def main():
parser = argparse.ArgumentParser()
parser.add_argument('--input_dir', required=True, type=str, \
help='Input directory with images.')
parser.add_argument('--output_dir', required=True, type=str, \
help='Output directory with images.')
parser.add_argument('--max_epsilon', default=16.0, type=float, \
help='Maximum size of adversarial perturbation.')
parser.add_argument('--input_dir_mode', default='test', type=str, \
help='Either flat or hierarchy, how the input dir is organised')
parser.add_argument('--meta', default='/dev/shm/dev_dataset.csv', type=str, \
help='True labels for dev set')
parser.add_argument('--num_samples', default=-1, type=int, \
help='Number of samples, -1 for all samples')
parser.add_argument('--train_batch_size', default=24, type=int, \
help='How many images process at one time.')
args = parser.parse_args()
input_dir_abs = os.path.abspath(args.input_dir)
if args.input_dir_mode == 'flat':
df_meta = pd.read_csv(args.meta)
df_meta = df_meta[['ImageId', 'TrueLabel']]
df_meta['ImageId'] = os.path.abspath(
args.input_dir) + '/' + df_meta['ImageId'] + '.png'
df_meta['TrueLabel'] = df_meta['TrueLabel'] - 1
df_meta = df_meta.set_index('ImageId')
meta_dict = df_meta.to_dict()['TrueLabel']
images_info = image_generators.get_names_and_labels(\
args.input_dir, mode='flat', meta_dict=meta_dict)
elif args.input_dir_mode == 'flat_targeted':
df_meta = pd.read_csv(args.meta)
df_meta = df_meta[['ImageId', 'TrueLabel', 'TargetClass']]
df_meta['ImageId'] = os.path.abspath(
args.input_dir) + '/' + df_meta['ImageId'] + '.png'
df_meta['TrueLabel'] = df_meta['TrueLabel'] - 1
df_meta['TargetClass'] = df_meta['TargetClass'] - 1
df_meta = df_meta.set_index('ImageId')
meta_dict = df_meta.to_dict()['TrueLabel']
meta_target_dict = df_meta.to_dict()['TargetClass']
images_info = image_generators.get_names_and_labels(\
args.input_dir, mode='flat_targeted', meta_dict=meta_dict, meta_target_dict=meta_target_dict)
elif args.input_dir_mode == 'test':
images_info = image_generators.get_names_and_labels(\
args.input_dir, mode='test')
elif args.input_dir_mode == 'test_targeted':
df_meta = pd.read_csv(os.path.join(os.path.abspath(args.input_dir), 'target_class.csv'), \
header=None, names=['ImageId', 'TargetClass'])
df_meta['TargetClass'] = df_meta['TargetClass'] - 1
df_meta = df_meta.set_index('ImageId')
meta_target_dict = df_meta.to_dict()['TargetClass']
images_info = image_generators.get_names_and_labels(\
args.input_dir, mode='test_targeted', meta_target_dict=meta_target_dict)
else:
images_info = image_generators.get_names_and_labels(\
args.input_dir, mode='hierarchy')
image_generators.initialize_hierarchy(args.output_dir)
print('Total images:', len(images_info))
eps = args.max_epsilon
if args.num_samples < 0:
num_samples = len(images_info)
else:
num_samples = args.num_samples
is_targeted = ('targeted' in args.input_dir_mode)
print('Is targeted', is_targeted)
use_avg_pred = True
print('Loading images')
# each image_data element: filename, img, lbl
image_data = image_generators.load_images_into_batches(\
images_info, args.train_batch_size, \
max_samples=num_samples, to_buffer=True)
print('Generating placeholders')
placeholders = {}
for i, data in enumerate(image_data):
imgmat_size = (len(data[0]), 299, 299, 3)
lblmat_size = (len(data[0]), 1000)
data_min = task_utils.get_raw_array(init=np.clip(
task_utils.get_array(data[1], imgmat_size) - eps, 0, None))
data_max = task_utils.get_raw_array(init=np.clip(
task_utils.get_array(data[1], imgmat_size) + eps, None, 255))
data_actual = data[1]
grad_mat = task_utils.get_raw_array(dims=imgmat_size)
lbls_mat = task_utils.get_raw_array(
dims=lblmat_size) if not is_targeted else data[3]
# data, min, max, grad, lbl, pseudo-label matrix
placeholders[i] = (\
data[0], \
data_actual, \
data_min, \
data_max, \
grad_mat, \
data[2], \
lbls_mat)
print('Preparing functions')
source_models = [
'incresv2ensadv', 'resnet50', 'inceptionv3adv', 'inceptionv3'
]
pred_models = ['incresv2ensadv', 'resnet50', 'inceptionv3adv']
step_size = 1.8
noise_size = 0.
