def train(self, sess: tf.Session, data: DataHelper, restore_file=None, log_path='data'):
actor_loss, critic_loss = self._build_loss()
actor_opt = self.opt(actor_loss, 'actor')
critic_opt = self.opt(critic_loss, 'critic')
initial_op = tf.global_variables_initializer()
sess.run(initial_op)
time_string = strftime("%a%d%b%Y-%H%M%S", gmtime())
summary_path = os.path.join(log_path, 'log', time_string) + os.sep
save_path = os.path.join(log_path, 'model') + os.sep
if not os.path.exists(save_path):
os.mkdir(save_path)
if restore_file is not None:
self._restore(sess, restore_file)
writer = tf.summary.FileWriter(summary_path)
actor_summary = tf.summary.merge(tf.get_collection(tf.GraphKeys.SUMMARIES, scope='actor'))
critic_summary = tf.summary.merge(tf.get_collection(tf.GraphKeys.SUMMARIES, scope='critic'))
for i in range(50000):
train_batch = data.next_batch('train')
train_dict = {self.image_in: train_batch['batch_image']}
critic_value, critic_summary_value, critic_step = sess.run(
[critic_loss, critic_summary, self.global_step], feed_dict=train_dict)
_, actor_value, actor_summary_value, code_value, actor_step = sess.run(
[actor_opt, actor_loss, actor_summary, self.net['codes'], self.global_step], feed_dict=train_dict)
writer.add_summary(critic_summary_value, critic_step)
writer.add_summary(actor_summary_value, actor_step)
data.update(code_value)
if (i + 1) % 100 == 0:
hook_train = data.hook_train()
hook_summary = tf.Summary(value=[tf.Summary.Value(tag='hook/train', simple_value=hook_train)])
print('batch {}: actor {}, critic {}'.format(i, actor_value, critic_value))
writer.add_summary(hook_summary, actor_step)
if (i + 1) % 3000 == 0:
print('Testing!!!!!!!!')
test_batch = data.next_batch('test')
test_dict = {self.image_in: test_batch['batch_image']}
test_code = sess.run(self.net['codes'], feed_dict=test_dict)
data.update(test_code, phase='test')
hook_test = data.hook_test()
hook_summary = tf.Summary(value=[tf.Summary.Value(tag='hook/test', simple_value=hook_test)])
writer.add_summary(hook_summary, actor_step)
if (i + 1) % 3000 == 0:
self._save(sess, save_path, actor_step)
def extract_image(self, sess: tf.Session, data: DataHelper):
import numpy as np
import scipy.io as sio
rslt = np.zeros((data.training_data.batch_num * data.training_data.batch_size, self.input_length))
for i in range(data.training_data.batch_num + 1):
this_batch = data.next_batch('train')
this_dict = {self.image_in: this_batch['batch_image']}
code = sess.run(self.net['decode_result'], feed_dict=this_dict)
rslt[this_batch['batch_start']:this_batch['batch_end']] = code
to_save = {'set_image': rslt,
'set_label': data.training_data.label}
sio.savemat('/home/ymcidence/Workspace/CodeGeass/MatlabWorkspace/sgh_code/mnist_sgh.mat', to_save)
def extract(self,
sess: tf.Session,
data: DataHelper,
log_path='data',
task='cifar'):
for i in range(data.training_data.batch_num + 1):
this_batch = data.next_batch('train')
this_dict = {self.image_in: this_batch['batch_image']}
code = sess.run(self.net['codes'], feed_dict=this_dict)
data.update(code)
for i in range(data.test_data.batch_num + 1):
this_batch = data.next_batch('test')
this_dict = {self.image_in: this_batch['batch_image']}
code = sess.run(self.net['codes'], feed_dict=this_dict)
data.update(code, 'test')
data.save(task, self.code_length, folder=log_path)
if (i + 1) % 3000 == 0:
self._save(sess, save_path, actor_step)
if __name__ == '__main__':
from util.data.dataset import MatDataset
batch_size = 200
code_length = 32
train_file = '/home/ymcidence/Workspace/CodeGeass/MatlabWorkspace/train_mnist.mat'
test_file = '/home/ymcidence/Workspace/CodeGeass/MatlabWorkspace/test_mnist.mat'
model_config = {'batch_size': batch_size, 'code_length': code_length, 'input_length': 784}
train_config = {'batch_size': batch_size, 'code_length': code_length, 'file_name': train_file, 'phase': 'train'}
train_config2 = {'batch_size': batch_size, 'code_length': code_length, 'file_name': train_file, 'phase': 'test'}
test_config = {'batch_size': batch_size, 'code_length': code_length, 'file_name': test_file, 'phase': 'test'}
this_sess = tf.Session()
model = SGH(**model_config)
train_data = MatDataset(**train_config)
test_data = MatDataset(**test_config)
data_helper = DataHelper(train_data, test_data)
train_data2 = MatDataset(**train_config2)
data_helper2 = DataHelper(train_data2, test_data)
model.train(this_sess, data_helper)
model.extract_image(this_sess, data_helper2)
'code_length': code_length,
'file_name': train_file,
'phase': 'train'
}
test_config = {
'batch_size': batch_size,
'code_length': code_length,
'file_name': test_file,
'phase': 'train'
}
base_config = {
'batch_size': batch_size,
'code_length': code_length,
'file_name': base_file,
'phase': 'train'
}
sess = tf.Session()
model = BasicModel(**model_config)
train_data = H5Dataset(**train_config)
test_data = H5Dataset(**test_config)
test_data1 = H5Dataset(**test_config)
base_data = H5Dataset(**base_config)
data_helper = DataHelper(train_data, test_data)
base_helper = DataHelper(base_data, test_data1)
model.train(sess, data_helper, log_path='../../Log')
model.extract(sess, base_helper, log_path='../../Log', task=task)