def __init__(self, config, rng):
self.config = config
self.rng = rng
self.task = config.task
self.model_dir = config.model_dir
self.gpu_memory_fraction = config.gpu_memory_fraction
self.log_step = config.log_step
self.max_step = config.max_step
self.num_log_samples = config.num_log_samples
self.checkpoint_secs = config.checkpoint_secs
if config.task.lower().startswith('tsp'):
self.data_loader = TSPDataLoader(config, rng=self.rng)
else:
raise Exception("[!] Unknown task: {}".format(config.task))
self.summary_writer = tf.summary.FileWriter(self.model_dir)
self.model = Model(config,
inputs=self.data_loader.x,
labels=self.data_loader.y,
enc_seq_length=self.data_loader.seq_length,
dec_seq_length=self.data_loader.seq_length,
mask=self.data_loader.mask)
self.build_session()
show_all_variables()
def __init__(self, config, rng):
self.config = config
self.rng = rng
self.task = config.task # tsp
self.model_dir = config.model_dir
self.gpu_memory_fraction = config.gpu_memory_fraction # 1.0
self.log_step = config.log_step # 50
self.max_step = config.max_step # 1000000
self.num_log_samples = config.num_log_samples # 3
self.checkpoint_secs = config.checkpoint_secs # 300
if config.task.lower().startswith('tsp'):
self.data_loader = TSPDataLoader(config, rng=self.rng)
else:
raise Exception("[!] Unknown task: {}".format(config.task))
self.model = Model(
config,
inputs=self.data_loader.x,
labels=self.data_loader.y,
enc_seq_length=self.data_loader.seq_length, # 128
dec_seq_length=self.data_loader.seq_length,
mask=self.data_loader.mask) # dict, dict_keys(['train', 'test']
self.build_session()
show_all_variables()
def __init__(self, config, rng):
self.config = config
self.rng = rng
self.task = config.task
self.model_dir = config.model_dir
self.gpu_memory_fraction = config.gpu_memory_fraction
self.log_step = config.log_step
self.max_step = config.max_step
self.num_log_samples = config.num_log_samples
self.checkpoint_secs = config.checkpoint_secs
if config.task.lower().startswith('tsp'):
self.data_loader = TSPDataLoader(config, rng=self.rng)
else:
raise Exception("[!] Unknown task: {}".format(config.task))
self.model = Model(
config,
inputs=self.data_loader.x,
labels=self.data_loader.y,
enc_seq_length=self.data_loader.seq_length,
dec_seq_length=self.data_loader.seq_length,
mask=self.data_loader.mask)
self.build_session()
show_all_variables()
def __init__(self, config, rng):
self.config = config
self.rng = rng
self.task = config.task
self.model_dir = config.model_dir
self.gpu_memory_fraction = config.gpu_memory_fraction
self.log_step = config.log_step # default 50
self.max_step = config.max_step # default 1e6
self.num_log_samples = config.num_log_samples # default 3
self.checkpoint_secs = config.checkpoint_secs # defaut 300
self.data_loader = TSPDataLoader(config, rng=self.rng)
self.summary_writer = tf.summary.FileWriter(self.model_dir)
self.model = Model(config, summary_writer=self.summary_writer)
self._build_session()
show_all_variables()
class Trainer(object):
def __init__(self, config, rng):
self.config = config
self.rng = rng
self.task = config.task # tsp
self.model_dir = config.model_dir
self.gpu_memory_fraction = config.gpu_memory_fraction # 1.0
self.log_step = config.log_step # 50
self.max_step = config.max_step # 1000000
self.num_log_samples = config.num_log_samples # 3
self.checkpoint_secs = config.checkpoint_secs # 300
if config.task.lower().startswith('tsp'):
self.data_loader = TSPDataLoader(config, rng=self.rng)
else:
raise Exception("[!] Unknown task: {}".format(config.task))
self.model = Model(
config,
inputs=self.data_loader.x,
labels=self.data_loader.y,
enc_seq_length=self.data_loader.seq_length, # 128
dec_seq_length=self.data_loader.seq_length,
mask=self.data_loader.mask) # dict, dict_keys(['train', 'test']
self.build_session()
show_all_variables()
def build_session(self):
self.saver = tf.train.Saver()
self.summary_writer = tf.summary.FileWriter(self.model_dir)
sv = tf.train.Supervisor(logdir=self.model_dir,
is_chief=True,
saver=self.saver,
summary_op=None,
summary_writer=self.summary_writer,
save_summaries_secs=300,
save_model_secs=self.checkpoint_secs,
global_step=self.model.global_step)
gpu_options = tf.GPUOptions(
per_process_gpu_memory_fraction=self.gpu_memory_fraction,
allow_growth=True) # seems to be not working
sess_config = tf.ConfigProto(allow_soft_placement=True,
gpu_options=gpu_options)
self.sess = sv.prepare_or_wait_for_session(config=sess_config)
def train(self):
tf.logging.info("Training starts...")