grad_aug_scale = 0.
plan = [\
(('RST', 'ir2ea-i3a-i3-r50-t', ('incresv2ensadv', 'inceptionv3adv', 'inceptionv3', 'resnet50')), \
(('incresv2ensadv', 1), ('inceptionv3adv', 1), ('inceptionv3', 1), step_size, noise_size), \
(('incresv2ensadv', 1), ('inceptionv3adv', 1), ('inceptionv3', 1), step_size, noise_size), \
(('incresv2ensadv', 1), ('inceptionv3adv', 1), ('inceptionv3', 1), step_size, noise_size), \
(('incresv2ensadv', 1), step_size, noise_size), \
(('incresv2ensadv', 1), ('inceptionv3adv', 1), ('inceptionv3', 1), step_size, noise_size), \
(('incresv2ensadv', 1), ('inceptionv3adv', 1), step_size, noise_size), \
(('incresv2ensadv', 1), step_size, noise_size), \
(('incresv2ensadv', 1), ('inceptionv3adv', 1), step_size, noise_size), \
(('incresv2ensadv', 1), ('inceptionv3adv', 1), step_size, noise_size), \
(('incresv2ensadv', 1), ('inceptionv3adv', 1), step_size, noise_size), \
(('incresv2ensadv', 1), ('inceptionv3adv', 1), step_size*0.5, noise_size), \
(('incresv2ensadv', 1), step_size*0.5, noise_size), \
(('incresv2ensadv', 1), step_size*0.5, noise_size), \
(('incresv2ensadv', 1), ('inceptionv3', 1), step_size*0.5, noise_size), \
(('incresv2ensadv', 1), ('inceptionv3adv', 1), step_size*0.5, noise_size), \
(('incresv2ensadv', 1), step_size*0.5, noise_size), \
(('incresv2ensadv', 1), ('inceptionv3adv', 1), ('inceptionv3', 1), step_size*0.25, noise_size), \
(('incresv2ensadv', 1), ('inceptionv3adv', 1), ('inceptionv3', 1), step_size*0.25, noise_size), \
(('incresv2ensadv', 1), step_size*0.25, noise_size), \
(('incresv2ensadv', 1), step_size*0.25, noise_size), \
(('incresv2ensadv', 1), step_size*0.25, noise_size), \
(('incresv2ensadv', 1), ('inceptionv3adv', 1), ('inceptionv3', 1), step_size*0.25, noise_size), \
(('incresv2ensadv', 1), ('inceptionv3adv', 1), ('inceptionv3', 1), step_size*0.25, noise_size), \
(('incresv2ensadv', 1), step_size*0.1, noise_size), \
(('incresv2ensadv', 1), ('inceptionv3', 1), step_size*0.1, noise_size), \
(('incresv2ensadv', 1), step_size*0.1, noise_size), \
(('incresv2ensadv', 1), step_size*0.1, noise_size), \
(('incresv2ensadv', 1), ('inceptionv3', 1), step_size*0.1, noise_size), \
(('incresv2ensadv', 1), ('inceptionv3', 1), step_size*0.1, noise_size), \
(('incresv2ensadv', 1), step_size*0.1, noise_size), \
(('incresv2ensadv', 1), ('inceptionv3', 1), step_size*0.1, noise_size), \
(('incresv2ensadv', 1), ('inceptionv3adv', 1), step_size*0.1, noise_size), \
(('incresv2ensadv', 1), step_size*0.1, noise_size), \
(('incresv2ensadv', 1), ('inceptionv3', 1), step_size*0.1, noise_size), \
(('incresv2ensadv', 1), ('inceptionv3', 1), step_size*0.1, noise_size), \
(('incresv2ensadv', 1), ('inceptionv3adv', 1), step_size*0.1, noise_size), \
(('incresv2ensadv', 1), ('inceptionv3adv', 1), step_size*0.1, noise_size), \
(('incresv2ensadv', 1), step_size*0.1, noise_size), \
(('incresv2ensadv', 1), step_size*0.1, noise_size), \
(('incresv2ensadv', 1), ('inceptionv3', 1), step_size*0.1, noise_size), \
(('incresv2ensadv', 1), ('inceptionv3adv', 1), step_size*0.1, noise_size), \
(('incresv2ensadv', 1), ('inceptionv3', 1), step_size*0.1, noise_size), \
(('incresv2ensadv', 1), ('inceptionv3', 1), step_size*0.1, noise_size)), \
]
if not task_utils.verify_plan(source_models, pred_models, plan):
