def main():
### Backbone network (Arcface)
args = get_args()
images = tf.placeholder(name='img_inputs', shape=[112, 112, 3], dtype=tf.float32)
train_phase_dropout = tf.placeholder(dtype=tf.bool, shape=None, name='train_phase')
train_phase_bn = tf.placeholder(dtype=tf.bool, shape=None, name='train_phase_last')
config = yaml.load(open('../configs/config_ms1m_100.yaml'))
images = tf.roll(images, shift=1, axis=2)
image_ = images / 127.5 - 1
# image_ = tf.roll(image_, shift=2, axis=3)
inputs = tf.reshape(image_, [-1, *args.image_size])
embedding_tensor, _ = get_embd(inputs, train_phase_dropout, train_phase_bn, config)
t_vars = tf.global_variables()
restore_vars = [var for var in t_vars if 'embd_extractor' in var.name]
### Graph & load setup
tf_config = tf.ConfigProto(allow_soft_placement=True)
tf_config.gpu_options.allow_growth = True
with tf.Session(config=tf_config) as sess:
tf.global_variables_initializer().run()
saver = tf.train.Saver(var_list=restore_vars)
saver.restore(sess, '../model/best-m-1006000')
generate_tfrecord(sess, images, train_phase_dropout, train_phase_bn, embedding_tensor)
def load_ytf_pairs(basedir, pair_txt, n):
with open(pair_txt, 'r') as csvfile:
trainrows = list(csv.reader(csvfile, delimiter=',', quoting=csv.QUOTE_NONE, skipinitialspace=True))[1:]
train_images = load_images(basedir, trainrows, n)
train_labels = np.array(list(map(lambda r: loadLabelsFromRow(r, n), trainrows)))
### Graph
args = get_args()
images = tf.placeholder(name='img_inputs', shape=[2, 112, 112, 3], dtype=tf.float32)
train_phase_dropout = tf.placeholder(dtype=tf.bool, shape=None, name='train_phase')
train_phase_bn = tf.placeholder(dtype=tf.bool, shape=None, name='train_phase_last')
config = yaml.load(open('./configs/config_ms1m_100.yaml'))
images = tf.roll(images, shift=1, axis=3)
image_ = images / 127.5 - 1
# image_ = tf.roll(image_, shift=2, axis=3)
inputs = tf.reshape(image_, [-1, *args.image_size])
embedding_tensor, _ = get_embd(inputs, train_phase_dropout, train_phase_bn, config)
t_vars = tf.global_variables()
restore_vars = [var for var in t_vars if 'embd_extractor' in var.name]
### Graph & load setup
tf_config = tf.ConfigProto(allow_soft_placement=True)
tf_config.gpu_options.allow_growth = True
with tf.Session(config=tf_config) as sess:
tf.global_variables_initializer().run()
saver = tf.train.Saver(var_list=restore_vars)
saver.restore(sess, './model/best-m-1006000')
train_embeddings = np.zeros((5000, 2, 512))
for i, imgs in enumerate(train_images):
embd_dict = {
images: imgs,
train_phase_dropout: False,
train_phase_bn: False
}
embeddings = sess.run(embedding_tensor, feed_dict=embd_dict)
train_embeddings[i] = embeddings
return train_embeddings, np.ravel(train_labels)
def load_fs_pairs(num):
def get_imgfile(name, file_name):
img_file = name + '/' + file_name
return imread(img_file)
rootpath = '/home/hy/Dataset/face/FaceScrub/'
readname = 'facescrub_mtcnn_112/'
lst_name = os.listdir(rootpath + readname)
match_list = []
nonmatch_list = []
for i in range(num):
match_name, nonmatch_name = random.sample(lst_name, 2)
read_matchfolder = rootpath + readname + match_name
lst_idx = os.listdir(read_matchfolder)
m1, m2, nm1 = random.sample(lst_idx, 3)
match_list.append([
get_imgfile(read_matchfolder, m1),
get_imgfile(read_matchfolder, m2)
])
read_nonmatchfolder = rootpath + readname + nonmatch_name
lst_idx = os.listdir(read_nonmatchfolder)
nm2 = random.sample(lst_idx, 1)[0]
nonmatch_list.append([
get_imgfile(read_matchfolder, nm1),
get_imgfile(read_nonmatchfolder, nm2)
])
total_list = np.concatenate(
(np.array(match_list), np.array(nonmatch_list)), axis=0)
total_label = np.concatenate((np.ones(num), np.zeros(num)))
# Graph
### Graph
args = get_args()
images = tf.placeholder(name='img_inputs',
shape=[2, 112, 112, 3],
dtype=tf.float32)
train_phase_dropout = tf.placeholder(dtype=tf.bool,
shape=None,
name='train_phase')
train_phase_bn = tf.placeholder(dtype=tf.bool,
shape=None,
name='train_phase_last')
config = yaml.load(open('./configs/config_ms1m_100.yaml'))
image_ = images / 127.5 - 1
# image_ = tf.roll(image_, shift=2, axis=3)
inputs = tf.reshape(image_, [-1, *args.image_size])
embedding_tensor, _ = get_embd(inputs, train_phase_dropout, train_phase_bn,
config)
t_vars = tf.global_variables()
restore_vars = [var for var in t_vars if 'embd_extractor' in var.name]
### Graph & load setup
tf_config = tf.ConfigProto(allow_soft_placement=True)
tf_config.gpu_options.allow_growth = True
with tf.Session(config=tf_config) as sess:
tf.global_variables_initializer().run()
saver = tf.train.Saver(var_list=restore_vars)
saver.restore(sess, './model/best-m-1006000')
train_embeddings = np.zeros((num * 2, 2, 512))
