def construct_model_and_data(args):
model = ImageModel(args.arch, args.dataset)
loader = ImageData(args.data, args.dataset,
num_samples=args.num_samples).data_loader
target, label = next(iter(loader))
target, label = target.to(device), label.to(device)
x_data, y_data = split_data(target, label, model, args.num_classes)
result = {
'data_model': model,
'x_data': x_data,
'y_data': y_data,
'clip_min': 0.0,
'clip_max': 1.0
}
if args.attack_type == 'targeted':
np.random.seed(0)
length = y_data.shape[0]
idx = [
np.random.choice([j for j in range(length) if j != label])
for label in range(length)
]
result['target_labels'] = y_data[idx]
result['target_images'] = x_data[idx]
return result
def main(args):
# Ensure training, testing, and manip are not all turned off
assert (
args.train or args.test or args.manip
), 'Cannot have train, test, and manip all set to 0, Nothing to do.'
# Load the training, validation, and testing data
try:
train_list, val_list, test_list = load_data(args.data_root_dir,
args.split_num)
except:
# Create the training and test splits if not found
logging.info(
'\nNo existing training, validate, test files...System will generate it.'
)
split_data(args.data_root_dir, num_splits=args.Kfold)
train_list, val_list, test_list = load_data(args.data_root_dir,
args.split_num)
# Get image properties from first image. Assume they are all the same.
logging.info('\nRead image files...%s' %
(join(args.data_root_dir, 'imgs', train_list[0][0])))
# Get image shape from the first image.
image = sitk.GetArrayFromImage(
sitk.ReadImage(join(args.data_root_dir, 'imgs', train_list[0][0])))
img_shape = image.shape # # (x, y, channels)
if args.dataset == 'luna16':
net_input_shape = (img_shape[1], img_shape[2], args.slices)
else:
args.slices = 1
if GRAYSCALE:
net_input_shape = (RESOLUTION, RESOLUTION, 1) # only one channel
else:
net_input_shape = (RESOLUTION, RESOLUTION, 3
) # Only access RGB 3 channels.
# Create the model for training/testing/manipulation
# enable_decoder = False only for SegCaps R3 to disable recognition image output on evaluation model
# to speed up performance.
model_list = create_model(args=args,
input_shape=net_input_shape,
enable_decoder=True)
print_summary(model=model_list[0], positions=[.38, .65, .75, 1.])
args.output_name = 'split-' + str(args.split_num) + '_batch-' + str(args.batch_size) + \
'_shuff-' + str(args.shuffle_data) + '_aug-' + str(args.aug_data) + \
'_loss-' + str(args.loss) + '_slic-' + str(args.slices) + \
'_sub-' + str(args.subsamp) + '_strid-' + str(args.stride) + \
'_lr-' + str(args.initial_lr) + '_recon-' + str(args.recon_wei)
# args.output_name = 'sh-' + str(args.shuffle_data) + '_a-' + str(args.aug_data)
args.time = time
if platform.system() == 'Windows':
args.use_multiprocessing = False
else:
args.use_multiprocessing = True
args.check_dir = join(args.data_root_dir, 'saved_models', args.net)
try:
makedirs(args.check_dir)
except:
pass
args.log_dir = join(args.data_root_dir, 'logs', args.net)
try:
makedirs(args.log_dir)
except:
pass
args.tf_log_dir = join(args.log_dir, 'tf_logs')
try:
makedirs(args.tf_log_dir)
except:
pass
args.output_dir = join(args.data_root_dir, 'plots', args.net)
try:
makedirs(args.output_dir)
except:
pass
if args.train:
from train import train
# Run training
train(args, train_list, val_list, model_list[0], net_input_shape)
if args.test:
from test import test
# Run testing
test(args, test_list, model_list, net_input_shape)
if args.manip:
from manip import manip
# Run manipulation of segcaps
manip(args, test_list, model_list, net_input_shape)
'''
The file is used to split the training dataset.
The test data of the original datset should not be touched.
'''
import utils.load_data as datama
import argparse
if __name__ == '__main__':
ap = argparse.ArgumentParser()
ap.add_argument("-t", "--toy", type=bool, default=False)
args = vars(ap.parse_args())
if args['toy']:
print("Toy Data")
datama.spli_toy_data()
else:
print("Not Toy Data")
datama.split_data(10000)
fig.savefig('./{}.png'.format(save_path), bbox_inches='tight')
print('saved')
def draw_boxplot(df, save_path=None, is_save=False):
fig = plt.figure(figsize=(6, 6))
ax = fig.add_subplot(111)
ax.set_xlabel('Initial AUC', fontsize=18)
ax.set_ylabel('lambda', fontsize=18)
sns.boxplot(data=df, orient='h', palette='Pastel1', width=0.5, sym='')
if save_path is not None and is_save == True:
save_graph(fig, save_path)
# %%
train_X, test_X, train_y, test_y = split_data(df_insect, False)
print('{}insect{}'.format('-' * 10, '-' * 10))
df_wj = pd.DataFrame([])
for lambd in lambds:
oracle_insect = Oracle(eta=eta, lambd=lambd)
min_w = oracle_insect.estimate_min_w(pd.concat([train_X, test_X]),
pd.concat([train_y, test_y]))
# print('[min_w] {}: {}'.format(min_w, predict(test_X, test_y, min_w)))
W_init = oracle_insect.make_W_init(J=J)
tmp = pd.Series(predict_wj(test_X, test_y, W_init))
df_wj = pd.concat([df_wj, tmp], axis=1)
# print("lambd = {} : {}".format(lambd, predict_by_W(test_X, test_y, W_init)))
df_wj.columns = ['0', '1', '2', '3', '4', '5', '10']
draw_boxplot(df_wj)
# %%