return
task_list = []
phase_info = {}
next_task = {}
for i_phase, phase in enumerate(plan):
last_task_in_batch = {}
if phase[0][0] == 'RST':
task = ('phase {0} restore from {1} model {2}'.format(i_phase, phase[0][1], phase[0][2]), \
(), 7, (phase[0][1],))
task_list.append(task)
phase_info[i_phase] = (len(task_list) - 1, set(), set())
# COPIED
if (not is_targeted) and (i_phase == 0):
for phid, (fname_list, imgs, imgs_min, imgs_max, grads, lbls,
lbls_mat) in placeholders.items():
task = ('batch {0}, fname [{1}], pred'.format(phid, ' '.join([os.path.split(f)[1][:-4] for f in fname_list])), \
(fname_list,imgs,None,grads,imgs_min,imgs_max,lbls_mat),3,(None,))
task_list.append(task)
last_task_in_batch[phid] = len(task_list) - 1
phase_info[i_phase][1].add(len(task_list) - 1)
phase_info[i_phase][2].add(len(task_list) - 1)
else:
if i_phase:
task = ('phase {0} reload {1}'.format(i_phase, phase[0]), (), 5, \
(phase[0][0], use_avg_pred, is_targeted))
task_list.append(task)
# id of loading task, all tasks other than load, first tasks after load
phase_info[i_phase] = (len(task_list) - 1, set(), set())
else:
task = ('phase {0} initial load {1} pred {2}'.format(i_phase, phase[0][0], phase[0][1]), (), 6, \
(phase[0][0], phase[0][1], use_avg_pred, is_targeted))
task_list.append(task)
phase_info[0] = (len(task_list) - 1, set(), set())
if not is_targeted:
for phid, (fname_list, imgs, imgs_min, imgs_max, grads,
lbls, lbls_mat) in placeholders.items():
task = ('batch {0}, fname [{1}], pred'.format(phid, ' '.join([os.path.split(f)[1][:-4] for f in fname_list])), \
(fname_list,imgs,None,grads,imgs_min,imgs_max,lbls_mat),3,(None,))
task_list.append(task)
last_task_in_batch[phid] = len(task_list) - 1
phase_info[i_phase][1].add(len(task_list) - 1)
phase_info[i_phase][2].add(len(task_list) - 1)
for phid, (fname_list, imgs, imgs_min, imgs_max, grads, lbls,
lbls_mat) in placeholders.items():
for i_pl, pl in enumerate(phase[1:]):
if pl[-1] > 1e-5:
task = ('phase {0} batch {1} step {2}, noise {3:.2f}'.format(i_phase, phid, i_pl, pl[-1]), \
(fname_list,imgs,None,grads,imgs_min,imgs_max,lbls_mat),4,\
(pl[-1],))
task_list.append(task)
# COPIED
if phid not in last_task_in_batch:
phase_info[i_phase][2].add(len(task_list) - 1)
else:
next_task[
last_task_in_batch[phid]] = len(task_list) - 1
last_task_in_batch[phid] = len(task_list) - 1
phase_info[i_phase][1].add(len(task_list) - 1)
for j_stp, stp in enumerate(pl[:-2]):
task = ('phase {0} batch {1} step {2}, grad {3} repeat {4}'.format(i_phase, phid, i_pl, stp[0], stp[1]), \
(fname_list,imgs,None,grads,imgs_min,imgs_max,lbls_mat),0,\
(None, stp[0], grad_aug_scale, stp[1]))
task_list.append(task)
# COPIED
if phid not in last_task_in_batch:
phase_info[i_phase][2].add(len(task_list) - 1)
else:
next_task[
last_task_in_batch[phid]] = len(task_list) - 1
last_task_in_batch[phid] = len(task_list) - 1
phase_info[i_phase][1].add(len(task_list) - 1)
task = ('phase {0} batch {1} step {2}, FGSM step, step size {3:.2f}'.format(i_phase, phid, i_pl, pl[-2]), \
(fname_list,imgs,None,grads,imgs_min,imgs_max,lbls_mat),1,\
(pl[-2],))
task_list.append(task)
# COPIED
if phid not in last_task_in_batch:
phase_info[i_phase][2].add(len(task_list) - 1)
else:
