def extract_feature(train_img, test_img, path=None):
"""
This help to compute feature for knn from pretrained network
:param FLAGS:
:param ckpt_path:
:return:
"""
# check if a certain variable has been saved in the model
if config.extract_feature == 'feature':
dir_path = os.path.join(config.save_model, config.dataset)
dir_path = os.path.join(dir_path,
'knn_num_neighbor_' + str(config.nb_teachers))
filename = str(config.nb_teachers) + '_stdnt_resnet.checkpoint.pth.tar'
filename = os.path.join(dir_path, filename)
train_feature = network.pred(train_img, filename, return_feature=True)
test_feature = network.pred(test_img, filename, return_feature=True)
print('shape of extract feature', train_feature.shape)
return train_feature, test_feature
#return utils.pca(test_feature, train_feature)
if config.extract_feature == 'hog':
# usually the file to save all hog is too large. we decompose it into 10 pieces.
train_data = None
each_length = int((9 + len(train_img)) / 10)
for idx in range(10):
#to save pkl into several small pieces
train_hog_path = os.path.join(
config.hog_path, config.dataset + str(idx) + '_train_hog.pkl')
if os.path.exists(train_hog_path) == False:
p1 = idx * each_length
p2 = min((idx + 1) * each_length, len(train_img))
print('save_hog_pkl for interval{} : {}'.format(p1, p2))
utils.save_hog(train_img[p1:p2], train_hog_path)
with open(train_hog_path, 'rb') as f:
if train_data is not None:
train_data = np.vstack((train_data, pickle.load(f)))
else:
train_data = pickle.load(f)
print('load hog feature shape', train_data.shape)
test_hog_path = os.path.join(config.hog_path,
config.dataset + '_test_hog.pkl')
if os.path.exists(test_hog_path) == False:
utils.save_hog(test_img, test_hog_path)
with open(test_hog_path, 'rb') as f:
test_data = pickle.load(f)
return train_data, test_data
if config.extract_feature == 'pca':
return utils.pca(test_img, train_img)
def extract_feature(train_img, test_img, path=None):
"""
This help to compute feature for knn from pretrained network
:param FLAGS:
:param ckpt_path:
:return:
"""
# check if a certain variable has been saved in the model
if config.extract_feature == 'feature':
# Update the feature extractor using the student model(filename) in the last iteration.
# Replace the filename with the saved student model, the following in an example of the checkpoint
filename = 'save_model/svhn/knn_num_neighbor_800/800_stdnt_.checkpoint.pth.tar'
train_feature = network.pred(train_img, filename, return_feature=True)
test_feature = network.pred(test_img, filename, return_feature=True)
return train_feature, test_feature
train_img = [np.asarray(data) for data in train_img]
test_img = [np.asarray(data) for data in test_img]
if config.extract_feature == 'hog':
# usually the file to save all hog is too large. we decompose it into 10 pieces.
train_data = None
each_length = int((9 + len(train_img)) / 10)
for idx in range(10):
#Save pkl into several small pieces, incase the size of private dataset is too large
train_hog_path = os.path.join(
config.hog_path, config.dataset + str(idx) + '_train_hog.pkl')
if os.path.exists(train_hog_path) == False:
p1 = idx * each_length
p2 = min((idx + 1) * each_length, len(train_img))
print('save_hog_pkl for interval{} : {}'.format(p1, p2))
utils.save_hog(train_img[p1:p2], train_hog_path)
with open(train_hog_path, 'rb') as f:
if train_data is not None:
train_data = np.vstack((train_data, pickle.load(f)))
else:
train_data = pickle.load(f)
print('load hog feature shape', train_data.shape)
test_hog_path = os.path.join(config.hog_path,
config.dataset + '_test_hog.pkl')
if os.path.exists(test_hog_path) == False:
utils.save_hog(test_img, test_hog_path)
with open(test_hog_path, 'rb') as f:
test_data = pickle.load(f)
return train_data, test_data
if config.extract_feature == 'pca':
return utils.pca(test_img, train_img)
def train_student(nb_teachers):
"""
This function trains a student using predictions made by an ensemble of
teachers. The student and teacher models are trained using the same
neural network architecture.
:param dataset: string corresponding to mnist, cifar10, or svhn
:param nb_teachers: number of teachers (in the ensemble) to learn from
:return: True if student training went well
"""
# Call helper function to prepare student data using teacher predictions
stdnt_dataset = prepare_student_data(nb_teachers, save=True)
# Unpack the student dataset
stdnt_data, stdnt_labels, stdnt_test_data, stdnt_test_labels = stdnt_dataset
if config.resnet:
dir_path = os.path.join(config.save_model, config.dataset)
dir_path = os.path.join(dir_path,
'pate_num_teacher_' + str(config.nb_teachers))
#dir_path = os.path.join(config.save_model,'pate_'+str(config.nb_teachers))
utils.mkdir_if_missing(dir_path)
filename = os.path.join(dir_path, '_stndent_resnet.checkpoint.pth.tar')
print('stdnt_label used for train', stdnt_labels.shape)
network.train_each_teacher(config.student_epoch, stdnt_data, stdnt_labels,
stdnt_test_data, stdnt_test_labels, filename)
final_preds = network.pred(stdnt_test_data, filename)
precision = hamming_accuracy(final_preds, stdnt_test_labels, torch=False)
print('Precision of student after training: ' + str(precision))
return True
def ensemble_preds(nb_teachers, stdnt_data):
result_shape = (nb_teachers, len(stdnt_data))
# Create array that will hold result
result = np.zeros(result_shape, dtype=np.float32)
dir_path = os.path.join(config.save_model,
'pate_' + str(config.nb_teachers))
for teacher_id in range(config.nb_teachers):
filename = os.path.join(
dir_path,
str(config.nb_teachers) + '_teachers_' + str(teacher_id) +
config.arch + '.checkpoint.pth.tar')
result[teacher_id] = network.pred(stdnt_data, filename)
print("Computed Teacher " + str(teacher_id) + " softmax predictions")
result = np.asarray(result, dtype=np.int32)
return result.transpose()
def ensemble_preds(nb_teachers, stdnt_data):
"""
Given a dataset, a number of teachers, and some input data, this helper
function queries each teacher for predictions on the data and returns
all predictions in a single array. (That can then be aggregated into
one single prediction per input using aggregation.py (cf. function
prepare_student_data() below)
:param dataset: string corresponding to mnist, cifar10, or svhn
:param nb_teachers: number of teachers (in the ensemble) to learn from
:param stdnt_data: unlabeled student training data
:return: 3d array (teacher id, sample id, probability per class)
"""
result_shape = (nb_teachers, len(stdnt_data), config.nb_labels)
# Create array that will hold result
result = np.zeros(result_shape, dtype=np.float32)
# Get predictions from each teacher
for teacher_id in range(nb_teachers):
# Compute path of checkpoint file for teacher model with ID teacher_id
if config.dataset == 'celeba':
dir_path = os.path.join(
config.save_model,
'pate_num_teacher_' + str(config.nb_teachers))
elif config.dataset == 'market':
dir_path = os.path.join(
config.save_model,
'pate_' + config.dataset + str(config.nb_teachers))
utils.mkdir_if_missing(dir_path)
filename = os.path.join(
dir_path,
str(config.nb_teachers) + '_teachers_' + str(teacher_id) +
config.arch + '.checkpoint.pth.tar')
result[teacher_id] = network.pred(stdnt_data, filename)
# This can take a while when there are a lot of teachers so output status
print("Computed Teacher " + str(teacher_id) + " softmax predictions")
return result
