def train(config='configs/cifar10_regression.json',
save_root_path='/cluster/work/math/fanyang-broglil/CoreRepo',
experiment_json_fname='experiments.json',
local_json_dir_name='local_json_files',
worstofk=None,
attack_style=None,
attack_limits=None,
fo_epsilon=None,
fo_step_size=None,
fo_num_steps=None,
lambda_core=None,
num_ids = None,
group_size=None,
use_core=None,
seed=None,
save_in_local_json=True,
this_repo=None):
# reset default graph (needed for running locally with run_jobs_ray.py)
tf.reset_default_graph()
# get configs
config_dict = utilities.get_config(config)
config_dict_copy = copy.deepcopy(config_dict)
config = utilities.config_to_namedtuple(config_dict)
# seeding randomness
if seed == None:
seed = config.training.tf_random_seed
else:
config_dict_copy['training']['tf_random_seed'] = seed
tf.set_random_seed(seed)
np.random.seed(seed)
# Setting up training parameters
max_num_training_steps = config.training.max_num_training_steps
step_size_schedule = config.training.step_size_schedule
weight_decay = config.training.weight_decay
momentum = config.training.momentum
if group_size == None:
group_size = config.training.group_size
else:
config_dict_copy['training']['group_size'] = int(group_size)
if num_ids == None:
num_ids = config.training.num_ids
else:
config_dict_copy['training']['num_ids'] = int(num_ids)
if lambda_core == None:
lambda_core = config.training.lambda_
else:
config_dict_copy['training']['lambda_'] = float(lambda_core)
if use_core == None:
use_core = config.training.use_core
else:
config_dict_copy['training']['use_core'] = use_core
batch_size = config.training.batch_size
# number of groups with group size > 1
num_grouped_ids = batch_size - num_ids
# number of unique ids needs to be larger than half the desired batch size
# so that full batch can be filled up
assert num_ids >= batch_size/2
# currently, code is designed for groups of size 2
assert config.training.group_size == 2
adversarial_training = config.training.adversarial_training
eval_during_training = config.training.eval_during_training
if eval_during_training:
num_eval_steps = config.training.num_eval_steps
# Setting up output parameters
num_output_steps = config.training.num_output_steps
num_summary_steps = config.training.num_summary_steps
num_checkpoint_steps = config.training.num_checkpoint_steps
num_easyeval_steps = config.training.num_easyeval_steps
# Setting up the data and the model
data_path = config.data.data_path
if config.data.dataset_name == "cifar-10":
raw_iterator = cifar10_input.CIFAR10Data(data_path)
elif config.data.dataset_name == "cifar-100":
raw_iterator = cifar100_input.CIFAR100Data(data_path)
elif config.data.dataset_name == "svhn":
raw_iterator = svhn_input.SVHNData(data_path)
else:
raise ValueError("Unknown dataset name.")
global_step = tf.train.get_or_create_global_step()
model_family = config.model.model_family
if model_family == "resnet":
if config.attack.use_spatial and config.attack.spatial_method == 'fo':
diffable = True
else:
diffable = False
model = resnet_reg.Model(config.model, num_ids, diffable,
config.training.adversarial_ce)
elif model_family == "vgg":
if config.attack.use_spatial and config.attack.spatial_method == 'fo':
diffable = True
else:
diffable = False
if config.training.adversarial_ce:
raise NotImplementedError
model = vgg.Model(config.model, num_ids, diffable)
# uncomment to get a list of trainable variables
# model_vars = tf.trainable_variables()
# Setting up the optimizer
boundaries = [int(sss[0]) for sss in step_size_schedule]
boundaries = boundaries[1:]
values = [sss[1] for sss in step_size_schedule]
learning_rate = tf.train.piecewise_constant(
tf.cast(global_step, tf.int32),
boundaries,
values)
if use_core and lambda_core > 0:
print("WARNING: in regression task, should not enter this section!\n")
total_loss = (model.reg_loss + weight_decay * model.weight_decay_loss +
lambda_core * model.core_loss2)
else:
total_loss = model.reg_loss + weight_decay * model.weight_decay_loss
optimizer = tf.train.MomentumOptimizer(learning_rate, momentum)
train_step = optimizer.minimize(total_loss, global_step=global_step)
# Set up adversary
if worstofk == None:
worstofk = config.attack.random_tries
else:
config_dict_copy['attack']['random_tries'] = worstofk
if fo_epsilon == None:
fo_epsilon = config.attack.epsilon
else:
config_dict_copy['attack']['epsilon'] = fo_epsilon
if fo_step_size == None:
fo_step_size = config.attack.step_size
else:
config_dict_copy['attack']['step_size'] = fo_step_size
if fo_num_steps == None:
fo_num_steps = config.attack.num_steps
else:
config_dict_copy['attack']['num_steps'] = fo_num_steps
# @ Luzius: incorporate being able to choose multiple transformations
if attack_style == None:
attack_style = 'rotate'
simple_train = config.attack.simple_train
if simple_train == False:
# Training attack == denfense
# L-inf attack if use_spatial is False and use_linf is True
# spatial attack if use_spatial is True and use_linf is False
# spatial random attack if spatial_method is 'random'
# spatial PGD attack if spatial_method is 'fo'
attack = SpatialAttack(model, config.attack, config.attack.spatial_method,
worstofk, attack_limits, fo_epsilon,
fo_step_size, fo_num_steps)
# Different eval attacks
# Random attack
# L-inf attack if use_spatial is False and use_linf is True
# random (worst-of-1) spatial attack if use_spatial is True
# and use_linf is False
attack_eval_random = SpatialAttack(model, config.attack, 'random', 1,
attack_limits, fo_epsilon,
fo_step_size, fo_num_steps)
# First order attack
# L-inf attack if use_spatial is False and use_linf is True
# first-order spatial attack if use_spatial is True and use_linf is False
attack_eval_fo = SpatialAttack(model, config.attack, 'fo', 1,
attack_limits, fo_epsilon,
fo_step_size, fo_num_steps)
# Grid attack
# spatial attack if use_spatial is True and use_linf is False
# not executed for L-inf attacks
attack_eval_grid = SpatialAttack(model, config.attack, 'grid', None,
attack_limits)
else:
attack = SpatialAttack(model, config.attack, config.attack.spatial_method,
worstofk, attack_limits, fo_epsilon,
fo_step_size, fo_num_steps)
# TODO(christina): add L-inf attack with random restarts
# ------------------START EXPERIMENT -------------------------
# Initialize the Repo
print("==> Creating repo..")