next_task[last_task_in_batch[phid]] = len(task_list) - 1
last_task_in_batch[phid] = len(task_list) - 1
phase_info[i_phase][1].add(len(task_list) - 1)
task = ('phase {0} batch {1}, finish'.format(i_phase, phid), \
(fname_list,imgs,None,grads,imgs_min,imgs_max,lbls_mat),2,(len(input_dir_abs), args.output_dir))
task_list.append(task)
# COPIED
if phid not in last_task_in_batch:
phase_info[i_phase][2].add(len(task_list) - 1)
else:
next_task[last_task_in_batch[phid]] = len(task_list) - 1
last_task_in_batch[phid] = len(task_list) - 1
phase_info[i_phase][1].add(len(task_list) - 1)
print('Preparation finished', time.time() - start_time)
print(len(task_list), 'tasks')
task_utils.run_tasks(task_list,
phase_info,
next_task,
verbose=False,
cpu_worker=1)
print('Time:', time.time() - start_time)
print('Total images:', len(images_info))
print('Augmentation:', grad_aug_scale)
print('\n'.join(str(u) for u in plan))
print()
def main():
parser = argparse.ArgumentParser()
parser.add_argument('--input_dir', required=True, type=str, \
help='Input directory with images.')
parser.add_argument('--output_file', required=True, type=str, \
help='Output directory with images.')
parser.add_argument('--input_dir_mode', default='test', type=str, \
help='Either flat or hierarchy, how the input dir is organised')
parser.add_argument('--train_batch_size', default=24, type=int, \
help='How many images process at one time.')
args = parser.parse_args()
input_dir_abs = os.path.abspath(args.input_dir)
if args.input_dir_mode == 'flat':
df_meta = pd.read_csv(args.meta)
df_meta = df_meta[['ImageId', 'TrueLabel']]
df_meta['ImageId'] = os.path.abspath(
args.input_dir) + '/' + df_meta['ImageId'] + '.png'
df_meta['TrueLabel'] = df_meta['TrueLabel'] - 1
df_meta = df_meta.set_index('ImageId')
meta_dict = df_meta.to_dict()['TrueLabel']
images_info = image_generators.get_names_and_labels(\
args.input_dir, mode='flat', meta_dict=meta_dict)
elif args.input_dir_mode == 'flat_targeted':
df_meta = pd.read_csv(args.meta)
df_meta = df_meta[['ImageId', 'TrueLabel', 'TargetClass']]
df_meta['ImageId'] = os.path.abspath(
args.input_dir) + '/' + df_meta['ImageId'] + '.png'
df_meta['TrueLabel'] = df_meta['TrueLabel'] - 1
df_meta['TargetClass'] = df_meta['TargetClass'] - 1
df_meta = df_meta.set_index('ImageId')
meta_dict = df_meta.to_dict()['TrueLabel']
meta_target_dict = df_meta.to_dict()['TargetClass']
images_info = image_generators.get_names_and_labels(\
args.input_dir, mode='flat_targeted', meta_dict=meta_dict, meta_target_dict=meta_target_dict)
elif args.input_dir_mode == 'test':
images_info = image_generators.get_names_and_labels(\
args.input_dir, mode='test')
elif args.input_dir_mode == 'test_targeted':
df_meta = pd.read_csv(os.path.join(obs.path.abspath(args.input_dir), 'target_class.csv'), \
header=None, names=['ImageId', 'TargetClass'])
df_meta = df_meta.set_index('ImageId')
meta_target_dict = df_meta.to_dict()['TargetClass']
images_info = image_generators.get_names_and_labels(\
args.input_dir, mode='test_targeted', meta_target_dict=meta_target_dict)
else:
images_info = image_generators.get_names_and_labels(\
args.input_dir, mode='hierarchy')
image_generators.initialize_hierarchy(args.output_dir)
print('Total images:', len(images_info))
eps = 8.