self.data_loader.run_input_queue(self.sess)
summary_writer = None
for k in trange(self.max_step, desc="train"):
fetch = {
'optim': self.model.optim,
}
# fetch是配置的graph的最终得到的值，可以理解为整个图, 所以这里最后要得到的就是model.optim的值
result = self.model.train(self.sess, fetch, summary_writer)
if result['step'] % self.log_step == 0:
self._test(self.summary_writer)
summary_writer = self._get_summary_writer(result)
self.data_loader.stop_input_queue()
def test(self):
tf.logging.info("Testing starts...")
self.data_loader.run_input_queue(self.sess)
for idx in range(10):
self._test(None)
self.data_loader.stop_input_queue()
def _test(self, summary_writer):
fetch = {
'loss': self.model.total_inference_loss,
'pred': self.model.dec_inference,
'true': self.model.dec_targets,
}
result = self.model.test(self.sess, fetch, summary_writer)
tf.logging.info("")
tf.logging.info("test loss: {}".format(result['loss']))
for idx in range(self.num_log_samples):
pred, true = result['pred'][idx], result['true'][idx]
tf.logging.info("test pred: {}".format(pred))
tf.logging.info("test true: {} ({})".format(true, np.array_equal(pred, true)))
if summary_writer:
summary_writer.add_summary(result['summary'], result['step'])
def _get_summary_writer(self, result):
if result['step'] % self.log_step == 0:
return self.summary_writer
else:
return None
class Trainer(object):
def __init__(self, config, rng):
self.config = config
self.rng = rng
self.task = config.task
self.model_dir = config.model_dir
self.gpu_memory_fraction = config.gpu_memory_fraction
self.log_step = config.log_step
self.max_step = config.max_step
self.num_log_samples = config.num_log_samples
self.checkpoint_secs = config.checkpoint_secs
if config.task.lower().startswith('tsp'):
self.data_loader = TSPDataLoader(config, rng=self.rng)
else:
raise Exception("[!] Unknown task: {}".format(config.task))
self.summary_writer = tf.summary.FileWriter(self.model_dir)
self.model = Model(config,
inputs=self.data_loader.x,
labels=self.data_loader.y,
enc_seq_length=self.data_loader.seq_length,
dec_seq_length=self.data_loader.seq_length,
mask=self.data_loader.mask)
self.build_session()
show_all_variables()
def build_session(self):
self.saver = tf.train.Saver()
sv = tf.train.Supervisor(logdir=self.model_dir,
is_chief=True,
saver=self.saver,
summary_op=None,
summary_writer=self.summary_writer,
save_summaries_secs=300,
save_model_secs=self.checkpoint_secs,
global_step=self.model.global_step)
gpu_options = tf.GPUOptions(
per_process_gpu_memory_fraction=self.gpu_memory_fraction,
allow_growth=True) # seems to be not working
sess_config = tf.ConfigProto(allow_soft_placement=True,
gpu_options=gpu_options)
sess_config.gpu_options.allow_growth = True
self.sess = sv.prepare_or_wait_for_session(config=sess_config)
if not self.config.is_train:
self.saver.restore(self.sess, self.config.load_path)
def train(self):
tf.logging.info("Training starts...")