for i, imgs in enumerate(total_list):
embd_dict = {
images: imgs,
train_phase_dropout: False,
train_phase_bn: False
}
embeddings = sess.run(embedding_tensor, feed_dict=embd_dict)
train_embeddings[i] = embeddings
return train_embeddings, total_label
### Graph
args = get_args()
images = tf.placeholder(name='img_inputs',
shape=[2, 112, 112, 3],
dtype=tf.float32)
train_phase_dropout = tf.placeholder(dtype=tf.bool,
shape=None,
name='train_phase')
train_phase_bn = tf.placeholder(dtype=tf.bool,
shape=None,
name='train_phase_last')
config = yaml.load(open('./configs/config_ms1m_100.yaml'))
images = tf.roll(images, shift=1, axis=3)
image_ = images / 127.5 - 1
# image_ = tf.roll(image_, shift=2, axis=3)
inputs = tf.reshape(image_, [-1, *args.image_size])
embedding_tensor, _ = get_embd(inputs, train_phase_dropout, train_phase_bn,
config)
t_vars = tf.global_variables()
restore_vars = [var for var in t_vars if 'embd_extractor' in var.name]
### Graph & load setup
tf_config = tf.ConfigProto(allow_soft_placement=True)
tf_config.gpu_options.allow_growth = True
with tf.Session(config=tf_config) as sess:
tf.global_variables_initializer().run()
saver = tf.train.Saver(var_list=restore_vars)
saver.restore(sess, './model/best-m-1006000')
train_embeddings = np.zeros((5000, 2, 512))
for i, imgs in enumerate(train_images):
embd_dict = {
images: imgs,
train_phase_dropout: False,
train_phase_bn: False
}
embeddings = sess.run(embedding_tensor, feed_dict=embd_dict)
train_embeddings[i] = embeddings
return train_embeddings, np.ravel(train_labels)
import os
from sklearn.metrics import roc_curve
import matplotlib.pyplot as plt
from src.model import get_embd, get_args
from src.youtubeface import load_ytf_data
from src.wrapper_basicImg import wrapper_basicImg
if __name__ == '__main__':
os.environ["CUDA_VISIBLE_DEVICES"] = '0'
m = 32
q = 32
n = 5
batch_size = 256
class_num = 1595
args = get_args()
### Fold
fold = tf.placeholder(tf.string)
### Get image and label from tfrecord
image, label, iterator = {}, {}, {}
image['train'], label['train'], iterator['train'] = load_ytf_data(
batch_size, 'train')
image['gallery'], label['gallery'], iterator['gallery'] = load_ytf_data(
batch_size, 'train', eval=True)
image['test'], label['test'], iterator['test'] = load_ytf_data(
batch_size, 'test')
### Get eer evaluation dataset. Wrapper
wrapper = wrapper_basicImg('LFW')
def load_lfw_pairs():
basename = '/home/hy/Dataset/face/LFW/lfw-deepfunneled'
pair_txt = '/home/hy/Dataset/face/LFW/pairs.txt'
tgz_filename = "{}.tgz".format(basename)
tar_filename = "{}.tar".format(basename)
tar_subdir = "lfw-deepfunneled"
# it will be faster to decompress this tar file all at once
print("--> Converting {} to tar".format(tgz_filename))
with gzip.open(tgz_filename, 'rb') as f_in, open(tar_filename,
'wb') as f_out:
shutil.copyfileobj(f_in, f_out)
tar = tarfile.open(tar_filename)
with open(pair_txt, 'r') as csvfile:
trainrows = list(csv.reader(csvfile, delimiter='\t'))[1:]
train_images = load_images(tar, tar_subdir, trainrows)
train_labels = np.array(
list(map(lambda r: loadLabelsFromRow(r), trainrows)))
train_images = np.asarray([crop_and_downsample(x) for x in train_images])
# train_images = np.reshape(train_images, (10, 600, 2, 112, 112, 3))
# train_labels = np.reshape(train_labels, (10, 600))
train_images = np.reshape(train_images, (6000, 2, 112, 112, 3))
train_labels = np.reshape(train_labels, (6000))
### Graph
args = get_args()
images = tf.placeholder(name='img_inputs',
shape=[2, 112, 112, 3],
dtype=tf.float32)
train_phase_dropout = tf.placeholder(dtype=tf.bool,
shape=None,
name='train_phase')
train_phase_bn = tf.placeholder(dtype=tf.bool,
shape=None,
name='train_phase_last')
config = yaml.load(open('./configs/config_ms1m_100.yaml'))
image_ = images / 127.5 - 1
# image_ = tf.roll(image_, shift=2, axis=3)
inputs = tf.reshape(image_, [-1, *args.image_size])
embedding_tensor, _ = get_embd(inputs, train_phase_dropout, train_phase_bn,
config)
t_vars = tf.global_variables()
restore_vars = [var for var in t_vars if 'embd_extractor' in var.name]
### Graph & load setup
tf_config = tf.ConfigProto(allow_soft_placement=True)
tf_config.gpu_options.allow_growth = True
with tf.Session(config=tf_config) as sess:
tf.global_variables_initializer().run()
saver = tf.train.Saver(var_list=restore_vars)
saver.restore(sess, './model/best-m-1006000')
train_embeddings = np.zeros((6000, 2, 512))
for i, imgs in enumerate(train_images):
embd_dict = {
images: imgs,
train_phase_dropout: False,
train_phase_bn: False
}
embeddings = sess.run(embedding_tensor, feed_dict=embd_dict)
train_embeddings[i] = embeddings
return train_embeddings, train_labels