def train(self, dataframe, shuffle=False):
"""See distkeras.trainers.Trainer.train
# Arguments
dataframe: dataframe. Spark Dataframe
shuffle: boolean. Tells to shuffle the dataframe before training.
Warning: this will tell Spark to shuffle all partitions over
the network. It is recommended to shuffle the dataset before
training and store it.
"""
# Assign the dataset.
dataset = dataframe
# Build the dataset with the number of epochs.
for i in range(0, self.num_epoch):
dataset = dataset.unionAll(dataframe)
# Check if the data needs to be shuffled.
if shuffle:
dataset = shuffle(dataset)
# Collect the dataset on a single worker node.
dataset = dataset.coalesce(1)
# Cache the dataset.
dataset.cache()
# Allocate a worker.
worker = self.allocate_worker()
# Start recording training time.
self.record_training_start()
# Fetch the trained model.
self.master_model = dataset.rdd.mapPartitionsWithIndex(
worker.train).collect()[0]
# Stop recording of training time.
self.record_training_end()
return deserialize_keras_model(self.master_model)
def train(self, dataframe, shuffle=False):
"""See distkeras.trainers.Trainer.train
# Arguments
dataframe: dataframe. A Spark Dataframe containing the training data.
shuffle: boolean. Tells to shuffle the dataframe before training.
Warning: this will tell Spark to shuffle all partitions over
the network. It is recommended to shuffle the dataframe before
training and store it.
"""
# Check if the data needs to be shuffled.
if shuffle:
dataframe = shuffle(dataframe)
# Collect the dataframe on a single worker node.
dataframe = dataframe.coalesce(1)
# Cache the dataframe.
dataframe.cache()
# Allocate a worker.
worker = self.allocate_worker()
# Set the maximum number of mini-batches.
worker.set_max_prefetch(self.max_mini_batches_prefetch)
# Start recording training time.
self.record_training_start()
# Fetch the trained model.
self.master_model = dataframe.rdd.mapPartitionsWithIndex(
worker.train).collect()[0]
# Stop recording of training time.
self.record_training_end()
return deserialize_keras_model(self.master_model)
def train(self, dataframe, shuffle=False):
"""Applies model averaging to the model replicas distributed over the specified
number of Spark executors.
# Arguments
dataframe: dataframe: A Spark Dataframe containing the training data.
shuffle: boolean. Tells to shuffle the dataframe before training.
Warning: this will tell Spark to shuffle all partitions over
the network. It is recommended to shuffle the dataframe before
training and store it.
"""
# Repartition the data in order to fit the number of workers.
num_partitions = dataframe.rdd.getNumPartitions()
# Check if the dataframe needs to be shuffled.
if shuffle:
dataframe = shuffle(dataframe)
# Check if we need to repartition the dataframe.
if num_partitions >= self.num_workers:
dataframe = dataframe.coalesce(self.num_workers)
else:
dataframe = dataframe.repartition(self.num_workers)
# Start the training procedure.
self.record_training_start()
for i in range(0, self.num_epoch):
worker = self.allocate_worker()
# Set the maximum number of mini-batches.
worker.set_max_prefetch(self.max_mini_batches_prefetch)
models = dataframe.rdd.mapPartitionsWithIndex(
worker.train).collect()
self.average_models(models)
# End the training procedure.
self.record_training_end()
return deserialize_keras_model(self.master_model)
def train(self, dataframe, shuffle=False):
"""Applies model averaging to the model replicas distributed over the specified
number of Spark executors.
# Arguments
dataframe: dataframe: A Spark Dataframe containing the training data.
shuffle: boolean. Tells to shuffle the dataframe before training.
Warning: this will tell Spark to shuffle all partitions over
the network. It is recommended to shuffle the dataframe before
training and store it.
"""
# Repartition the data in order to fit the number of workers.
num_partitions = dataframe.rdd.getNumPartitions()
# Check if the dataframe needs to be shuffled.
if shuffle:
dataframe = shuffle(dataframe)
# Check if we need to repartition the dataframe.
if num_partitions >= self.num_workers:
dataframe = dataframe.coalesce(self.num_workers)
else:
dataframe = dataframe.repartition(self.num_workers)
# Start the training procedure.
self.record_training_start()
for i in range(0, self.num_epoch):
worker = self.allocate_worker()
# Set the maximum number of mini-batches.
worker.set_max_prefetch(self.max_mini_batches_prefetch)
models = dataframe.rdd.mapPartitionsWithIndex(worker.train).collect()
self.average_models(models)
# End the training procedure.
self.record_training_end()
return deserialize_keras_model(self.master_model)
def train(self, dataframe, shuffle=False):
"""See distkeras.trainers.Trainer.train
# Arguments
dataframe: dataframe. A Spark Dataframe containing the training data.
shuffle: boolean. Tells to shuffle the dataframe before training.
