def train(self, counter=1, gen_dirs=()):
if self.conf.need_to_load:
self.load(self.conf.checkpoint_dir, step=counter)
data = self.data
logger.info('Total amount of images: %s' % len(data))
# np.random.shuffle(data)
tf.initialize_all_variables().run()
# counter = 1
start_time = time.time()
batch_idxs = min(len(data), self.conf.train_size) / self.conf.batch_size
stego_accuracy = 0
accuracies = []
accuracies_steps = []
logger.debug('Starting updating')
for epoch in range(self.conf.epoch):
losses = []
np.random.shuffle(data)
logger.info('Starting epoch %s' % epoch)
for idx in range(0, int(batch_idxs)):
batch_files = data[idx * self.conf.batch_size:(idx + 1) * self.conf.batch_size]
batch = [get_image(batch_file, self.conf.image_size)
for batch_file in batch_files]
batch_images = np.array(batch).astype(np.float32)
batch_targets = self.get_targets(batch_files)
self.sess.run(self.optimize, feed_dict={self.images: batch_images, self.target: batch_targets})
loss = self.loss.eval({self.images: batch_images, self.target: batch_targets})
losses.append(loss)
# logger.debug("[ITERATION] Epoch [%2d], iteration [%4d/%4d] time: %4.4f, Loss: %8f, accuracy: %8f" %
# (epoch, idx, batch_idxs, time.time() - start_time, loss, stego_accuracy))
counter += 1
if counter % 300 == 0:
logger.info('------')
stego_accuracy = self.accuracy(n_files=-1, test_dir=self.test_dir)
logger.info('[TEST] Epoch {:2d} accuracy: {:3.1f}%'.format(epoch + 1, 100 * stego_accuracy))
for gen_dir in gen_dirs:
gen_accuracy = self.accuracy(n_files=-1, test_dir=gen_dir)
logger.info('[GEN_TEST] Folder {}, accuracy: {:3.1f}%'.format(gen_dir, 100 * gen_accuracy))
def save(self, checkpoint_dir, step):
model_dir = "%s_%s" % (self.conf.model_name, self.conf.batch_size)
checkpoint_dir = os.path.join(checkpoint_dir, model_dir)
if not os.path.exists(checkpoint_dir):
os.makedirs(checkpoint_dir)
ckpt_name = '%s_%s.ckpt' % (self.conf.model_name, step)
logger.info('[SAVING] step: %s, name: %s' % (step, ckpt_name))
self.saver.save(self.sess, os.path.join(checkpoint_dir, ckpt_name), global_step=step)
def main(_):
logger.info('====================================================')
logger.info('===================NEW EXPERIMENT===================')
logger.info('====================================================')
logger.info(flags.FLAGS.__flags)
if not os.path.exists(FLAGS.checkpoint_dir):
os.makedirs(FLAGS.checkpoint_dir)
if not os.path.exists(FLAGS.sample_dir):
os.makedirs(FLAGS.sample_dir)
with tf.Session() as sess:
steganalisys = Steganalyzer(config=FLAGS,
sess=sess,
stego_algorithm=LSBMatching)
# dirs = [
# '/home/dvolkhonskiy/SGAN/code/data/10_seeds/test',
# '/home/dvolkhonskiy/SGAN/code/data/10_seeds/other_seeds',
# '/home/dvolkhonskiy/SGAN/code/data/10_seeds/more_train',
# '/home/dvolkhonskiy/SGAN/code/data/10_seeds/more_train_other_seeds',
# ]
# dirs = [
# '/home/dvolkhonskiy/SGAN/code/data/overtraining/plus_0',
# '/home/dvolkhonskiy/SGAN/code/data/overtraining/plus_1',
# '/home/dvolkhonskiy/SGAN/code/data/overtraining/plus_2',
# '/home/dvolkhonskiy/SGAN/code/data/overtraining/plus_3',
# '/home/dvolkhonskiy/SGAN/code/data/overtraining/plus_4',
# '/home/dvolkhonskiy/SGAN/code/data/overtraining/plus_5',
# '/home/dvolkhonskiy/SGAN/code/data/overtraining/plus_6',
# '/home/dvolkhonskiy/SGAN/code/data/overtraining/plus_7',
# '/home/dvolkhonskiy/SGAN/code/data/overtraining/plus_8',
# '/home/dvolkhonskiy/SGAN/code/data/overtraining/plus_9',
# ]
dirs = [
'/home/dvolkhonskiy/SGAN/code/data/overtraining/plus_%s' % i