self.data_loader.run_input_queue(self.sess)
test_data = np.load("test.data.npy")
loss_list = []
distance_list = []
accuracy_list = []
n_classes = 20
fetch = {
'optim': self.model.optim,
'loss': self.model.total_loss,
'logits': self.model.dec_pred_logits
}
data = self.model.get_data(self.sess, train=True)
for k in trange(self.max_step):
data = self.model.get_data(self.sess, train=True)
optimal_distance = np.zeros(128)
predict_distance = np.zeros(128)
num_nodes = 20
data_x = data["x"]
data_y = data["y"]
mask = np.ones((128, 20), dtype="bool")
distance = np.zeros(128)
for i in range(18):
decoder_input = data_x[np.arange(128).astype("int32"),
data_y[:, i].astype("int32")].reshape(
128, 1, 2)
step_data_x = data_x[mask].reshape(128, -1, 2)
one_hot_targets = np.eye(n_classes)[data_y[:, i +
1]][mask].reshape(
128, -1)
step_data_y = np.argmax(one_hot_targets, -1)
if num_nodes < 20:
step_data_x_padding = np.concatenate(
[step_data_x,
np.zeros((128, 20 - num_nodes, 2))], 1)
else:
step_data_x_padding = step_data_x
mask[np.arange(128).astype("int32"),
data_y[:, i + 1].astype("int32")] = 0
step_mask = np.zeros((128, 20))
step_mask[:, 1:num_nodes] = 1
feed_dict = {
self.model.x: step_data_x_padding,
self.model.y: step_data_y,
self.model.seq_length: np.ones(128) * num_nodes,
self.model.decoder_input: decoder_input,
self.model.step_mask: step_mask
}
result = self.model.run(self.sess, fetch, feed_dict)
num_nodes -= 1
loss_list.append(result['loss'])
accuracy_list.append(
np.mean(step_data_y == np.argmax(
result['logits'].reshape(128, 20) * step_mask, -1)))
next_input = data_x[np.arange(128).astype("int32"),
data_y[:, i + 1].astype("int32")].reshape(
128, 1, 2)
distance += np.sqrt(
np.sum(((next_input.reshape(128, 2) -
decoder_input.reshape(128, 2))**2), -1))
input1 = data_x[np.arange(128).astype("int32"),
data_y[:, -2].astype("int32")].reshape(128, 1, 2)
input2 = data_x[np.arange(128).astype("int32"),
data_y[:, -1].astype("int32")].reshape(128, 1, 2)
distance += np.sqrt(
np.sum((input1.reshape(128, 2) - input2.reshape(128, 2))**2,
-1))
input1 = data_x[np.arange(128).astype("int32"),
data_y[:, -1].astype("int32")].reshape(128, 1, 2)
input2 = data_x[np.arange(128).astype("int32"),
data_y[:, 0].astype("int32")].reshape(128, 1, 2)
distance += np.sqrt(
np.sum((input1.reshape(128, 2) - input2.reshape(128, 2))**2,
-1))
distance_list.append(np.mean(distance))
if k % self.log_step == 0:
print("")
print("loss", np.mean(loss_list), "opt_dist",
np.mean(distance_list), "accuracy",
np.mean(accuracy_list))
loss_list = []
distance_list = []
accuracy_list = []
self._test(test_data)
self.data_loader.stop_input_queue()
def test(self):
tf.logging.info("Testing starts...")