# Create repo object if it wasn't passed, comment out if repo has issues
if this_repo == None:
this_repo = exprepo.ExperimentRepo(
save_in_local_json=save_in_local_json,
json_filename=experiment_json_fname,
local_dir_name=local_json_dir_name,
root_dir=save_root_path)
# Create new experiment
if this_repo != None:
exp_id = this_repo.create_new_experiment(config.data.dataset_name,
model_family,
worstofk,
attack_style,
attack_limits,
lambda_core,
num_grouped_ids,
group_size,
config_dict_copy)
# Setting up the Tensorboard and checkpoint outputs
model_dir = '%s/logdir/%s' % (save_root_path, exp_id)
# We add accuracy and xent twice so we can easily make three types of
# comparisons in Tensorboard:
# - train vs eval (for a single run)
# - train of different runs
# - eval of different runs
saver = tf.train.Saver(max_to_keep=3)
tf.summary.scalar('regression loss function value', model.reg_loss, collections= ['err'])
tf.summary.scalar('avg_abs_err_x', model.avg_abs_err_transX, collections=['err'])
tf.summary.scalar('avg_abs_err_y', model.avg_abs_err_transY, collections=['err'])
tf.summary.scalar('avg_abs_err_rot', model.avg_abs_err_rot, collections=['err'])
tf.summary.scalar('avg_rel_err_x', model.avg_rel_err_transX, collections=['err'])
tf.summary.scalar('avg_rel_err_y', model.avg_rel_err_transY, collections=['err'])
tf.summary.scalar('avg_rel_err_rot', model.avg_rel_err_rot, collections=['err'])
tf.summary.scalar('learning_rate', learning_rate, collections=['err'])
tf.summary.image('before_reflect_padding', model.before_reflect_x, collections=['err'])
tf.summary.image('after_reflect_padding', model.reflect_x, collections=['err'])
err_summaries = tf.summary.merge_all('err')
tf.summary.scalar('full_batch_avg_abs_err_x', model.avg_abs_err_transX, collections=['eval'])
tf.summary.scalar('full_batch_avg_abs_err_y', model.avg_abs_err_transY, collections=['eval'])
tf.summary.scalar('full_batch_avg_abs_err_rot', model.avg_abs_err_rot, collections=['eval'])
tf.summary.scalar('full_batch_avg_rel_err_x', model.avg_rel_err_transX, collections=['eval'])
tf.summary.scalar('full_batch_avg_rel_err_y', model.avg_rel_err_transY, collections=['eval'])
tf.summary.scalar('full_batch_avg_rel_err_rot', model.avg_rel_err_rot, collections=['eval'])
tf.summary.scalar('full_batch_avg_worst', model.avg_worst_err, collections=['eval'])
# tf.summary.scalar('learning_rate', learning_rate, collections=['eval'])
eval_summaries = tf.summary.merge_all('eval')
# data augmentation used if config.training.data_augmentation_core is True
x_input_placeholder = tf.placeholder(tf.float32, shape=[None, 32, 32, 3])
flipped = tf.map_fn(lambda img: tf.image.random_flip_left_right(img),
x_input_placeholder)
with tf.Session() as sess:
# initialize standard data augmentation
if config.training.data_augmentation:
if config.data.dataset_name == "cifar-10":
data_iterator = cifar10_input.AugmentedCIFAR10Data(raw_iterator, sess)
elif config.data.dataset_name == "cifar-100":
data_iterator = cifar100_input.AugmentedCIFAR100Data(raw_iterator, sess)
elif config.data.dataset_name == "svhn":
data_iterator = svhn_input.AugmentedSVHNData(raw_iterator, sess)
else:
raise ValueError("Unknown dataset name.")
else:
data_iterator = raw_iterator
if simple_train:
# attack = SpatialAttack(model, config.attack, config.attack.spatial_method,
# worstofk, attack_limits, fo_epsilon,
# fo_step_size, fo_num_steps)
# attack.simple_train_perturb should return a list of parameters (len(x_batch),3)
x_eval_batch = data_iterator.eval_data.xs
x_batch_eval = x_eval_batch
# the evaluation batch labels are 3 dim transformations now
y_batch_eval = data_iterator.eval_data.ys
# we pass the label values to the model. as the transformation
trans_eval = y_batch_eval
eval_dict = {model.x_input: x_batch_eval,
model.y_input: y_batch_eval,
# group is not used in simple train
model.group: np.arange(0, batch_size, 1, dtype="int32"),
model.transform: trans_eval,
model.is_training: False}
else:
eval_dict = {model.x_input: data_iterator.eval_data.xs,
model.y_input: data_iterator.eval_data.ys,
model.group: np.arange(0, batch_size, 1, dtype="int32"),
model.transform: np.zeros([data_iterator.eval_data.n, 3]),
model.is_training: False}
# Initialize the summary writer, global variables, and our time counter.
summary_writer = tf.summary.FileWriter(model_dir, sess.graph)
# if eval_during_training:
eval_dir = os.path.join(model_dir, 'eval')
os.makedirs(eval_dir, exist_ok=True)
# eval_summary_writer = tf.summary.FileWriter(eval_dir)
sess.run(tf.global_variables_initializer())
training_time = 0.0
run_time_without_eval = 0.0
run_time_adv_ex_creation = 0.0
run_time_train_step = 0.0
####################################
# Main training loop
####################################
start_time = time.time()
no_epochs_done = 0 # the same as epoch_count, need to merge
start_epoch = timer()
it_count = 0
epoch_count = 0
acc_sum = 0
printFlag = 0
for ii in range(max_num_training_steps+1):
# original batch
x_batch, y_batch, epoch_done = data_iterator.train_data.get_next_batch(
num_ids, multiple_passes=True)
no_epochs_done += epoch_done
# noop trans
noop_trans = np.zeros([len(x_batch), 3])
# id_batch starts with IDs of original examples
id_batch = np.arange(0, num_ids, 1, dtype="int32")
if use_core:
print("*********Warning: should not be using core in train_reg.py!*********\n")
else:
if adversarial_training:
start = timer()
if simple_train:
# only generating 1 tilted image per original image, which is equivalent to wo-1
x_batch_inp = x_batch
# using the labels as the transformation, hope the model will learn it
trans_inp = y_batch
else:
print("shouldn't be entering here in regression task!\n")
quit()
x_batch_inp, trans_inp = attack.perturb(x_batch, y_batch,
sess)
end = timer()
training_time += end - start
run_time_without_eval += end - start
run_time_adv_ex_creation += end - start
else:
x_batch_inp, trans_inp = x_batch, noop_trans
# if simple_train:
# y_batch_inp = y_batch
# y_batch_adv = transform_parameters
# trans_adv = transform_parameters
# x_batch_adv = x_batch_inp
# id_batch_inp = id_batch
# id_batch_adv = id_batch
# for adversarial training and plain training, the following
# variables coincide
# else:
y_batch_inp = y_batch
y_batch_adv = y_batch
trans_adv = trans_inp
x_batch_adv = x_batch_inp