for i in range(0, 66)
]
if FLAGS.is_train:
steganalisys.train(counter=1, gen_dirs=dirs)
else:
# steganalisys.load(FLAGS.checkpoint_dir, step=28481) # LSB matching clf
steganalisys.load(FLAGS.checkpoint_dir, step=59)
for gen_dir in dirs:
acc, std = steganalisys.accuracy(n_files=-1, test_dir=gen_dir)
logger.info('[GEN_TEST] Folder %s, accuracy: %.4f, std: %.4f' %
(gen_dir, acc, std))
def load(self, checkpoint_dir, step):
model_dir = "%s_%s" % (self.conf.model_name, self.conf.batch_size)
checkpoint_dir = os.path.join(checkpoint_dir, model_dir)
try:
ckpt_name = '%s_%s.ckpt-%s' % (self.conf.model_name, step, step)
logger.info('[LOADING] step: %s, name: %s' % (step, ckpt_name))
self.saver.restore(self.sess, os.path.join(checkpoint_dir, ckpt_name))
except Exception as e:
logger.debug(e)
ckpt_name = 'StegoDCGAN-%s' % (step)
logger.info('[LOADING] step: %s, name: %s' % (step, ckpt_name))
self.saver.restore(self.sess, os.path.join(checkpoint_dir, ckpt_name))
def accuracy(self, test_dir='test', abs=False, n_files=2 ** 12):
logger.info('[TEST], test data folder: %s, n_files: %s' % (test_dir, 2 * n_files))
X_test = self.get_images_names('%s/*.%s' % (test_dir, self.conf.img_format), abs=abs)[:n_files]
accuracies = []
batch_idxs = min(len(X_test), self.conf.train_size) / self.conf.batch_size
# logger.debug('Starting iteration')
for idx in range(0, int(batch_idxs)):
batch_files_stego = X_test[idx * self.conf.batch_size:(idx + 1) * self.conf.batch_size]
batch = [get_image(batch_file, self.conf.image_size) for batch_file in batch_files_stego]
batch_images = np.array(batch).astype(np.float32)
batch_targets = self.get_targets(batch_files_stego)
accuracies.append(self.get_accuracy(batch_images, batch_targets))
return np.mean(accuracies)
def train(self, counter=1, gen_dirs=()):
if self.conf.need_to_load:
self.load(self.conf.checkpoint_dir, step=counter)
data = self.data
logger.info('Total amount of images: %s' % len(data))
# np.random.shuffle(data)
tf.initialize_all_variables().run()
# counter = 1
start_time = time.time()
batch_idxs = min(len(data),
self.conf.train_size) / self.conf.batch_size
logger.debug('Starting updating')
for epoch in range(self.conf.epoch):
losses = []
np.random.shuffle(data)
logger.info('Starting epoch %s' % epoch)
for idx in range(0, int(batch_idxs)):
batch_files = data[idx * self.conf.batch_size:(idx + 1) *
self.conf.batch_size]
batch = [
get_image(batch_file, self.conf.image_size)
for batch_file in batch_files
]
batch_images = np.array(batch).astype(np.float32)
batch_targets = self.get_targets(batch_files)
_, loss = self.sess.run([self.optimize, self.loss],
feed_dict={
self.images: batch_images,
self.target: batch_targets
})
losses.append(loss)
# logger.debug("[ITERATION] Epoch [%2d], iteration [%4d/%4d] time: %4.4f, Loss: %8f, accuracy: %8f" %
# (epoch, idx, batch_idxs, time.time() - start_time, loss, stego_accuracy))
counter += 1
if epoch % 5 == 0 and epoch > 0:
for gen_dir in gen_dirs:
acc, std = self.accuracy(n_files=-1, test_dir=gen_dir)
logger.info(
'[GEN_TEST] Folder %s, accuracy: %.4f, std: %.4f' %
(gen_dir, acc, std))
# SAVE after each epoch
self.save(self.conf.checkpoint_dir, epoch)
def main(_):
logger.info('====================================================')
logger.info('===================NEW EXPERIMENT===================')
logger.info('====================================================')
logger.info(flags.FLAGS.__flags)
if not os.path.exists(FLAGS.checkpoint_dir):
os.makedirs(FLAGS.checkpoint_dir)
if not os.path.exists(FLAGS.sample_dir):
