 2).op
#create an operation to updated the validated step
self.update_validated_step = validated_step.assign(
self.global_step).op
#variable to hold the best validation loss so far
self.best_validation_all_tasks = [
tf.get_variable(
name='best_validation_task_%i' % ind,
shape=[],
dtype=tf.float32,
initializer=tf.constant_initializer(1.79e+308),
trainable=False)
for ind in range(len(self.val_task_trainers))
]
#op to update the best validation loss
self.update_best_all_tasks = \
[best_val_task.assign(self.val_loss_all_tasks[ind])
for ind,best_val_task in enumerate(self.best_validation_all_tasks)]
#a variable that holds the amount of workers at the
#validation point
waiting_workers = tf.get_variable(
name='waiting_workers',
shape=[],
dtype=tf.int32,
initializer=tf.constant_initializer(0),
trainable=False)
#an operation to signal a waiting worker
self.waiting = waiting_workers.assign_add(1).op
#an operation to set the waiting workers to zero
self.reset_waiting = waiting_workers.initializer
#an operation to check if all workers are waiting
self.all_waiting = tf.equal(waiting_workers,
num_replicas - 1)
tf.summary.scalar('validation loss',
self.validation_loss)
else:
self.process_val_batch = None
tf.summary.scalar('learning rate', self.learning_rate)
#create a histogram for all trainable parameters
for param in tf.trainable_variables():
tf.summary.histogram(param.name, param)
#create the scaffold
self.scaffold = tf.train.Scaffold()