id_batch_inp = id_batch
id_batch_adv = id_batch
# feed_dict for training step
inp_dict = {model.x_input: x_batch_inp,
model.y_input: y_batch_inp,
model.group: id_batch_inp,
model.transform: trans_inp,
model.is_training: False}
# separate natural and adversarially transformed examples for eval
# nat_dict = {model.x_input: x_batch,
# model.y_input: y_batch,
# model.group: id_batch,
# model.transform: noop_trans,
# model.is_training: False}
#
# adv_dict = {model.x_input: x_batch_adv,
# model.y_input: y_batch_adv,
# model.group: id_batch_adv,
# model.transform: trans_adv,
# model.is_training: False}
loss = sess.run(model.reg_loss, feed_dict=inp_dict)
if ii % num_easyeval_steps == 0 or ii == max_num_training_steps:
print("\nin resnet_reg")
print("y_input")
y_input = sess.run(model.y_input, feed_dict=inp_dict)
print(y_input[0:5])
print("\nprediction")
prediction = sess.run(model.prediction, feed_dict=inp_dict)
print(prediction[0:5])
print(' Easy Evalutaion Step training time error Training Set ')
print('avg_abs_err_transX : {}'.format(sess.run(model.avg_abs_err_transX, feed_dict=inp_dict)))
print('avg_abs_err_transY : {}'.format(sess.run(model.avg_abs_err_transY, feed_dict=inp_dict)))
print('avg_abs_err_rot : {}'.format(sess.run(model.avg_abs_err_rot, feed_dict=inp_dict)))
print('avg_rel_err_transX : {}'.format(sess.run(model.avg_rel_err_transX, feed_dict=inp_dict)))
print('avg_rel_err_transY : {}'.format(sess.run(model.avg_rel_err_transY, feed_dict=inp_dict)))
print('avg_rel_err_rot : {}'.format(sess.run(model.avg_rel_err_rot, feed_dict=inp_dict)))
print(' Easy Evalutaion Step training time error Evaluation Set ')
print('avg_abs_err_transX : {}'.format(sess.run(model.avg_abs_err_transX, feed_dict=eval_dict)))
print('avg_abs_err_transY : {}'.format(sess.run(model.avg_abs_err_transY, feed_dict=eval_dict)))
print('avg_abs_err_rot : {}'.format(sess.run(model.avg_abs_err_rot, feed_dict=eval_dict)))
print('avg_rel_err_transX : {}'.format(sess.run(model.avg_rel_err_transX, feed_dict=eval_dict)))
print('avg_rel_err_transY : {}'.format(sess.run(model.avg_rel_err_transY, feed_dict=eval_dict)))
print('avg_rel_err_rot : {}'.format(sess.run(model.avg_rel_err_rot, feed_dict=eval_dict)))
# Output to stdout
if epoch_done:
epoch_time = timer() - start_epoch
# ToDo: Log this to file as well
# Training accuracy over epoch
print('Epoch {}: ({})'.format(epoch_count, datetime.now()))
print(' training loss {:.4}'.format(loss))
print(' {:.4} seconds per epoch'.format(epoch_time))
print(' training time error Training Set ')
print('avg_abs_err_transX : {}'.format(sess.run(model.avg_abs_err_transX, feed_dict=inp_dict)))
print('avg_abs_err_transY : {}'.format(sess.run(model.avg_abs_err_transY, feed_dict=inp_dict)))
print('avg_abs_err_rot : {}'.format(sess.run(model.avg_abs_err_rot, feed_dict=inp_dict)))
print('avg_rel_err_transX : {}'.format(sess.run(model.avg_rel_err_transX, feed_dict=inp_dict)))
print('avg_rel_err_transY : {}'.format(sess.run(model.avg_rel_err_transY, feed_dict=inp_dict)))
print('avg_rel_err_rot : {}'.format(sess.run(model.avg_rel_err_rot, feed_dict=inp_dict)))
# if ii % config.eval.full_batch_eval_steps == 0:
# print_eval_fullbatch(ii, sess, y_batch_eval, batch_size, trans_eval, model, attack, model_dir, global_step, summary_writer)
epoch_done = 0
epoch_count += 1
start_epoch = timer()
it_count = 1
else:
it_count += 1
# acc_sum += nat_acc_tr
# Output to stdout
if ii % num_output_steps == 0:
if ii != 0:
training_time = 0.0
# Tensorboard summaries and heavy checkpoints
if ii % num_summary_steps == 0:
summary = sess.run(err_summaries, feed_dict=inp_dict)
summary_writer.add_summary(summary, global_step.eval(sess))
# Write a checkpoint and eval if it's time
if ii % num_checkpoint_steps == 0 or ii == max_num_training_steps:
# Save checkpoint data (weights)
saver.save(sess,
os.path.join(model_dir, 'checkpoint'),
global_step=global_step)
full_eval_dict = fetch_full_eval_dict(model, batch_size)
print_eval_fullbatch(ii,sess, model, model_dir, full_eval_dict)
summary_eval = sess.run(eval_summaries, feed_dict=full_eval_dict)
summary_writer.add_summary(summary_eval, global_step.eval(sess))
# Evaluation on full evaluation batch
# if ((eval_during_training and ii % num_eval_steps == 0
# and ii > 0 and config.attack.use_spatial) or
# (eval_during_training and ii == max_num_training_steps and
# config.attack.use_spatial)):
# print_eval_fullbatch(ii,sess, y_batch_eval, batch_size, trans_eval, model, attack, model_dir, global_step, summary_writer)
# Actual training step
start = timer()
inp_dict[model.is_training] = True
sess.run(train_step, feed_dict=inp_dict)
end = timer()
training_time += end - start
run_time_without_eval += end - start
run_time_train_step += end - start
runtime = time.time() - start_time
# this_repo.mark_experiment_as_completed(
# runtime=runtime, runtime_wo_eval=run_time_without_eval,
# runtime_train_step=run_time_train_step,
# runtime_adv_ex_creation=run_time_adv_ex_creation)
return 0
def main():
# parse args and get configs
args = parser.parse_args()
logging.set_verbosity(logging.INFO)
random.seed(args.seed)
# reload model from checkpoint or train from scratch
if args.reload_ckpt != "None":
checkpoint_path = os.path.join(local_settings.MODEL_PATH,
args.all_checkpoints_folder)
checkpoint_folders = os.listdir(checkpoint_path)
checkpoint_folder = [
f for f in checkpoint_folders if args.reload_ckpt in f
]
if len(checkpoint_folder) == 0:
raise Exception("No matching folder found.")
elif len(checkpoint_folder) > 1:
logging.info(checkpoint_folder)
raise Exception("More than one matching folder found.")
else:
checkpoint_folder = checkpoint_folder[0]
logging.info("Restoring from {}".format(checkpoint_folder))
checkpoint_dir = os.path.join(checkpoint_path, checkpoint_folder)
if not args.overwrite_configs:
# reload configs from file
with open(os.path.join(checkpoint_dir, "hparams.pkl"), 'rb') as f:
config_dict = pickle.load(f)
else:
# get configs
config_dict = util.get_config(args.config)
config_dict = util.update_config(config_dict, args)
else:
# get configs
config_dict = util.get_config(args.config)
config_dict = util.update_config(config_dict, args)
config_dict_copy = copy.deepcopy(config_dict)
config = util.config_to_namedtuple(config_dict)
# Initialize the repo
logging.info("==> Creating repo..")
