def train(config):
model = cls.Cls(config, loss_mode=sys.argv[1])
model.initialize_weights()
max_training_step = config.max_training_step
best_acc = 0
endurance = 0
i = 0
logger.info('lambda1: {}'.format(config.lambda1_stud))
while i < max_training_step and endurance < config.max_endurance_stud:
train_loss, train_acc = model.train()
if i % config.valid_frequency_stud == 0:
endurance += 1
valid_loss, valid_acc, _, _ = model.valid()
logger.info('====Step: {}===='.format(i))
logger.info('train_loss: {}, valid_loss: {}'\
.format(train_loss, valid_loss))
logger.info('train_acc: {}, valid_acc: {}'\
.format(train_acc, valid_acc))
if valid_acc > best_acc:
best_acc = valid_acc
_, test_acc, _, _ = model.valid(model.test_dataset)
endurance = 0
i += 1
logger.info('lambda1: {}, lambda2: {}'.format(config.lambda1_stud,
config.lambda2_stud))
logger.info('valid_acc: {}'.format(best_acc))
logger.info('test_acc: {}'.format(test_acc))
return test_acc
def train(config):
g = tf.Graph()
gpu_options = tf.GPUOptions(allow_growth=True)
configProto = tf.ConfigProto(gpu_options=gpu_options)
sess = tf.InteractiveSession(config=configProto, graph=g)
model = cls.Cls(config, g, loss_mode=sys.argv[1])
sess.run(model.init)
max_training_step = config.max_training_step
best_acc = 0
endurance = 0
i = 0
while i < max_training_step and endurance < config.max_endurance_stud:
train_loss, train_acc = model.train(sess)
if i % config.valid_frequence_stud == 0:
endurance += 1
valid_loss, valid_acc, _, _ = model.valid(sess)
#logger.info('====Step: {}===='.format(i))
#logger.info('train_loss: {}, train_acc: {}'\
# .format(train_loss, train_acc))
#logger.info('valid_loss: {}, valid_acc: {}'\
# .format(valid_loss, valid_acc))
if valid_acc > best_acc:
best_acc = valid_acc
_, test_acc, _, _ = model.valid(sess, model.test_dataset)
endurance = 0
i += 1
logger.info('lambda1: {}, lambda2: {}'.format(config.lambda1_stud,
config.lambda2_stud))
logger.info('valid_acc: {}'.format(best_acc))
logger.info('test_acc: {}'.format(test_acc))
return test_acc
def __init__(self, config, args):
self.config = config
#hostname = config.hostname
#hostname = '-'.join(hostname.split('.')[0:2])
#datetime = strftime('%m-%d-%H-%M', gmtime())
#exp_name = '{}_{}_{}'.format(hostname, datetime, args.exp_name)
exp_name = args.exp_name
logger.info('exp_name: {}'.format(exp_name))
if config.rl_method == 'reinforce':
self.model_ctrl = controller_reinforce.Controller(config, exp_name+'_ctrl')
elif config.rl_method == 'ppo':
self.model_ctrl = controller_ppo.MlpPPO(config, exp_name+'_ctrl')
if args.task_name == 'reg':
from models import reg
self.model_task = reg.Reg(config, exp_name+'_reg')
elif args.task_name == 'cls':
from models import cls
self.model_task = cls.Cls(config, exp_name+'_cls')
elif args.task_name == 'gan':
from models import gan
self.model_task = gan.Gan(config, exp_name+'_gan')
elif args.task_name == 'gan_cifar10':
from models import gan_cifar10
self.model_task = gan_cifar10.Gan_cifar10(config,
exp_name+'_gan_cifar10')
elif args.task_name == 'nmt':
from models import nmt
self.model_task = nmt.Nmt(config, exp_name+'_nmt')
else:
raise NotImplementedError
def __init__(self, config, exp_name=None, arch=None):
self.config = config
hostname = socket.gethostname()
hostname = '-'.join(hostname.split('.')[0:2])
datetime = strftime('%m-%d-%H-%M', gmtime())
if not exp_name:
exp_name = '{}_{}'.format(hostname, datetime)
logger.info('exp_name: {}'.format(exp_name))
self.model_ctrl = controller.Controller(config, exp_name+'_ctrl')
if config.student_model_name == 'toy':
self.model_stud = toy.Toy(config, exp_name+'_reg')
elif config.student_model_name == 'cls':
self.model_stud = cls.Cls(config, exp_name+'_cls')
elif config.student_model_name == 'gan':
self.model_stud = gan.Gan(config, exp_name+'_gan', arch=arch)
elif config.student_model_name == 'gan_grid':
self.model_stud = gan_grid.Gan_grid(config, exp_name+'_gan_grid')
elif config.student_model_name == 'gan_cifar10':
self.model_stud = gan_cifar10.Gan_cifar10(config,
exp_name+'_gan_cifar10')
else:
raise NotImplementedError