def __getattribute__(user_class, attribute_name):
# Overriding train in order to minimize the changes necessary in the user
# code to go from local to remote execution.
# That way, by simply commenting or uncommenting the Train decorator
# Model.train() will execute either on the local setup or in kubernetes
if attribute_name != 'train' or user_class.is_training_initialized:
return super(UserClass,
user_class).__getattribute__(attribute_name)
if attribute_name == 'train' and not is_runtime_phase():
return super(
UserClass,
user_class).__getattribute__('_deploy_training')
user_class.is_training_initialized = True
self.trainer.start_training(user_class)
return super(UserClass,
user_class).__getattribute__('_noop_attribute')
def __getattribute__(user_object, attribute_name):
# Overriding train in order to minimize the changes
# necessary in the user code to go from local to remoten't be here probably
# execution.
# That way, by simply commenting or uncommenting the
# Train decorator model.train() will execute either on the local
# setup or in kubernetes.
if (attribute_name != 'train'
or user_object.is_training_initialized):
return super(UserClass, user_object).__getattribute__(
attribute_name)
if attribute_name == 'train' and not utils.is_runtime_phase():
return super(UserClass, user_object).__getattribute__(
'_deploy_training')
user_object.is_training_initialized = True
self._train(user_object)
return super(UserClass, user_object).__getattribute__(
'_noop_attribute')
def wrapped():
if is_runtime_phase():
return self.serve(func)
ast = self.backend.compile_serving_ast(self.image,
self.package.name,
self.port, self.replicas)
self.builder.write_dockerfile(self.package)
self.builder.build(self.image)
if self.package.publish:
self.builder.publish(self.image)
def signal_handler(signal, frame):
mp.cancel(self.package.name)
sys.exit(0)
signal.signal(signal.SIGINT, signal_handler)
mp.run(ast)
logger.warn("Server deployed.")
return mp.logs(self.package.name)
import logging
import types
import redis
import os
import json
import math
import numpy as np
import time
from ..basic import BasicTrainingStrategy
from .exploit import Truncation
from .explore import Resample, Perturb
from fairing.utils import is_runtime_phase
#TODO: can we make this class framework agnostic?
if is_runtime_phase():
import tensorflow as tf
logger = logging.getLogger('fairing')
class PopulationBasedTraining(BasicTrainingStrategy):
def __init__(
self,
model_path,
population_size,
exploit_count,
steps_per_exploit,
pvc_name,
exploiter=Truncation(),
explorer=Perturb(),
def test_is_runtime_phase(runtime_phase, monkeypatch):
if runtime_phase:
monkeypatch.setenv("FAIRING_RUNTIME", True)
assert is_runtime_phase() == runtime_phase