def train_epoch(self):
avg_loss_all, avg_loss_law, avg_loss_accu, avg_loss_impris = 0., 0., 0., 0.
avg_acc_impris = 0.
avg_macro_f1_law, avg_micro_f1_law, avg_weighted_f1_law = 0., 0., 0.
avg_macro_f1_accu, avg_micro_f1_accu, avg_weighted_f1_accu = 0., 0., 0.
while not self.sv.should_stop():
loss_law, loss_accu, loss_impris, loss_l2, loss_all, \
macro_f1_law, micro_f1_law, weighted_f1_law, \
macro_f1_accu, micro_f1_accu, weighted_f1_accu, \
acc_impris, summaries = self.train_step()
global_steps = self.model.global_step_tensor.eval(self.sess)
avg_loss_law = running_avg_loss(avg_loss_law, loss_law)
avg_loss_accu = running_avg_loss(avg_loss_accu, loss_accu)
avg_loss_impris = running_avg_loss(avg_loss_impris, loss_impris)
avg_loss_all = running_avg_loss(avg_loss_all, loss_all)
avg_macro_f1_law = running_avg_loss(avg_macro_f1_law, macro_f1_law)
avg_micro_f1_law = running_avg_loss(avg_micro_f1_law, micro_f1_law)
avg_weighted_f1_law = running_avg_loss(avg_weighted_f1_law,
weighted_f1_law)
avg_macro_f1_accu = running_avg_loss(avg_macro_f1_accu,
macro_f1_accu)
avg_micro_f1_accu = running_avg_loss(avg_micro_f1_accu,
micro_f1_accu)
avg_weighted_f1_accu = running_avg_loss(avg_weighted_f1_accu,
weighted_f1_accu)
avg_acc_impris = running_avg_loss(avg_acc_impris, acc_impris)
self.summary_writer.add_summary(summaries,
global_step=global_steps)
if global_steps % 100 == 0:
print(
"global_steps:{}\tloss_all:{:.3f}\tloss_law:{:.3f}\tloss_accu:{:.3f}\tloss_impris:{:.3f}"
"\tmacro_f1_law:{:.3f}\tmicro_f1_law:{:.3f}\tweighted_f1_law:{:.3f}"
"\tavg_macro_f1_accu:{:.3f}\tavg_micro_f1_accu:{:.3f}\tavg_weighted_f1_accu:{:.3f}\t"
"acc_impris:{}".format(
global_steps, avg_loss_all, avg_loss_law,
avg_loss_accu, avg_loss_impris, avg_macro_f1_law,
avg_micro_f1_law, avg_weighted_f1_law,
avg_macro_f1_accu, avg_micro_f1_accu,
avg_weighted_f1_accu, avg_acc_impris))
self.summary_writer.flush()
self.sv.Stop()
def eval_epoch(self):
avg_loss_all, avg_loss_law, avg_loss_accu, avg_loss_impris = 0., 0., 0., 0.
best_avg_loss_all, best_avg_loss_law, best_avg_loss_accu, best_avg_loss_impris = \
float('inf'), float('inf'), float('inf'), float('inf')
avg_acc_impris = 0.
avg_macro_f1_law, avg_micro_f1_law, avg_weighted_f1_law = 0., 0., 0.
avg_macro_f1_accu, avg_micro_f1_accu, avg_weighted_f1_accu = 0., 0., 0.
self.sess.run(tf.local_variables_initializer())
coord = tf.train.Coordinator()
threads = tf.train.start_queue_runners(sess=self.sess, coord=coord)
try:
while not coord.should_stop():
time.sleep(self.config.eval_interval_secs)
self.model.load_ckpt(self.sess)
loss_law, loss_accu, loss_impris, loss_l2, loss_all, \
macro_f1_law, micro_f1_law, weighted_f1_law, \
macro_f1_accu, micro_f1_accu, weighted_f1_accu, \
acc_impris, summaries = self.eval_step()
global_steps = self.model.global_step_tensor.eval(self.sess)
avg_loss_law = running_avg_loss(avg_loss_law, loss_law)
avg_loss_accu = running_avg_loss(avg_loss_accu, loss_accu)
avg_loss_impris = running_avg_loss(avg_loss_impris,
loss_impris)
avg_loss_all = running_avg_loss(avg_loss_all, loss_all)
avg_macro_f1_law = running_avg_loss(avg_macro_f1_law,
macro_f1_law)
avg_micro_f1_law = running_avg_loss(avg_micro_f1_law,
micro_f1_law)
avg_weighted_f1_law = running_avg_loss(avg_weighted_f1_law,
weighted_f1_law)
avg_macro_f1_accu = running_avg_loss(avg_macro_f1_accu,
macro_f1_accu)
avg_micro_f1_accu = running_avg_loss(avg_micro_f1_accu,
micro_f1_accu)
avg_weighted_f1_accu = running_avg_loss(
avg_weighted_f1_accu, weighted_f1_accu)
avg_acc_impris = running_avg_loss(avg_acc_impris, acc_impris)
self.summary_writer.add_summary(summaries,
global_step=global_steps)
print(
"global_steps:{}\tloss_all:{:.3f}\tloss_law:{:.3f}\tloss_accu:{:.3f}\tloss_impris:{:.3f}"
"\tmacro_f1_law:{:.3f}\tmicro_f1_law:{:.3f}\tweighted_f1_law:{:.3f}"
"\tavg_macro_f1_accu:{:.3f}\tavg_micro_f1_accu:{:.3f}\tavg_weighted_f1_accu:{:.3f}\t"
"acc_impris:{}".format(
global_steps, avg_loss_all, avg_loss_law,
avg_loss_accu, avg_loss_impris, avg_macro_f1_law,
avg_micro_f1_law, avg_weighted_f1_law,
avg_macro_f1_accu, avg_micro_f1_accu,
avg_weighted_f1_accu, avg_acc_impris))
self.summary_writer.flush()
if global_steps == 0:
continue
if "impris" in self.config.train_tasks and best_avg_loss_impris > avg_loss_impris:
tf.logging.info('Found new best model (%f vs. %f)',
running_avg_loss, best_avg_loss_all)
best_avg_loss_impris = avg_loss_impris
self.model.saver.save(self.sess,
self.config.checkpoint_dir +
'/model.best',
write_state=False)
elif "law" in self.config.train_tasks and best_avg_loss_law > avg_loss_law:
tf.logging.info('Found new best model (%f vs. %f)',
running_avg_loss, best_avg_loss_all)
best_avg_loss_law = avg_loss_law
self.model.saver.save(self.sess,
self.config.checkpoint_dir +
'/model.best',
write_state=False)
elif "accu" in self.config.train_tasks and best_avg_loss_accu > avg_loss_accu:
tf.logging.info('Found new best model (%f vs. %f)',
running_avg_loss, best_avg_loss_all)
best_avg_loss_accu = avg_loss_accu
self.model.saver.save(self.sess,
self.config.checkpoint_dir +
'/model.best',
write_state=False)
elif "all" in self.config.train_tasks and best_avg_loss_all > avg_loss_all:
tf.logging.info('Found new best model (%f vs. %f)',
running_avg_loss, best_avg_loss_all)
best_avg_loss_all = avg_loss_all
self.model.saver.save(self.sess,
self.config.checkpoint_dir +
'/model.best',
write_state=False)
except tf.errors.OutOfRangeError as e:
tf.logging.info('\nDone Training, step limit reached.')
finally:
coord.request_stop()
coord.join(threads)