test_data = np.load("test.data.npy")
self._test(test_data)
print("------ done -------")
def _test(self, test_data):
start = time.time()
distance_list = []
fetch = {'logits': self.model.dec_pred_logits}
batch_size = 128
for j in trange(test_data.shape[0] // batch_size):
data = test_data[j * batch_size:(j + 1) * batch_size]
num_nodes = data.shape[1]
mask = np.ones((batch_size, num_nodes), dtype="bool")
decoder_input = data[:, 0].reshape(batch_size, 1, 2)
index_matrix = np.concatenate(
[np.arange(20).reshape(1, 20) for i in range(batch_size)], 0)
result_index = [np.zeros(data.shape[0])]
for i in range(18):
step_data_x = data[mask].reshape(batch_size, -1, 2)
if num_nodes < 20:
step_data_x_padding = np.concatenate([
step_data_x,
np.zeros((batch_size, 20 - num_nodes, 2))
], 1)
else:
step_data_x_padding = step_data_x
index_matrix_step = index_matrix[mask].reshape(batch_size, -1)
step_mask = np.zeros((batch_size, 20))
step_mask[:, 1:num_nodes] = 1
feed_dict = {
self.model.x: step_data_x_padding,
self.model.seq_length: np.ones(batch_size) * num_nodes,
self.model.decoder_input: decoder_input,
self.model.step_mask: step_mask
}
result = self.model.run(self.sess, fetch, feed_dict)
logits = result['logits'].reshape(batch_size, 20) * step_mask
prediction = np.argmax(logits, 1)
prediction = index_matrix_step[
np.arange(batch_size).astype("int32"),
prediction.astype("int32")]
mask[np.arange(batch_size).astype("int32"),
prediction.astype("int32")] = 0
num_nodes -= 1
next_input = data[np.arange(batch_size).astype("int32"),
prediction.astype("int32")].reshape(
batch_size, 1, 2)
decoder_input = next_input
result_index.append(prediction.reshape(-1))
mask[:, 0] = 0
last_ones_index = np.argmax(mask, -1)
last_selected_ones = data[np.arange(batch_size).astype("int"),
last_ones_index.astype("int")].reshape(
batch_size, 1, 2)
result_index.append(last_ones_index.reshape(-1))
result_index = np.stack(result_index, -1)
assert np.array_equal(
np.arange(20).reshape(1, -1).repeat(data.shape[0], 0),
np.sort(result_index, 1)), "Invalid tour"
d = data[np.arange(data.shape[0]).reshape(-1, 1),
result_index.astype(int)]
distance = np.sum(np.sqrt(np.sum((d - np.roll(d, 1, 1))**2, -1)),
-1)
distance_list.append(np.mean(distance))
print("test result", np.mean(distance_list), time.time() - start)
def _get_summary_writer(self, result):
if result['step'] % self.log_step == 0:
return self.summary_writer
else:
return None
class Trainer(object):
def __init__(self, config, rng):
self.config = config
self.rng = rng
self.task = config.task
self.model_dir = config.model_dir
self.gpu_memory_fraction = config.gpu_memory_fraction
self.log_step = config.log_step
self.max_step = config.max_step
self.num_log_samples = config.num_log_samples
self.checkpoint_secs = config.checkpoint_secs
self.summary_ops = {}
if config.task.lower().startswith('tsp'):
self.data_loader = TSPDataLoader(config, rng=self.rng)
else:
raise Exception("[!] Unknown task: {}".format(config.task))
self.models = {}
self.model = Model(config,
inputs=self.data_loader.x,
labels=self.data_loader.y,
enc_seq_length=self.data_loader.seq_length,
dec_seq_length=self.data_loader.seq_length,
mask=self.data_loader.mask)
self.build_session()
show_all_variables()
def build_session(self):
self.saver = tf.train.Saver()
self.summary_writer = tf.summary.FileWriter(self.model_dir)
sv = tf.train.Supervisor(logdir=self.model_dir,
is_chief=True,
saver=self.saver,
summary_op=None,
summary_writer=self.summary_writer,
save_summaries_secs=300,
save_model_secs=self.checkpoint_secs,
global_step=self.model.global_step)
gpu_options = tf.GPUOptions(
per_process_gpu_memory_fraction=self.gpu_memory_fraction,
allow_growth=True) # seems to be not working
sess_config = tf.ConfigProto(allow_soft_placement=True,
gpu_options=gpu_options)
self.sess = sv.prepare_or_wait_for_session(config=sess_config)
def train(self):
tf.logging.info("Training starts...")
