def test():
parser = argparse.ArgumentParser()
parser.add_argument('--config',
'-c',
default='configs/celebA_YSBBB_Classifier.yaml')
args = parser.parse_args()
# ============= Load config =============
config_path = args.config
config = yaml.load(open(config_path))
print(config)
# ============= Experiment Folder=============
output_dir = os.path.join(config['log_dir'], config['name'])
classifier_output_path = os.path.join(output_dir, 'classifier_output')
try:
os.makedirs(classifier_output_path)
except:
pass
past_checkpoint = output_dir
# ============= Experiment Parameters =============
BATCH_SIZE = config['batch_size']
channels = config['num_channel']
input_size = config['input_size']
N_CLASSES = config['num_class']
# ============= Data =============
try:
categories, file_names_dict = read_data_file(
config['image_label_dict'])
except:
print("Problem in reading input data file : ",
config['image_label_dict'])
sys.exit()
data_train = np.load(config['train'])
data_test = np.load(config['test'])
print("The classification categories are: ")
print(categories)
print('The size of the training set: ', data_train.shape[0])
print('The size of the testing set: ', data_test.shape[0])
# ============= placeholder =============
with tf.name_scope('input'):
x_ = tf.placeholder(tf.float32,
[None, input_size, input_size, channels],
name='x-input')
y_ = tf.placeholder(tf.int64, [None, N_CLASSES], name='y-input')
isTrain = tf.placeholder(tf.bool)
# ============= Model =============
if N_CLASSES == 1:
y = tf.reshape(y_, [-1])
y = tf.one_hot(y, 2, on_value=1.0, off_value=0.0, axis=-1)
logit, prediction = pretrained_classifier(x_,
n_label=2,
reuse=False,
name='classifier',
isTrain=isTrain)
else:
logit, prediction = pretrained_classifier(x_,
n_label=N_CLASSES,
reuse=False,
name='classifier',
isTrain=isTrain)
y = y_
classif_loss = tf.losses.sigmoid_cross_entropy(multi_class_labels=y,
logits=logit)
loss = tf.losses.get_total_loss()
# ============= Variables =============
# Note that this list of variables only include the weights and biases in the model.
lst_vars = []
for v in tf.global_variables():
lst_vars.append(v)
# ============= Session =============
sess = tf.InteractiveSession()
saver = tf.train.Saver(var_list=lst_vars)
tf.global_variables_initializer().run()
# ============= Load Checkpoint =============
if past_checkpoint is not None:
ckpt = tf.train.get_checkpoint_state(past_checkpoint + '/')
if ckpt and ckpt.model_checkpoint_path:
print("HERE...................lod checkpoint.........")
print(str(ckpt.model_checkpoint_path))
saver.restore(sess,
tf.train.latest_checkpoint(past_checkpoint + '/'))
else:
sys.exit()
else:
sys.exit()
# ============= Testing Save the Output =============
names = np.empty([0])
prediction_y = np.empty([0])
true_y = np.empty([0])
for epoch in range(1):
num_batch = int(data_train.shape[0] / BATCH_SIZE)
for i in range(0, num_batch):
start = i * BATCH_SIZE
ns = data_train[start:start + BATCH_SIZE]
xs, ys = load_images_and_labels(ns,
config['image_dir'],
N_CLASSES,
file_names_dict,
input_size,
channels,
do_center_crop=True)
[_pred] = sess.run([prediction],
feed_dict={
x_: xs,
isTrain: False,
y_: ys
})
if i == 0:
names = np.asarray(ns)
prediction_y = np.asarray(_pred)
true_y = np.asarray(ys)
else:
names = np.append(names, np.asarray(ns), axis=0)
prediction_y = np.append(prediction_y,
np.asarray(_pred),
axis=0)
true_y = np.append(true_y, np.asarray(ys), axis=0)
np.save(classifier_output_path + '/name_train1.npy', names)
np.save(classifier_output_path + '/prediction_y_train1.npy',
prediction_y)
np.save(classifier_output_path + '/true_y_train1.npy', true_y)
names = np.empty([0])
prediction_y = np.empty([0])
true_y = np.empty([0])
num_batch = int(data_test.shape[0] / BATCH_SIZE)
for i in range(0, num_batch):
start = i * BATCH_SIZE
ns = data_test[start:start + BATCH_SIZE]
xs, ys = load_images_and_labels(ns,
config['image_dir'],
N_CLASSES,
file_names_dict,
input_size,
channels,
do_center_crop=True)
[_pred] = sess.run([prediction],
feed_dict={
x_: xs,
isTrain: False,
y_: ys
})
if i == 0:
names = np.asarray(ns)
prediction_y = np.asarray(_pred)
true_y = np.asarray(ys)
else:
names = np.append(names, np.asarray(ns), axis=0)
prediction_y = np.append(prediction_y,
np.asarray(_pred),
axis=0)
true_y = np.append(true_y, np.asarray(ys), axis=0)
np.save(classifier_output_path + '/name_test1.npy', names)
np.save(classifier_output_path + '/prediction_y_test1.npy',
prediction_y)
np.save(classifier_output_path + '/true_y_test1.npy', true_y)
n_neighbors=1,
return_distance=True)
class2 = classes[ind2[0][0]]
[class1, round(dist1[0][0]), class2, round(dist2[0][0], 3)]
return [class1, round(dist1[0][0]), class2, round(dist2[0][0], 3)]
# =============================================================================
# LOAD IMAGES
# =============================================================================
path = 'C:/Users/Pablo/Google Drive/TFM/'