exp_repo = repo.ExperimentRepo(local_dir_name=config.local_json_dir_name,
root_dir=local_settings.MODEL_PATH)
if args.reload_ckpt != "None":
exp_id = config_dict["id"]
else:
exp_id = None
# Create new experiment
exp_id = exp_repo.create_new_experiment(config.dataset, config_dict_copy,
exp_id)
config_dict_copy["id"] = exp_id
# Set up model directory
current_time = datetime.datetime.now().strftime(r"%y%m%d_%H%M")
ckpt_dir_name = args.all_checkpoints_folder if not DEBUG else 'checkpoints_tmp'
ckpt_dir = os.path.join(local_settings.MODEL_PATH, ckpt_dir_name)
os.makedirs(ckpt_dir, exist_ok=True)
if args.reload_ckpt != "None":
model_dir = checkpoint_dir
else:
model_dir = os.path.join(ckpt_dir,
"ckpt_{}_{}".format(current_time, exp_id))
# Save hyperparameter settings
os.makedirs(model_dir, exist_ok=True)
if not os.path.exists(os.path.join(model_dir, "hparams.json")):
with open(os.path.join(model_dir, "hparams.json"), 'w') as f:
json.dump(config_dict_copy, f, indent=2, sort_keys=True)
with open(os.path.join(model_dir, "hparams.pkl"), 'wb') as f:
pickle.dump(config_dict_copy, f)
# Set optimizers
# learning_rate = tf.keras.optimizers.schedules.ExponentialDecay(
# config.learning_rate, config.decay_every,
# config.decay_base, staircase=True)
# learning rate = 0.02 in paper
optimizer = tf.keras.optimizers.RMSprop(learning_rate=0.02)
if args.reload_ckpt != "None":
# TODO: fix this hack
epoch_start = int(
sorted([f for f in os.listdir(checkpoint_dir)
if 'ckpt-' in f])[-1].split('ckpt-')[1].split('.')[0])
init_gs = 0
else:
epoch_start = 0
init_gs = 0
global_step = tf.Variable(initial_value=init_gs,
name="global_step",
trainable=False,
dtype=tf.int64)
# Get model
model_domain = get_model(config.name_classifier_domain, config)
model_label = get_model(config.name_classifier_label, config)
# Get datasets
if DEBUG:
num_batches = 5
else:
num_batches = None
ds_train1 = _get_dataset(
config.dataset,
model_label,
split=tfds.Split.TRAIN.subsplit(tfds.percent[:50]),
batch_size=tf.cast(config.batch_size / 2, tf.int64),
num_batches=num_batches,
domain=tf.constant(0),
e=0.2)
ds_train2 = _get_dataset(
config.dataset,
model_label,
split=tfds.Split.TRAIN.subsplit(tfds.percent[-50:]),
batch_size=tf.cast(config.batch_size / 2, tf.int64),
num_batches=num_batches,
domain=tf.constant(1),
e=0.1)
ds_val = _get_dataset(config.dataset,
model_label,
split=tfds.Split.TEST,
batch_size=config.batch_size,
num_batches=num_batches,
domain=tf.constant(2),
e=0.9)
# TODO: add test set - done
show_inputs = iter(ds_train1)
_ = model_label(next(show_inputs)["image"])
# Set up checkpointing
if args.reload_ckpt != "None":
ckpt = tf.train.Checkpoint(model=model_label, global_step=global_step)
manager = tf.train.CheckpointManager(ckpt,
checkpoint_dir,
max_to_keep=3)
status = ckpt.restore(manager.latest_checkpoint)
status.assert_consumed()
else:
ckpt = tf.train.Checkpoint(model=model_label, global_step=global_step)
manager = tf.train.CheckpointManager(ckpt, model_dir, max_to_keep=3)
writer = tf.summary.create_file_writer(manager._directory)
with writer.as_default(), tf.summary.record_if(
lambda: int(global_step.numpy()) % 100 == 0):
for epoch in range(epoch_start, config.num_epochs):
start_time = time.time()
# random = np.array([0, 1, 2])
# np.random.shuffle(random)
# rand_inputs = [ds_train1, ds_train2, ds_train3]
train_one_epoch(model_domain=model_domain,
model_label=model_label,
train_input1=ds_train1,
train_input2=ds_train2,
optimizer=optimizer,
global_step=global_step,
config=config)
train1_metr = eval_one_epoch(model_label=model_label,
dataset=ds_train1,
summary_directory=os.path.join(
manager._directory, "train1"),
global_step=global_step,
config=config,
training=False)
train2_metr = eval_one_epoch(model_label=model_label,
dataset=ds_train2,
summary_directory=os.path.join(
manager._directory, "train2"),
global_step=global_step,
config=config,
training=False)
val_metr = eval_one_epoch(model_label=model_label,
dataset=ds_val,
summary_directory=os.path.join(
manager._directory, "val"),
global_step=global_step,
config=config,
training=False)
# if epoch == (config.num_epochs - 1):
# # full training set
# train_metr = eval_one_epoch(model_classifier=model_classifier, dataset=ds_train_complete,
# summary_directory=os.path.join(manager._directory, "train"),
# global_step=global_step, config=config, training=False)
# # full test_out set
# test_out_metr = eval_one_epoch(model_classifier=model_classifier, dataset=ds_val_out,
# summary_directory=os.path.join(manager._directory, "val_out"),
# global_step=global_step, config=config, training=False)
# # full test_in set
# test_in_metr = eval_one_epoch(model_classifier=model_classifier, dataset=ds_val_in,
# summary_directory=os.path.join(manager._directory, "val_in"),
# global_step=global_step, config=config, training=False)
manager.save()
logging.info("\n #### \n epoch: %d, time: %0.2f" %
(epoch, time.time() - start_time))
logging.info("Global step: {}".format(global_step.numpy()))
logging.info("train1_accuracy: {:2f}, train1_loss: {:4f}".format(
train1_metr['accuracy'], train1_metr['loss']))
logging.info("train2_accuracy: {:2f}, train2_loss: {:4f}".format(
train2_metr['accuracy'], train2_metr['loss']))
logging.info("val_accuracy: {:2f}, val_loss: {:4f}".format(
val_metr['accuracy'], val_metr['loss']))
if epoch == epoch_start:
dir_path = os.path.dirname(os.path.realpath(__file__))
copy_source(dir_path, manager._directory)
# Mark experiment as completed
# TODO: add other metrics - done
exp_repo.mark_experiment_as_completed(
exp_id,
train1_accuracy=train1_metr['accuracy'],
train2_accuracy=train2_metr['accuracy'],
val_accuracy=val_metr['accuracy'])
def train(config='configs/fannyconfig.json',
save_root_path='/cluster/work/math/fanyang-broglil/CoreRepo',
experiment_json_fname='experiments.json',
local_json_dir_name='local_json_files',
worstofk=None,
attack_style=None,
attack_limits=None,
fo_epsilon=None,
fo_step_size=None,
fo_num_steps=None,
lambda_core=None,
num_ids=None,
group_size=None,
use_core=None,
seed=None,
save_in_local_json=True,
this_repo=None):
# reset default graph (needed for running locally with run_jobs_ray.py)
tf.reset_default_graph()
# get configs
config_dict = utilities.get_config(config)
config_dict_copy = copy.deepcopy(config_dict)
config = utilities.config_to_namedtuple(config_dict)
# seeding randomness
if seed == None:
seed = config.training.tf_random_seed
else:
config_dict_copy['training']['tf_random_seed'] = seed
tf.set_random_seed(seed)
np.random.seed(seed)
# Setting up training parameters
max_num_training_steps = config.training.max_num_training_steps
step_size_schedule = config.training.step_size_schedule
weight_decay = config.training.weight_decay
momentum = config.training.momentum
if group_size == None:
group_size = config.training.group_size
else:
config_dict_copy['training']['group_size'] = int(group_size)
if num_ids == None:
num_ids = config.training.num_ids
else:
config_dict_copy['training']['num_ids'] = int(num_ids)
if lambda_core == None:
lambda_core = config.training.lambda_
else:
config_dict_copy['training']['lambda_'] = float(lambda_core)
if use_core == None:
use_core = config.training.use_core
else:
config_dict_copy['training']['use_core'] = use_core
batch_size = config.training.batch_size
# number of groups with group size > 1
num_grouped_ids = batch_size - num_ids
# number of unique ids needs to be larger than half the desired batch size
# so that full batch can be filled up
assert num_ids >= batch_size / group_size
# currently, code is designed for groups of size 2
# assert batch_size % group_size == 0
adversarial_training = config.training.adversarial_training
eval_during_training = config.training.eval_during_training
if eval_during_training:
num_eval_steps = config.training.num_eval_steps
# Setting up output parameters
num_output_steps = config.training.num_output_steps
num_summary_steps = config.training.num_summary_steps
num_checkpoint_steps = config.training.num_checkpoint_steps
num_easyeval_steps = config.training.num_easyeval_steps
# Setting up the data and the model
data_path = config.data.data_path
if config.data.dataset_name == "cifar-10":
raw_iterator = cifar10_input.CIFAR10Data(data_path)
elif config.data.dataset_name == "cifar-100":
raw_iterator = cifar100_input.CIFAR100Data(data_path)
elif config.data.dataset_name == "svhn":
raw_iterator = svhn_input.SVHNData(data_path)
else:
raise ValueError("Unknown dataset name.")