self.data_loader.run_input_queue(self.sess)
summary_writer = None
for k in trange(self.max_step, desc="train"):
fetch = {
'optim': self.model.optim,
}
result = self.model.train(self.sess, fetch, summary_writer)
if result['step'] % self.log_step == 0:
fetch = {
'loss': self.model.total_loss,
'pred': self.model.dec_inference,
'targets': self.model.dec_targets,
}
result = self.model.test(self.sess, fetch, self.summary_writer)
tf.logging.info("loss: {}".format(result['loss']))
for idx in range(self.num_log_samples):
tf.logging.info("preds: {}".format(result['preds'][idx]))
tf.logging.info("targets: {}".format(
result['targets'][idx]))
self.summary_writer.add_summary(result['summary'],
result['step'])
summary_writer = self._get_summary_writer(result)
def test(self):
tf.logging.info("Testing starts...")
def _inject_summary(self, tag, feed_dict, step):
summaries = self.sess.run(self.summary_ops[tag], feed_dict)
self.summary_writer.add_summary(summaries['summary'], step)
path = os.path.join(self.config.sample_model_dir,
"{}.png".format(step))
imwrite(
path,
img_tile(summaries['output'],
tile_shape=self.config.sample_image_grid)[:, :, 0])
def _get_summary_writer(self, result):
if result['step'] % self.log_step == 0:
return self.summary_writer
else:
return None
class Trainer(object):
def __init__(self, config, rng):
self.config = config
self.rng = rng
self.task = config.task
self.model_dir = config.model_dir
self.gpu_memory_fraction = config.gpu_memory_fraction
self.log_step = config.log_step
self.max_step = config.max_step
self.num_log_samples = config.num_log_samples
self.checkpoint_secs = config.checkpoint_secs
if config.task.lower().startswith('tsp'):
self.data_loader = TSPDataLoader(config, rng=self.rng)
else:
raise Exception("[!] Unknown task: {}".format(config.task))
self.model = Model(
config,
inputs=self.data_loader.x,
labels=self.data_loader.y,
enc_seq_length=self.data_loader.seq_length,
dec_seq_length=self.data_loader.seq_length,
mask=self.data_loader.mask)
self.build_session()
show_all_variables()
def build_session(self):
self.saver = tf.train.Saver()
self.summary_writer = tf.summary.FileWriter(self.model_dir)
sv = tf.train.Supervisor(logdir=self.model_dir,
is_chief=True,
saver=self.saver,
summary_op=None,
summary_writer=self.summary_writer,
save_summaries_secs=300,
save_model_secs=self.checkpoint_secs,
global_step=self.model.global_step)
gpu_options = tf.GPUOptions(
per_process_gpu_memory_fraction=self.gpu_memory_fraction,
allow_growth=True) # seems to be not working
sess_config = tf.ConfigProto(allow_soft_placement=True,
gpu_options=gpu_options)
self.sess = sv.prepare_or_wait_for_session(config=sess_config)
def train(self):
tf.logging.info("Training starts...")
self.data_loader.run_input_queue(self.sess)
summary_writer = None
for k in trange(self.max_step, desc="train"):
fetch = {
'optim': self.model.optim,
}
result = self.model.train(self.sess, fetch, summary_writer)
if result['step'] % self.log_step == 0:
self._test(self.summary_writer)
summary_writer = self._get_summary_writer(result)
self.data_loader.stop_input_queue()
def test(self):
tf.logging.info("Testing starts...")