# Load images normal size
class_names = os.listdir(path + 'Images/')
class_numbers = list(map(lambda x: int(x[0]), class_names))
images, labels = utils.load_images_and_labels(path + 'Images/')
num_samples, height, width = images.shape
images_mean = []
images_closest = []
for i in np.arange(0, 61 * 6 + 1, 61):
images_mean.append(np.mean(images[i:i + 61], axis=0))
index = i + np.argmin(
np.sum(np.abs(images[i:i + 61] - images_mean[-1]), axis=(1, 2)))
images_closest.append(images[index])
if save_comp:
ims_closest = []
for im in images_closest:
ims_closest.append(cv2.cvtColor(im, cv2.COLOR_GRAY2BGR))
ims_closest.append(cv2.cvtColor(img_blank, cv2.COLOR_GRAY2BGR))
# Load images reduced
def extract_features_ORB(images, orb):
descriptors = []
for img in images:
descriptors.append(orb.detectAndCompute(img, None)[1])
return descriptors
# =============================================================================
# LOAD IMAGES
# =============================================================================
path = 'C:/Users/Pablo/Google Drive/TFM/'
# Load images
class_names = os.listdir(path + 'Images/')
class_numbers = list(map(lambda x: int(x[0]), class_names))
images, labels = utils.load_images_and_labels(path + 'Images/')
num_samples, height, width = images.shape
MAX_MATCHES_array = np.arange(150, 601, 50)
accuracies = np.zeros((len(video_names), len(MAX_MATCHES_array)))
for index, MAX_MATCHES in enumerate(MAX_MATCHES_array):
# =============================================================================
# CREATE / LOAD ORB FEATURES
# =============================================================================
path = 'C:/Users/Pablo/Google Drive/TFM/'
orb = cv2.ORB_create(MAX_MATCHES)
matcher = cv2.DescriptorMatcher_create(
cv2.DESCRIPTOR_MATCHER_BRUTEFORCE_HAMMING)
if not os.path.isfile(path + 'Models/orb_' + str(MAX_MATCHES) + '.pkl'):
def train():
parser = argparse.ArgumentParser()
parser.add_argument('--config',
'-c',
default='configs/celebA_YSBBB_Classifier.yaml')
args = parser.parse_args()
# ============= Load config =============
config_path = args.config
config = yaml.load(open(config_path))
print(config)
# ============= Experiment Folder=============
output_dir = os.path.join(config['log_dir'], config['name'])
try:
os.makedirs(output_dir)
except:
pass
try:
os.makedirs(os.path.join(output_dir, 'logs'))
except:
pass
# ============= Experiment Parameters =============
BATCH_SIZE = config['batch_size']
EPOCHS = config['epochs']
channels = config['num_channel']
input_size = config['input_size']
N_CLASSES = config['num_class']
ckpt_dir_continue = config['ckpt_dir_continue']
if ckpt_dir_continue == '':
continue_train = False
else:
continue_train = True
# ============= Data =============
try:
categories, file_names_dict = read_data_file(
config['image_label_dict'])
except:
print("Problem in reading input data file : ",
config['image_label_dict'])
sys.exit()
data_train = np.load(config['train'])
data_test = np.load(config['test'])
print("The classification categories are: ")
print(categories)
print('The size of the training set: ', data_train.shape[0])
print('The size of the testing set: ', data_test.shape[0])
fp = open(os.path.join(output_dir, 'setting.txt'), 'w')
fp.write('config_file:' + str(config_path) + '\n')
fp.close()
# ============= placeholder =============
with tf.name_scope('input'):
x_ = tf.placeholder(tf.float32,
[None, input_size, input_size, channels],
name='x-input')
y_ = tf.placeholder(tf.int64, [None, N_CLASSES], name='y-input')
isTrain = tf.placeholder(tf.bool)
# ============= Model =============
if N_CLASSES == 1:
y = tf.reshape(y_, [-1])
y = tf.one_hot(y, 2, on_value=1.0, off_value=0.0, axis=-1)
logit, prediction = pretrained_classifier(x_,
n_label=2,
reuse=False,
name='classifier',
isTrain=isTrain)
else:
logit, prediction = pretrained_classifier(x_,
n_label=N_CLASSES,
reuse=False,
name='classifier',
isTrain=isTrain)
y = y_
classif_loss = tf.losses.sigmoid_cross_entropy(multi_class_labels=y,
logits=logit)
loss = tf.losses.get_total_loss()
# ============= Optimization functions =============
train_step = tf.train.AdamOptimizer(0.0001).minimize(loss)
# ============= summary =============
cls_loss = tf.summary.scalar('classif_loss', classif_loss)
total_loss = tf.summary.scalar('total_loss', loss)
sum_train = tf.summary.merge([cls_loss, total_loss])
# ============= Variables =============
# Note that this list of variables only include the weights and biases in the model.
lst_vars = []
for v in tf.global_variables():
lst_vars.append(v)
# ============= Session =============
sess = tf.InteractiveSession()
saver = tf.train.Saver(var_list=lst_vars)
tf.global_variables_initializer().run()
writer = tf.summary.FileWriter(output_dir + '/train', sess.graph)
writer_test = tf.summary.FileWriter(output_dir + '/test', sess.graph)
# ============= Checkpoints =============
if continue_train:
print("Before training, Load checkpoint ")
print("Reading checkpoint...")