global_step = tf.train.get_or_create_global_step()
model_family = config.model.model_family
if model_family == "resnet":
if config.attack.use_spatial and config.attack.spatial_method == 'fo':
diffable = True
else:
diffable = False
model = resnet.Model(config.model, num_ids, diffable)
elif model_family == "vgg":
if config.attack.use_spatial and config.attack.spatial_method == 'fo':
# TODO: add differentiable transformer to vgg.py
raise NotImplementedError
model = vgg.Model(config.model, num_ids)
# uncomment to get a list of trainable variables
# model_vars = tf.trainable_variables()
# Setting up the optimizer
boundaries = [int(sss[0]) for sss in step_size_schedule]
boundaries = boundaries[1:]
values = [sss[1] for sss in step_size_schedule]
learning_rate = tf.train.piecewise_constant(tf.cast(global_step, tf.int32),
boundaries, values)
if use_core and lambda_core > 0:
total_loss = (model.mean_xent +
weight_decay * model.weight_decay_loss +
lambda_core * model.core_loss)
else:
total_loss = model.mean_xent + weight_decay * model.weight_decay_loss
optimizer = tf.train.MomentumOptimizer(learning_rate, momentum)
train_step = optimizer.minimize(total_loss, global_step=global_step)
# Set up adversary
if worstofk == None:
worstofk = config.attack.random_tries
else:
config_dict_copy['attack']['random_tries'] = worstofk
if fo_epsilon == None:
fo_epsilon = config.attack.epsilon
else:
config_dict_copy['attack']['epsilon'] = fo_epsilon
if fo_step_size == None:
fo_step_size = config.attack.step_size
else:
config_dict_copy['attack']['step_size'] = fo_step_size
if fo_num_steps == None:
fo_num_steps = config.attack.num_steps
else:
config_dict_copy['attack']['num_steps'] = fo_num_steps
# @ Luzius: incorporate being able to choose multiple transformations
if attack_style == None:
attack_style = 'rotate'
# Training attack
# L-inf attack if use_spatial is False and use_linf is True
# spatial attack if use_spatial is True and use_linf is False
# spatial random attack if spatial_method is 'random'
# spatial PGD attack if spatial_method is 'fo'
attack = SpatialAttack(model, config.attack, config.attack.spatial_method,
worstofk, attack_limits, fo_epsilon, fo_step_size,
fo_num_steps)
# Different eval attacks
# Random attack
# L-inf attack if use_spatial is False and use_linf is True
# random (worst-of-1) spatial attack if use_spatial is True
# and use_linf is False
attack_eval_random = SpatialAttack(model, config.attack, 'random', 1,
attack_limits, fo_epsilon, fo_step_size,
fo_num_steps)
# First order attack
# L-inf attack if use_spatial is False and use_linf is True
# first-order spatial attack if use_spatial is True and use_linf is False
attack_eval_fo = SpatialAttack(model, config.attack, 'fo', 1,
attack_limits, fo_epsilon, fo_step_size,
fo_num_steps)
# Grid attack
# spatial attack if use_spatial is True and use_linf is False
# not executed for L-inf attacks
attack_eval_grid = SpatialAttack(model, config.attack, 'grid', None,
attack_limits)
# TODO(christina): add L-inf attack with random restarts
# ------------------START EXPERIMENT -------------------------
# Initialize the Repo
print("==> Creating repo..")
# Create repo object if it wasn't passed, comment out if repo has issues
if this_repo == None:
this_repo = exprepo.ExperimentRepo(
save_in_local_json=save_in_local_json,
json_filename=experiment_json_fname,
local_dir_name=local_json_dir_name,
root_dir=save_root_path)
# Create new experiment
if this_repo != None:
exp_id = this_repo.create_new_experiment(config.data.dataset_name,
model_family, worstofk,
attack_style, attack_limits,
lambda_core, num_grouped_ids,
group_size, config_dict_copy)
# Setting up the Tensorboard and checkpoint outputs
model_dir = '%s/logdir/%s' % (save_root_path, exp_id)
# We add accuracy and xent twice so we can easily make three types of
# comparisons in Tensorboard:
# - train vs eval (for a single run)
# - train of different runs
# - eval of different runs
saver = tf.train.Saver(max_to_keep=3)
tf.summary.scalar('accuracy_nat_train',
model.accuracy,
collections=['nat'])
tf.summary.scalar('accuracy_nat', model.accuracy, collections=['nat'])
tf.summary.scalar('xent_nat_train',
model.xent / batch_size,
collections=['nat'])
tf.summary.scalar('xent_nat', model.xent / batch_size, collections=['nat'])
tf.summary.image('images_nat_train', model.x_image, collections=['nat'])
tf.summary.scalar('learning_rate', learning_rate, collections=['nat'])
nat_summaries = tf.summary.merge_all('nat')
# data augmentation used if config.training.data_augmentation_core is True
x_input_placeholder = tf.placeholder(tf.float32, shape=[None, 32, 32, 3])
flipped = tf.map_fn(lambda img: tf.image.random_flip_left_right(img),
x_input_placeholder)
with tf.Session() as sess:
# initialize standard data augmentation
if config.training.data_augmentation:
if config.data.dataset_name == "cifar-10":
data_iterator = cifar10_input.AugmentedCIFAR10Data(
raw_iterator, sess)
elif config.data.dataset_name == "cifar-100":
data_iterator = cifar100_input.AugmentedCIFAR100Data(
raw_iterator, sess)
elif config.data.dataset_name == "svhn":
data_iterator = svhn_input.AugmentedSVHNData(
raw_iterator, sess)
else:
raise ValueError("Unknown dataset name.")