self.data_loader.run_input_queue(self.sess)
for idx in range(10):
self._test(None)
self.data_loader.stop_input_queue()
def _test(self, summary_writer):
fetch = {
'loss': self.model.total_inference_loss,
'pred': self.model.dec_inference,
'true': self.model.dec_targets,
}
result = self.model.test(self.sess, fetch, summary_writer)
tf.logging.info("")
tf.logging.info("test loss: {}".format(result['loss']))
for idx in range(self.num_log_samples):
pred, true = result['pred'][idx], result['true'][idx]
tf.logging.info("test pred: {}".format(pred))
tf.logging.info("test true: {} ({})".format(true, np.array_equal(pred, true)))
if summary_writer:
summary_writer.add_summary(result['summary'], result['step'])
def _get_summary_writer(self, result):
if result['step'] % self.log_step == 0:
return self.summary_writer
else:
return None
class Trainer(object):
def __init__(self, config, rng):
self.config = config
self.rng = rng
self.task = config.task
self.model_dir = config.model_dir
self.gpu_memory_fraction = config.gpu_memory_fraction
self.log_step = config.log_step # default 50
self.max_step = config.max_step # default 1e6
self.num_log_samples = config.num_log_samples # default 3
self.checkpoint_secs = config.checkpoint_secs # defaut 300
self.data_loader = TSPDataLoader(config, rng=self.rng)
self.summary_writer = tf.summary.FileWriter(self.model_dir)
self.model = Model(config, summary_writer=self.summary_writer)
self._build_session()
show_all_variables()
def _build_session(self):
self.saver = tf.train.Saver()
sv = tf.train.Supervisor(logdir=self.model_dir,
is_chief=True,
saver=self.saver,
summary_op=None,
summary_writer=self.summary_writer,
save_summaries_secs=300,
save_model_secs=self.checkpoint_secs,
global_step=self.model.global_step)
gpu_options = tf.GPUOptions(
per_process_gpu_memory_fraction=self.gpu_memory_fraction,
allow_growth=True) # seems to be not working
sess_config = tf.ConfigProto(allow_soft_placement=True,
gpu_options=gpu_options)
self.sess = sv.prepare_or_wait_for_session(config=sess_config)
def rl_train(self):
tf.logging.info("Training starts...")
for k in trange(self.max_step, desc="train"):
x, _, y = self.data_loader.get_train_batch(self.config.batch_size)
do_summary = True if (k + 1) % self.log_step == 0 else None
# x_path: [batch_size, data_len]
# x_path = self.model.sample(self.sess, batch_points=x)
# 根据x和path生成reward = -len
# from data_loader import get_path_distance
# length = [get_path_distance(x[i,:,:], x_path[i,:])
# for i in range(self.config.batch_size)]
# self.model.train(self.sess, batch_points=x, target_path=x_path, td=y - length,
# do_summary=do_summary)
result = self.model.rl_sample_and_train(self.sess,
batch_points=x,
base_length=y,
do_summary=do_summary)
# self.data_loader.update_last_batch_critic(result['length']*0.3 + y*0.7)
# todo: better if data_loader单独提供一份测试数据集
if (k % 100) == 0:
x, _, y = self.data_loader.get_test_batch(
self.config.batch_size)
self.model.test(self.sess,
batch_points=x,
base_length=y,
do_summary=True)
def test(self):
"""用于训练完成以后单独test"""
tf.logging.info("Testing starts...")
loop_num = int(self.data_loader.data_num['test'] /
self.config.batch_size)
# 顺序遍历一遍数据集
for _ in trange(loop_num, desc="test"):
x, _, y = self.data_loader.get_test_batch(self.config.batch_size)
x_path, x_length = self.model.test(self.sess,
batch_points=x,
base_length=y,
do_summary=True)