os.makedirs(FLAGS.sample_dir)
with tf.Session() as sess:
dcgan = SGAN(sess, LSBMatching, config=FLAGS,
image_size=FLAGS.image_size,
batch_size=FLAGS.batch_size)
dcgan.load(FLAGS.checkpoint_dir, 4) # 18990 22155 25320 28485 31650
n_batches = 1999
for i in range(n_batches):
if i % 100 == 0:
print('Iteration', i)
z_sample = np.random.uniform(-1, 1, size=(FLAGS.batch_size, dcgan.z_dim))
samples = sess.run(dcgan.sampler, feed_dict={dcgan.z: z_sample})
save_images(samples, i, folder='./data/overtraining/')
for chpt_i, chpt in enumerate(range(4, 70, 1)): # [18990, 22155, 25320, 28485, 31650]
dcgan.load(FLAGS.checkpoint_dir, chpt)
n_batches = 199
for i in range(n_batches):
if i % 100 == 0:
print('Iteration', i)
z_sample = np.random.uniform(-1, 1, size=(FLAGS.batch_size, dcgan.z_dim))
samples = sess.run(dcgan.sampler, feed_dict={dcgan.z: z_sample})
folder = './data/overtraining/plus_%s' % (chpt_i)
if not os.path.exists(folder):
os.makedirs(folder)
save_images(samples, i, folder=folder)
def main(_):
logger.info('====================================================')
logger.info('===================NEW EXPERIMENT===================')
logger.info('====================================================')
logger.info(flags.FLAGS.__flags)
if not os.path.exists(FLAGS.checkpoint_dir):
os.makedirs(FLAGS.checkpoint_dir)
if not os.path.exists(FLAGS.sample_dir):
os.makedirs(FLAGS.sample_dir)
with tf.Session() as sess:
dcgan = SGAN(sess,
LSBMatching,
config=FLAGS,
image_size=FLAGS.image_size,
batch_size=FLAGS.batch_size)
if FLAGS.is_train:
dcgan.train()
else:
dcgan.load(FLAGS.checkpoint_dir, 50400) # 50400 50800
n_batches = 200
z_sample = np.random.uniform(-1,
1,
size=(FLAGS.batch_size, dcgan.z_dim))
for i in range(n_batches):
np.random.seed(int(time()))
z_sample = np.random.uniform(-1,
1,
size=(FLAGS.batch_size, dcgan.z_dim))
samples = sess.run(dcgan.sampler, feed_dict={dcgan.z: z_sample})
save_images(
samples,
i,
folder='/home/dvolkhonskiy/datasets/new/sgan_generated')
def main(_):
logger.info('====================================================')
logger.info('===================NEW EXPERIMENT===================')
logger.info('====================================================')
logger.info(flags.FLAGS.__flags)
if not os.path.exists(FLAGS.checkpoint_dir):
os.makedirs(FLAGS.checkpoint_dir)
if not os.path.exists(FLAGS.sample_dir):
os.makedirs(FLAGS.sample_dir)
with tf.Session() as sess:
steganalisys = Steganalyzer(config=FLAGS, sess=sess, stego_algorithm=LSBMatching, stego_name='lsb_matching')
if FLAGS.is_train:
steganalisys.train(counter=1, gen_dirs=['gen_test_seed_666', 'gen_test_more_train'])
else:
# steganalisys.load(FLAGS.checkpoint_dir, step=28481) # LSB matching clf
steganalisys.load(FLAGS.checkpoint_dir, step=122114)
print('ACCURACY:::::::::%s' % steganalisys.accuracy(test_dir='gen_test_more_train', n_files=-1))
def main(_):
logger.info('====================================================')
logger.info('===================NEW EXPERIMENT===================')
logger.info('====================================================')
logger.info(flags.FLAGS.__flags)
if not os.path.exists(FLAGS.checkpoint_dir):
os.makedirs(FLAGS.checkpoint_dir)
if not os.path.exists(FLAGS.sample_dir):
os.makedirs(FLAGS.sample_dir)
with tf.Session() as sess:
dcgan = SGAN(sess,
LSBMatching,
config=FLAGS,
image_size=FLAGS.image_size,
batch_size=FLAGS.batch_size)
if FLAGS.is_train:
dcgan.train(start_epoch=0)
else:
dcgan.load(FLAGS.checkpoint_dir, 50400) # 50400 50800