else:
data_iterator = raw_iterator
eval_dict = {
model.x_input: data_iterator.eval_data.xs,
model.y_input: data_iterator.eval_data.ys,
model.group: np.arange(0, batch_size, 1, dtype="int32"),
model.transform: np.zeros([data_iterator.eval_data.n, 3]),
model.is_training: False
}
# Initialize the summary writer, global variables, and our time counter.
summary_writer = tf.summary.FileWriter(model_dir, sess.graph)
# if eval_during_training:
eval_dir = os.path.join(model_dir, 'eval')
os.makedirs(eval_dir, exist_ok=True)
eval_summary_writer = tf.summary.FileWriter(eval_dir)
sess.run(tf.global_variables_initializer())
training_time = 0.0
run_time_without_eval = 0.0
run_time_adv_ex_creation = 0.0
run_time_train_step = 0.0
####################################
# Main training loop
####################################
start_time = time.time()
no_epochs_done = 0 # the same as epoch_count, need to merge
start_epoch = timer()
it_count = 0
epoch_count = 0
acc_sum = 0
for ii in range(max_num_training_steps + 1):
# original batch
x_batch, y_batch, epoch_done = data_iterator.train_data.get_next_batch(
num_ids, multiple_passes=True)
no_epochs_done += epoch_done
# noop trans
noop_trans = np.zeros([len(x_batch), 3])
# id_batch starts with IDs of original examples
id_batch = np.arange(0, num_ids, 1, dtype="int32")
if use_core:
# first num_id examples of batch are natural
x_batch_inp = x_batch
y_batch_inp = y_batch
trans_inp = noop_trans
id_batch_inp = id_batch
start = timer()
for _ in range(group_size - 1):
if config.training.data_augmentation_core:
raise NotImplementedError
# create rotated examples
x_batch_adv_i, trans_adv_i = attack.perturb(
x_batch, y_batch, sess)
# construct new batches including rotated examples
x_batch_inp = np.concatenate((x_batch_inp, x_batch_adv_i),
axis=0)
y_batch_inp = np.concatenate((y_batch_inp, y_batch),
axis=0)
trans_inp = np.concatenate((trans_inp, trans_adv_i),
axis=0)
id_batch_inp = np.concatenate((id_batch_inp, id_batch),
axis=0)
end = timer()
training_time += end - start
run_time_without_eval += end - start
run_time_adv_ex_creation += end - start
trans_adv = trans_inp[num_ids:, ...]
id_batch_adv = id_batch_inp[num_ids:]
y_batch_adv = y_batch_inp[num_ids:]
x_batch_adv = x_batch_inp[num_ids:, ...]
else:
if adversarial_training:
start = timer()
x_batch_inp, trans_inp = attack.perturb(
x_batch, y_batch, sess)
end = timer()
training_time += end - start
run_time_without_eval += end - start
run_time_adv_ex_creation += end - start
else:
x_batch_inp, trans_inp = x_batch, noop_trans
# for adversarial training and plain training, the following
# variables coincide
y_batch_inp = y_batch
y_batch_adv = y_batch
trans_adv = trans_inp
x_batch_adv = x_batch_inp
id_batch_inp = id_batch
id_batch_adv = id_batch
# feed_dict for training step
inp_dict = {
model.x_input: x_batch_inp,
model.y_input: y_batch_inp,
model.group: id_batch_inp,
model.transform: trans_inp,
model.is_training: False
}
# separate natural and adversarially transformed examples for eval
nat_dict = {
model.x_input: x_batch,
model.y_input: y_batch,
model.group: id_batch,
model.transform: noop_trans,
model.is_training: False
}
adv_dict = {
model.x_input: x_batch_adv,
model.y_input: y_batch_adv,
model.group: id_batch_adv,
model.transform: trans_adv,
model.is_training: False
}
########### Outputting/saving weights and evaluations ###########
acc_grid_te = -1.0
avg_xent_grid_te = -1.0
acc_fo_te = -1.0
avg_xent_fo_te = -1.0
saved_weights = 0
# Compute training accuracy on this minibatch
nat_acc_tr = 100 * sess.run(model.accuracy, feed_dict=nat_dict)
# Output to stdout
if epoch_done:
epoch_time = timer() - start_epoch
# Average
av_acc = acc_sum / it_count
# ToDo: Log this to file as well
# Training accuracy over epoch
print('Epoch {}: ({})'.format(epoch_count, datetime.now()))
print(' training natural accuracy {:.4}%'.format(av_acc))
print(' {:.4} seconds per epoch'.format(epoch_time))
# Accuracy on entire test set
nat_acc_te = 100 * sess.run(model.accuracy,
feed_dict=eval_dict)
print(
' test set natural accuracy {:.4}%'.format(nat_acc_te))
# Set loss sum, it count back to zero
acc_sum = nat_acc_tr
epoch_done = 0
epoch_count += 1
start_epoch = timer()
it_count = 1
else:
it_count += 1
acc_sum += nat_acc_tr
# Output to stdout
if ii % num_output_steps == 0:
# nat_acc_tr = 100 * sess.run(model.accuracy, feed_dict=nat_dict)
adv_acc_tr = 100 * sess.run(model.accuracy, feed_dict=adv_dict)
inp_acc_tr = 100 * sess.run(model.accuracy, feed_dict=inp_dict)
# print('Step {}: ({})'.format(ii, datetime.now()))
# print(' training nat accuracy {:.4}%'.format(nat_acc_tr))
# print(' training adv accuracy {:.4}%'.format(adv_acc_tr))
# print(' training inp accuracy {:.4}%'.format(inp_acc_tr))
if ii != 0:
# print(' {} examples per second'.format(
# num_output_steps * batch_size / training_time))
training_time = 0.0
# Tensorboard summaries and heavy checkpoints
if ii % num_summary_steps == 0:
summary = sess.run(nat_summaries, feed_dict=nat_dict)
summary_writer.add_summary(summary, global_step.eval(sess))
# Write a checkpoint and eval if it's time
if ii % num_checkpoint_steps == 0 or ii == max_num_training_steps:
# Save checkpoint data (weights)
saver.save(sess,
os.path.join(model_dir, 'checkpoint'),
global_step=global_step)
saved_weights = 1
# Write evaluation meta data for checkpoint
if ii % num_easyeval_steps == 0 or ii == max_num_training_steps:
# Get training accuracies
nat_acc_tr = 100 * sess.run(model.accuracy, feed_dict=nat_dict)
adv_acc_tr = 100 * sess.run(model.accuracy, feed_dict=adv_dict)
inp_acc_tr = 100 * sess.run(model.accuracy, feed_dict=inp_dict)
# Evaluation on random and natural
[
acc_nat_te, acc_rand_adv_te, avg_xent_nat_te,
avg_xent_adv_te
] = evaluate(model, attack_eval_random, sess, config, 'random',
data_path, None)
# Evaluation on grid (only for spatial attacks)
if ((eval_during_training and ii % num_eval_steps == 0
and ii > 0 and config.attack.use_spatial) or
(eval_during_training and ii == max_num_training_steps
and config.attack.use_spatial)):
if config.attack.use_spatial and config.attack.spatial_method == 'fo':
# Evaluation on first-order PDG attack (too expensive to
# evaluate more frequently on whole dataset)
[_, acc_fo_te, _,
avg_xent_fo_te] = evaluate(model, attack_eval_fo,
sess, config, 'fo',
data_path, None)
# Evaluation on grid
[_, acc_grid_te, _, avg_xent_grid_te
] = evaluate(model, attack_eval_grid, sess, config,
"grid", data_path, eval_summary_writer)
chkpt_id = this_repo.create_training_checkpoint(
exp_id,
training_step=ii,
epoch=no_epochs_done,
train_acc_nat=nat_acc_tr,
train_acc_adv=adv_acc_tr,
train_acc_inp=inp_acc_tr,
test_acc_nat=acc_nat_te,
test_acc_adv=acc_rand_adv_te,
test_acc_fo=acc_fo_te,
test_acc_grid=acc_grid_te,
test_loss_nat=avg_xent_nat_te,
test_loss_adv=avg_xent_adv_te,
test_loss_fo=avg_xent_fo_te,
test_loss_grid=avg_xent_grid_te)
if saved_weights == 0:
# Save checkpoint data (weights)
saver.save(
sess,
os.path.join(model_dir,
'{}_checkpoint'.format(chkpt_id)))
# Actual training step
start = timer()
inp_dict[model.is_training] = True
sess.run(train_step, feed_dict=inp_dict)
end = timer()
training_time += end - start
run_time_without_eval += end - start
run_time_train_step += end - start
runtime = time.time() - start_time
this_repo.mark_experiment_as_completed(
exp_id,
train_acc_nat=nat_acc_tr,
train_acc_adv=adv_acc_tr,
