def main(_):
loader = Loader(flag="azenuz_small")
config = Config(loader, flag="azenuz_small")
config.gpu = 0
if platform.system() == 'Linux':
gpuid = config.gpu
os.environ["CUDA_VISIBLE_DEVICES"] = '{}'.format(gpuid)
device = '/gpu:' + str(gpuid)
else:
device = '/cpu:0'
lr_updater = LearningRateUpdater(config.learning_rate, config.decay,
config.decay_epoch)
i = 0
graph = tf.Graph()
INFO_LOG("build model")
with graph.as_default():
trainm = CTR_ggnn(config, device, loader, "Train")
testm = CTR_ggnn(config, device, loader, "Valid")
INFO_LOG("finish build model")
session_config = tf.ConfigProto(allow_soft_placement=True,
log_device_placement=False)
session_config.gpu_options.allow_growth = True
with tf.Session(graph=graph, config=session_config) as session:
# session.run(tf.global_variables_initializer())
CTR_GNN_loader(session, config)
best_auc = 0.
best_logloss = 0.
best_epoch_auc = 0.
best_epoch_logloss = 0.
auc = 0.
for epoch in range(config.epoch_num):
trainm.update_lr(session, lr_updater.get_lr())
cost, auc = run(session, config, trainm, loader, verbose=True)
INFO_LOG("Epoch %d Train AUC %.3f" % (epoch + 1, auc))
INFO_LOG("Epoch %d Train costs %.3f" % (epoch, cost))
session.run(tf.local_variables_initializer())
cost, auc = run(session, config, testm, loader, verbose=True)
INFO_LOG("Epoch %d Valid AUC %.3f" % (epoch, auc))
INFO_LOG("Epoch %d Valid cost %.3f" % (epoch, cost))
# #
lr_updater.update(auc, epoch)
if best_auc < auc:
best_auc = auc
best_epoch_auc = epoch
CTR_GNN_saver(session, config, best_auc, best_epoch_auc)
if best_logloss > cost:
best_logloss = cost
best_epoch_logloss = epoch
# CTR_GNN_saver(session, config, best_epoch_logloss, best_epoch_logloss)
INFO_LOG("*** best AUC now is %.3f in %d epoch" %
(best_auc, best_epoch_auc))
INFO_LOG("*** best logloss now is %.3f in %d epoch" %
(best_logloss, best_epoch_logloss))
if epoch % 5 == 0 and epoch != 0:
loader.change_data_list(loader.increase_data_idx())
def main(_):
loader = Loader(flag="as")
config = Config(loader, flag="as")
if platform.system() == 'Linux':
gpuid = config.gpu
os.environ["CUDA_VISIBLE_DEVICES"] = '{}'.format(gpuid)
device = '/gpu:' + str(gpuid)
else:
device = '/cpu:0'
lr_updater = LearningRateUpdater(config.learning_rate, config.decay,
config.decay_epoch)
i = 0
graph = tf.Graph()
with graph.as_default():
trainm = DynGCN(config, device, loader, "Train")
testm = DynGCN(config, device, loader, "Valid")
session_config = tf.ConfigProto(allow_soft_placement=True,
log_device_placement=False)
session_config.gpu_options.allow_growth = True
with tf.Session(graph=graph, config=session_config) as session:
# print "!!!!!!!!!!!"
session.run(tf.global_variables_initializer())
# print "*********"
# trainm.load_last_time_embedding(loader.present_graph, session)
# CTR_GNN_loader(session, config)
best_f1_score = 0.
best_auc_score = 0.
best_epoch = 0
time_consume_t = 0.
sum_time_consume = 0.
for epoch in range(config.epoch_num):
trainm.update_lr(session, lr_updater.get_lr())
# session.run(tf.local_variables_initializer())
# #
cost, eavluation_result = run(session,
config,
trainm,
loader,
verbose=False)
INFO_LOG("Epoch %d Train " % epoch + str(eavluation_result),
epoch % 1 == 0)
INFO_LOG("Epoch %d Train costs %.3f" % (epoch, cost),
epoch % 100 == 0)
session.run(tf.local_variables_initializer())
cost, eavluation_result = run(session,
config,
testm,
loader,
verbose=False)
INFO_LOG("Epoch %d Valid " % epoch + str(eavluation_result),
epoch % 1 == 0)
INFO_LOG("Epoch %d Valid cost %.3f" % (epoch, cost),
epoch % 1 == 0)
# #
auc = eavluation_result['auc']
f1_score = eavluation_result["f1_score"]["micro_f1_score"]
lr_updater.update(f1_score, epoch)
if best_f1_score < f1_score:
best_f1_score = f1_score
best_epoch = epoch
DynGCN_saver(session, config, best_f1_score, best_epoch, "hep")
INFO_LOG(
"*** best f1_score now is %.5f in %d epoch" %
(best_f1_score, best_epoch), True)
INFO_LOG(
"BEST Epoch %d Valid " % epoch + str(eavluation_result),
True)
if best_auc_score < auc:
best_auc_score = auc
INFO_LOG(
"*** best auc now is %.5f in %d epoch" %
(best_auc_score, epoch), True)
INFO_LOG(
"*** best f1_score now is %.4f in %d epoch" %
(best_f1_score, best_epoch), epoch % 100 == 0)
# time_consume_t += 1.
# sum_time_consume += eavluation_result["time_consume"]
# print("TIME CONSUME *** ", sum_time_consume/ time_consume_t)
if epoch % 1 == 0 and epoch != 0:
loader.change_2_next_graph_date()