num_samples = len(images_info)
use_avg_pred = True
step_size = eps
noise_size = eps * 0.2
grad_aug_scale = 0.02
print('Loading images')
# each image_data element: filename, img, lbl
image_data = image_generators.load_images_into_batches(\
images_info, args.train_batch_size, args.input_dir, \
max_samples=num_samples, to_buffer=True)
print('Generating placeholders')
placeholders = {}
for i, data in enumerate(image_data):
imgmat_size = (len(data[0]), 299, 299, 3)
lblmat_size = (len(data[0]), 1000)
data_min = get_raw_array(
init=np.clip(get_array(data[1], imgmat_size) - eps, 0, None))
data_max = get_raw_array(
init=np.clip(get_array(data[1], imgmat_size) + eps, None, 255))
data_actual = data[1]
grad_mat = get_raw_array(dims=imgmat_size)
lbl_mat = get_raw_array(dims=lblmat_size)
# data, min, max, grad, lbl, pseudo-label matrix
placeholders[i] = (\
data[0], \
data_actual, \
data_min, \
data_max, \
grad_mat, \
data[2], \
lbl_mat)
print('Preparing functions')
# each tuple is a sequence of pairs followed by two floats
# the pairs describe which network to use and the rand_repeat for accumulating gradients
# the two floats are the step size and noise size of that FGSM step
source_models = [
'incresv2ensadv', 'inceptionv3adv', 'inceptionv3ens3adv',
'inceptionv3ens4adv'
]
dist_pairs = [('inceptionv3ens3adv', 'inceptionv3ens4adv'), \
('inceptionv3ens3adv', 'incresv2ensadv'), \
('inceptionv3ens4adv', 'incresv2ensadv'), \
('inceptionv3adv', 'incresv2ensadv'), \
('inceptionv3ens3adv', 'inceptionv3adv'), \
('inceptionv3ens4adv', 'inceptionv3adv'), \
]
step_size = 1.