train_acc_inp=inp_acc_tr,
test_acc_nat=acc_nat_te,
test_acc_adv=acc_rand_adv_te,
test_acc_fo=acc_fo_te,
test_acc_grid=acc_grid_te,
runtime=runtime,
runtime_wo_eval=run_time_without_eval,
runtime_train_step=run_time_train_step,
runtime_adv_ex_creation=run_time_adv_ex_creation)
return 0
def train(config='configs/cifar10_config_stn.json',
save_root_path='/cluster/work/math/fanyang-broglil/CoreRepo',
worstofk=None,
attack_style=None,
attack_limits=None,
lambda_core=None,
num_grouped_ids=None,
num_ids=None,
group_size=None,
use_core=None,
seed=None,
this_repo=None):
config_dict = utilities.get_config(config)
config_dict_copy = copy.deepcopy(config_dict)
# model_dir = config_dict['model']['output_dir']
# if not os.path.exists(model_dir):
# os.makedirs(model_dir)
# # keep the configuration file with the model for reproducibility
# with open(os.path.join(model_dir, 'config.json'), 'w') as f:
# json.dump(config_dict, f, sort_keys=True, indent=4)
config = utilities.config_to_namedtuple(config_dict)
# seeding randomness
if seed == None:
seed = config.training.tf_random_seed
else:
config_dict_copy['training']['tf_random_seed'] = seed
tf.set_random_seed(seed)
np.random.seed(seed)
# Setting up training parameters
max_num_epochs = config.training.max_num_epochs
step_size_schedule = config.training.step_size_schedule
weight_decay = config.training.weight_decay
momentum = config.training.momentum
num_ids = config.training.num_ids # number of IDs per minibatch
if group_size == None:
group_size = config.training.group_size
else:
config_dict_copy['training']['group_size'] = group_size
if num_grouped_ids == None:
num_grouped_ids = config.training.num_grouped_ids
else:
config_dict_copy['training']['num_grouped_ids'] = num_grouped_ids
if num_ids == None:
num_ids = config.training.num_ids
else:
config_dict_copy['training']['num_ids'] = num_ids
if lambda_core == None:
lambda_core = config.training.lambda_
else:
config_dict_copy['training']['lambda_'] = lambda_core
if use_core == None:
use_core = config.training.use_core
else:
config_dict_copy['training']['use_core'] = use_core
adversarial_training = config.training.adversarial_training
eval_during_training = config.training.eval_during_training
if eval_during_training:
num_eval_steps = config.training.num_eval_steps
# Setting up output parameters
num_summary_steps = config.training.num_summary_steps
num_checkpoint_steps = config.training.num_checkpoint_steps
num_easyeval_steps = config.training.num_easyeval_steps
# mini batch size per iteration
# ToDo: need to make this support variable number of num_grouped_ids
batch_size = num_ids + num_grouped_ids
# Setting up model and loss
model_family = config.model.model_family
with_transformer = config.model.transformer
translation_model = config.model.translation_model
if model_family == "resnet":
model = loc_net.Model(config.model)
else:
print("Model family does not exist")
exit()
if use_core:
total_loss = model.y_loss #model.mean_xent + weight_decay * model.weight_decay_loss + lambda_core * model.core_loss2
else:
total_loss = model.y_loss #model.mean_xent + weight_decay * model.weight_decay_loss
# Setting up the data and the model
data_path = config.data.data_path
if config.data.dataset_name == "cifar-10":
raw_cifar = cifar10_input.CIFAR10Data(data_path)
elif config.data.dataset_name == "cifar-100":
raw_cifar = cifar100_input.CIFAR100Data(data_path)
else:
raise ValueError("Unknown dataset name.")
# uncomment to get a list of trainable variables
# model_vars = tf.trainable_variables()
# slim.model_analyzer.analyze_vars(model_vars, print_info=True)
# Setting up the optimizer
boundaries = [int(sss[0]) for sss in step_size_schedule]
boundaries = boundaries[1:]
values = [sss[1] for sss in step_size_schedule]
global_step = tf.train.get_or_create_global_step()
learning_rate = tf.train.piecewise_constant(tf.cast(global_step, tf.int32),
boundaries, values)
optimizer = tf.train.MomentumOptimizer(learning_rate, momentum)
#optimizer = tf.train.AdamOptimizer(learning_rate=learning_rate,
# name="Adam")
train_step = optimizer.minimize(total_loss, global_step=global_step)
# Set up adversary
if worstofk == None:
worstofk = config.attack.random_tries
else:
config_dict_copy['attack']['random_tries'] = worstofk
# @ Luzius: incorporate being able to choose multiple transformations
if attack_style == None:
attack_style = 'rotate'
# Training attack
attack = SpatialAttack(model, config.attack, 'random', worstofk,
attack_limits)
# Different eval attacks
# Same attack as worstofk
# @ Luzius: currently the names are not clear/consistent since I wasn't sure if we actually want random or not since you originally had your attack like that but I feel like it should rather be worstofk?
# attack_eval_adv = SpatialAttack(model, config.attack, 'random', worstofk, attack_limits)
attack_eval_random = SpatialAttack(model, config.attack, 'random', 1,
attack_limits)
# Grid attack
attack_eval_grid = SpatialAttack(model, config.attack, 'grid', None,
attack_limits)
# ------------------START EXPERIMENT -------------------------
# Initialize the Repo
print("==> Creating repo..")
# Create repo object if it wasn't passed, comment out if repo has issues
if this_repo == None:
this_repo = exprepo.ExperimentRepo(root_dir=save_root_path)
# Create new experiment
if this_repo != None:
exp_id = this_repo.create_new_experiment('cifar-10', model_family,
worstofk, attack_style,
attack_limits, lambda_core,
num_grouped_ids, group_size,
config_dict_copy)
# Setting up the Tensorboard and checkpoint outputs
model_dir = '%s/logdir/%s' % (save_root_path, exp_id)
os.makedirs(model_dir, exist_ok=True)
# We add accuracy and xent twice so we can easily make three types of
# comparisons in Tensorboard:
# - train vs eval (for a single run)
# - train of different runs
# - eval of different runs
saver = tf.train.Saver(max_to_keep=3)
tf.summary.scalar('accuracy_nat_train',
model.accuracy,
collections=['nat'])
tf.summary.scalar('accuracy_nat', model.accuracy, collections=['nat'])
tf.summary.scalar('xent_nat_train',
model.xent / batch_size,
collections=['nat'])
tf.summary.scalar('xent_nat', model.xent / batch_size, collections=['nat'])
tf.summary.image('images_nat_train', model.x_image, collections=['nat'])
tf.summary.scalar('learning_rate', learning_rate, collections=['nat'])
nat_summaries = tf.summary.merge_all('nat')
#dataAugmentation
x_input_placeholder = tf.placeholder(tf.float32, shape=[None, 32, 32, 3])
flipped = tf.map_fn(lambda img: tf.image.random_flip_left_right(img),
x_input_placeholder)
tot_samp = raw_cifar.train_data.n
max_num_iterations = int(np.floor((tot_samp / num_ids) * max_num_epochs))
print("Total # of samples is: %d; This exp. will run %d iterations" %
(tot_samp, max_num_iterations))
# Compute the (epoch) gaps between summary, worstof1eval, checkpoints should happen
summary_gap = int(np.floor(max_num_epochs / num_summary_steps))
easyeval_gap = int(np.floor(max_num_epochs / num_easyeval_steps))
checkpoint_gap = int(np.floor(max_num_epochs / num_checkpoint_steps))
with tf.Session() as sess:
# initialize data augmentation
if config.training.data_augmentation:
if config.data.dataset_name == "cifar-10":
cifar = cifar10_input.AugmentedCIFAR10Data(raw_cifar, sess)
elif config.data.dataset_name == "cifar-100":
cifar = cifar100_input.AugmentedCIFAR100Data(raw_cifar, sess)
else:
raise ValueError("Unknown dataset name.")