Warning: this will tell Spark to shuffle all partitions over
the network. It is recommended to shuffle the dataframe before
training and store it.
"""
# Check if the data needs to be shuffled.
if shuffle:
dataframe = shuffle(dataframe)
# Collect the dataframe on a single worker node.
dataframe = dataframe.coalesce(1)
# Cache the dataframe.
dataframe.cache()
# Allocate a worker.
worker = self.allocate_worker()
# Set the maximum number of mini-batches.
worker.set_max_prefetch(self.max_mini_batches_prefetch)
# Start recording training time.
self.record_training_start()
# Fetch the trained model.
self.master_model = dataframe.rdd.mapPartitionsWithIndex(worker.train).collect()[0]
# Stop recording of training time.
self.record_training_end()
return deserialize_keras_model(self.master_model)
def prepare_model(self):
"""Prepares the model for training."""
# Deserialize the Keras model.
self.model = deserialize_keras_model(self.model)
# Compile the model with the specified loss and optimizer.
self.model.compile(loss=self.loss,
optimizer=self.optimizer,
metrics=['accuracy'])
def destroy_remote_job(self):
address = self.address + '/api/destroy?secret=' + self.secret
request = urllib2.Request(address)
response = urllib2.urlopen(request)
data = json.load(response)
model = unpickle_object(data['model'].decode('hex_codec'))
self.trained_model = deserialize_keras_model(model)
self.history = unpickle_object(data['history'].decode('hex_codec'))
def prepare_model(self):
"""Prepares the model for training."""
# Set the Keras directory.
set_keras_base_directory()
# Deserialize the Keras model.
self.model = deserialize_keras_model(self.model)
# Compile the model with the specified loss and optimizer.
self.model.compile(loss=self.loss, optimizer=self.optimizer, metrics=self.metrics)
def average_models(self, models):
"""Averages the specified list of Keras models, and assigns the
averaged model as the master model.
# Arguments:
models: list. A list of serialized Keras models.
"""
num_models = len(models)
# Get all weights of the models.
for i in range(0, num_models):
weights = np.asarray(deserialize_keras_model(models[i]).get_weights())
self.parameter_buffer += weights
# Average the parameters.
self.parameter_buffer /= num_models
temp_model = deserialize_keras_model(self.master_model)
temp_model.set_weights(self.parameter_buffer)
self.master_model = serialize_keras_model(temp_model)
def _predict(self, iterator):
"""Lambda method which will append a prediction column to the provided rows.
# Arguments:
iterator: iterator. Spark Row iterator.
"""
model = deserialize_keras_model(self.model)
for row in iterator:
features = [np.asarray([row[c]]) for c in self.features_column]
prediction = model.predict(features)
dense_prediction = DenseVector(prediction[0])
new_row = new_dataframe_row(row, self.output_column, dense_prediction)
yield new_row
def _predict(self, iterator):
"""Lambda method which will append a prediction column to the provided rows.
# Arguments:
iterator: iterator. Spark Row iterator.
"""
model = deserialize_keras_model(self.model)
for row in iterator:
features = np.asarray([row[self.features_column]])
prediction = model.predict(features)
dense_prediction = DenseVector(prediction[0])
new_row = new_dataframe_row(row, self.output_column,
dense_prediction)
yield new_row
def prepare_model(self):
"""Prepares the model for training."""
# Set the Keras directory.
set_keras_base_directory()
if K.backend() == 'tensorflow':
# set GPU option allow_growth to False for GPU-enabled tensorflow
config = tf.ConfigProto()
config.gpu_options.allow_growth = False
sess = tf.Session(config=config)
K.set_session(sess)
# Deserialize the Keras model.
self.model = deserialize_keras_model(self.model)
self.optimizer = deserialize(self.optimizer)
# Compile the model with the specified loss and optimizer.
self.model.compile(loss=self.loss, loss_weights = self.loss_weights,
optimizer=self.optimizer, metrics=self.metrics)
def __init__(self, model):
self.model = deserialize_keras_model(model)
self.num_updates = 1
def read_trained_model(self):
home = expanduser("~")
with open(home + "/models/" + self.secret, "r") as f:
self.trained_model = deserialize_keras_model(
unpickle_object(f.read()))