noise_size = 0.4
grad_aug_scale = 0.01
pred_aug_scale = 0.01
plan = [\
(('RST', 'def'), \
#((source_models, dist_pairs), \
((('inceptionv3ens3adv','inceptionv3ens4adv'), 1), step_size, noise_size), \
((('inceptionv3ens3adv','inceptionv3ens4adv'), 1), step_size, noise_size), \
((('inceptionv3ens4adv','incresv2ensadv'), 1), (('inceptionv3ens3adv','incresv2ensadv'), 1), step_size, noise_size), \
((('inceptionv3ens3adv','inceptionv3ens4adv'), 1), step_size, noise_size), \
((('inceptionv3ens4adv','incresv2ensadv'), 1), (('inceptionv3ens3adv','incresv2ensadv'), 1), step_size, noise_size), \
((('inceptionv3ens3adv','inceptionv3ens4adv'), 1), step_size, noise_size), \
((('inceptionv3ens4adv','inceptionv3adv'), 1), (('inceptionv3ens3adv','incresv2ensadv'), 1), step_size, noise_size)), \
]
task_list = []
phase_info = {}
next_task = {}
for i_phase, phase in enumerate(plan):
last_task_in_batch = {}
if phase[0][0] == 'RST':
task = ('phase {0} restore from {1}'.format(i_phase, phase[0][1]), \
(), 7, (phase[0][1],))
task_list.append(task)
phase_info[i_phase] = (len(task_list) - 1, set(), set())
else:
task = ('phase {0} load {1}'.format(i_phase, phase[0]), (), 5+1*(i_phase==0), \
(phase[0][0], phase[0][1]))
task_list.append(task)
# id of loading task, all tasks other than load, first tasks after load
phase_info[i_phase] = (len(task_list) - 1, set(), set())
for phid, (fname_list, imgs, imgs_min, imgs_max, grads, lbls,
lbls_mat) in placeholders.items():
for i_pl, pl in enumerate(phase[1:]):
if pl[-1] > 1e-5:
task = ('phase {0} batch {1} step {2}, noise {3:.2f}'.format(i_phase, phid, i_pl, pl[-1]), \
(fname_list,imgs,None,grads,imgs_min,imgs_max,lbls_mat),4,\
(pl[-1],))
task_list.append(task)
# COPIED
if phid not in last_task_in_batch:
phase_info[i_phase][2].add(len(task_list) - 1)
else:
next_task[
last_task_in_batch[phid]] = len(task_list) - 1
last_task_in_batch[phid] = len(task_list) - 1
phase_info[i_phase][1].add(len(task_list) - 1)
for j_stp, stp in enumerate(pl[:-2]):
task = ('phase {0} batch {1} step {2}, grad {3} repeat {4}'.format(i_phase, phid, i_pl, stp[0], stp[1]), \
(fname_list,imgs,None,grads,imgs_min,imgs_max,lbls_mat),0,\
(None, stp[0], grad_aug_scale, stp[1]))
task_list.append(task)
# COPIED
if phid not in last_task_in_batch:
phase_info[i_phase][2].add(len(task_list) - 1)
else:
next_task[
last_task_in_batch[phid]] = len(task_list) - 1
last_task_in_batch[phid] = len(task_list) - 1
phase_info[i_phase][1].add(len(task_list) - 1)
task = ('phase {0} batch {1} step {2}, FGSM step, step size {3:.2f}'.format(i_phase, phid, i_pl, pl[-2]), \
(fname_list,imgs,None,grads,imgs_min,imgs_max,lbls_mat),1,\
(pl[-2],))
task_list.append(task)
# COPIED
if phid not in last_task_in_batch:
phase_info[i_phase][2].add(len(task_list) - 1)
else:
next_task[last_task_in_batch[phid]] = len(task_list) - 1
last_task_in_batch[phid] = len(task_list) - 1
phase_info[i_phase][1].add(len(task_list) - 1)
task = ('phase {0} batch {1}, final pred'.format(i_phase, phid), \
(fname_list,imgs,None,grads,imgs_min,imgs_max,lbls_mat),3,(pred_aug_scale,))
task_list.append(task)
# COPIED
if phid not in last_task_in_batch:
phase_info[i_phase][2].add(len(task_list) - 1)
else:
next_task[last_task_in_batch[phid]] = len(task_list) - 1
last_task_in_batch[phid] = len(task_list) - 1
phase_info[i_phase][1].add(len(task_list) - 1)
print('Preparation finished', time.time() - start_time)
print(len(task_list), 'tasks')
run_tasks(task_list, phase_info, next_task, verbose=False, cpu_worker=1)
with open(args.output_file, 'w') as f:
for phid, (fnames, _, _, _, _, _, lbls_mat) in placeholders.items():
batch_preds = get_array(lbls_mat,
(len(fnames), 1000)).argmax(axis=1)
for img, pred in zip(fnames, batch_preds):
_ = f.write('{0},{1}\n'.format(
os.path.split(img)[1], pred + 1))
print('Time:', time.time() - start_time)
print('\n'.join(str(u) for u in plan))
print()