else:
cifar = raw_cifar
cifar_eval_dict = {
model.x_input: cifar.eval_data.xs,
model.y_input: cifar.eval_data.ys,
model.group: np.arange(0, batch_size, 1, dtype="int32"),
model.transform: np.zeros([cifar.eval_data.n, 3]),
model.is_training: False
}
# Initialize the summary writer, global variables, and our time counter.
summary_writer = tf.summary.FileWriter(model_dir, sess.graph)
#if eval_during_training:
eval_dir = os.path.join(model_dir, 'eval')
os.makedirs(eval_dir, exist_ok=True)
eval_summary_writer = tf.summary.FileWriter(eval_dir)
sess.run(tf.global_variables_initializer())
training_time = 0.0
####################################
# Main training loop
####################################
# Initialize cache variables
start_time = time.time()
start_epoch = timer()
it_count = 0
epoch_count = 0
acc_sum = 0
it_summary = 0
it_easyeval = 0
it_ckpt = 0
adv_time = 0
train_time = 0
for ii in range(max_num_iterations + 1):
x_batch, y_batch, epoch_done = cifar.train_data.get_next_batch(
num_ids, multiple_passes=True)
noop_trans = np.zeros([len(x_batch), 3])
x_batch_nat = x_batch
y_batch_nat = y_batch
id_batch = np.arange(0, num_ids, 1, dtype="int32")
if use_core:
# Create rotated examples
start = timer()
ids = np.arange(0, num_grouped_ids, 1, dtype="int32")
for i in range(config.training.group_size):
if config.training.data_augmentation_core:
x_batch_core = sess.run(
flipped,
feed_dict={
x_input_placeholder:
x_batch[0:num_grouped_ids, :, :, :]
})
else:
x_batch_core = x_batch[0:num_grouped_ids, :, :, :]
x_batch_group, trans_group = attack.perturb(
x_batch_core, y_batch[0:num_grouped_ids], sess)
#construct new batches including rotated examples
x_batch_nat = np.concatenate((x_batch_nat, x_batch_group),
axis=0)
y_batch_nat = np.concatenate((y_batch_nat, y_batch),
axis=0)
noop_trans = np.concatenate((noop_trans, trans_group),
axis=0)
id_batch = np.concatenate((id_batch, ids), axis=0)
end = timer()
training_time += end - start
adv_time += end - start
else:
if adversarial_training:
start = timer()
x_batch_nat, noop_trans = attack.perturb(
x_batch, y_batch, sess)
end = timer()
adv_time += end - start
else:
x_batch_nat, noop_trans = x_batch, noop_trans
nat_dict = {
model.x_input: x_batch_nat,
model.y_input: y_batch_nat,
model.group: id_batch,
model.transform: noop_trans,
model.is_training: False
}
################# Outputting/saving weights and evaluations ###############
nat_acc = -1.0
acc_grid = -1.0
avg_xent_grid = -1.0
saved_weights = 0
# Compute training accuracy on this minibatch
train_nat_acc = sess.run(model.accuracy, feed_dict=nat_dict)
# Output to stdout
if epoch_done:
epoch_time = timer() - start_epoch
# Average
av_acc = acc_sum / it_count
# ToDo: Log this to file as well
# Training accuracy over epoch
print('Epoch {}: ({})'.format(epoch_count, datetime.now()))
print(' training natural accuracy {:.4}%'.format(av_acc *
100))
print(' {:.4} seconds per epoch'.format(epoch_time))
# Accuracy on entire test set
test_nat_acc = sess.run(model.accuracy,
feed_dict=cifar_eval_dict)
print(' test set natural accuracy {:.4}%'.format(
test_nat_acc * 100))
# print(' {:.4} seconds for test evaluation'.format(test_time))
print("example TIME")
print(adv_time)
print("train TIME")
print(train_time)
########### Things to do every xxx epochs #############
# Check if worstof1 eval should be run
if it_summary == summary_gap - 1 or epoch_count == max_num_epochs - 1:
summary = sess.run(nat_summaries, feed_dict=nat_dict)
summary_writer.add_summary(summary, global_step.eval(sess))
it_summary = 0
else:
it_summary += 1
if it_easyeval == easyeval_gap - 1 or epoch_count == max_num_epochs - 1:
# Evaluation on adv and natural
[acc_nat, acc_adv, avg_xent_nat,
avg_xent_adv] = evaluate(model, attack_eval_random, sess,
config, "random", data_path,
None)
# Save in checkpoint
chkpt_id = this_repo.create_training_checkpoint(
exp_id,
training_step=ii,
epoch=epoch_count,
train_acc_nat=nat_acc,
test_acc_adv=acc_adv,
test_acc_nat=acc_nat,
test_loss_adv=avg_xent_adv,
test_loss_nat=avg_xent_nat)
it_easyeval = 0
else:
it_easyeval += 1
startt = timer()
if it_ckpt == checkpoint_gap - 1 or epoch_count == max_num_epochs - 1:
# Create checkpoint id if non-existent
if not chkpt_id:
chkpt_id = this_repo.create_training_checkpoint(
exp_id,
training_step=ii,
epoch=epoch_count,
train_acc_nat=train_nat_acc,
test_acc_nat=test_nat_acc)
# Save checkpoint data (weights)
saver.save(
sess,
os.path.join(model_dir,
'{}_checkpoint'.format(chkpt_id)))
print(' chkpt saving took {:.4}s '.format(timer() -
startt))
it_ckpt = 0
else:
it_ckpt += 1
# Set loss sum, it count back to zero
acc_sum = train_nat_acc
epoch_done = 0
epoch_count += 1
start_epoch = timer()
it_count = 1
else:
it_count += 1
acc_sum += train_nat_acc
# Actual training step
start = timer()
nat_dict[model.is_training] = True
sess.run(train_step, feed_dict=nat_dict)
training_time += timer() - start
train_time += timer() - start
runtime = time.time() - start_time
# Do all evaluations in the last step - on grid
[_, acc_grid, _,
avg_xent_grid] = evaluate(model, attack_eval_grid, sess, config,
"grid", data_path, eval_summary_writer)
this_repo.mark_experiment_as_completed(exp_id,
train_acc_nat=nat_acc,
test_acc_adv=acc_adv,
test_acc_nat=acc_nat,
test_acc_grid=acc_grid,
runtime=runtime)
return 0