ckpt = tf.train.get_checkpoint_state(ckpt_dir_continue)
if ckpt and ckpt.model_checkpoint_path:
ckpt_name = os.path.basename(ckpt.model_checkpoint_path)
saver.restore(sess, os.path.join(ckpt_dir_continue, ckpt_name))
print(ckpt_name)
print("Successful checkpoint upload")
else:
print("Failed checkpoint load")
sys.exit()
# ============= Training =============
train_loss = []
test_loss = []
itr_train = 0
itr_test = 0
for epoch in range(EPOCHS):
total_loss = 0.0
perm = np.arange(data_train.shape[0])
np.random.shuffle(perm)
data_train = data_train[perm]
num_batch = int(data_train.shape[0] / BATCH_SIZE)
for i in range(0, num_batch):
start = i * BATCH_SIZE
ns = data_train[start:start + BATCH_SIZE]
xs, ys = load_images_and_labels(ns,
config['image_dir'],
N_CLASSES,
file_names_dict,
input_size,
channels,
do_center_crop=True)
[_, _loss, summary_str] = sess.run([train_step, loss, sum_train],
feed_dict={
x_: xs,
isTrain: True,
y_: ys
})
writer.add_summary(summary_str, itr_train)
itr_train += 1
total_loss += _loss
total_loss /= i
print("Epoch: " + str(epoch) + " loss: " + str(total_loss) + '\n')
train_loss.append(total_loss)
total_loss = 0.0
perm = np.arange(data_test.shape[0])
np.random.shuffle(perm)
data_test = data_test[perm]
num_batch = int(data_test.shape[0] / BATCH_SIZE)
for i in range(0, num_batch):
start = i * BATCH_SIZE
ns = data_test[start:start + BATCH_SIZE]
xs, ys = load_images_and_labels(ns,
config['image_dir'],
N_CLASSES,
file_names_dict,
input_size,
channels,
do_center_crop=True)
[_loss, summary_str] = sess.run([loss, sum_train],
feed_dict={
x_: xs,
isTrain: False,
y_: ys
})
writer_test.add_summary(summary_str, itr_test)
itr_test += 1
total_loss += _loss
total_loss /= i
print("Epoch: " + str(epoch) + " Test loss: " + str(total_loss) + '\n')
test_loss.append(total_loss)
if epoch % 2 == 0:
checkpoint_name = os.path.join(output_dir,
'cp1_epoch' + str(epoch) + '.ckpt')
save_path = saver.save(sess, checkpoint_name)
np.save(os.path.join(output_dir, 'logs', 'train_loss.npy'),
np.asarray(train_loss))
np.save(os.path.join(output_dir, 'logs', 'test_loss.npy'),
np.asarray(test_loss))
checkpoint_name = os.path.join(output_dir,
'cp1_epoch' + str(epoch) + '.ckpt')
save_path = saver.save(sess, checkpoint_name)
np.save(os.path.join(output_dir, 'logs', 'train_loss.npy'),
np.asarray(train_loss))
np.save(os.path.join(output_dir, 'logs', 'test_loss.npy'),
np.asarray(test_loss))
for img_name in os.listdir(path + pathnew):
im = plt.imread(path + pathnew + '/' + img_name)
imNew = generic_filter(im, spatial_kernel_filter, size=(neighborhood,neighborhood))
imNew = np.stack((imNew,)*3, axis=2)
plt.imsave(path + '../Modified Images ' + technique + '/' + pathnew + '/' + img_name, imNew)
# =============================================================================
# LOAD IMAGES AND CREATE / LOAD MODEL
# =============================================================================
path = 'C:/Users/Pablo/Google Drive/TFM/'
# Load images
class_names = os.listdir(path + 'Modified Images ' + technique + '/')
class_numbers = list(map(lambda x: int(x[0]), class_names))
images, labels = utils.load_images_and_labels(path + 'Modified Images ' + technique + '/')
num_samples, height, width = images.shape
for model_name in model_names:
if model_name == 'eigen':
num_components = 40
elif model_name == 'fisher':
num_components = 6
else:
print('ERROR: Wrong model_name')
# Generate and save model or Load model if exists
if not os.path.isfile(path + 'Models/' + model_name + '_' + technique + '_' + str(num_components) + '_' + str(threshold) + '.yml'):
if model_name == 'eigen':
model = cv2.face.createEigenFaceRecognizer(num_components)