if epoch % self.args.save_freq == 0:
filename = 'epoch_%d_lfw_%.4f.pth.tar' % (epoch, max_lfw_acc)
savename = os.path.join(self.args.save_to, filename)
torch.save(
{
'epoch': epoch,
'backbone': self.model['backbone'].state_dict(),
'metric': self.model['metric'].state_dict(),
'lfw_acc': max_lfw_acc,
}, savename)
if self.args.is_debug:
break
def train_runner(self):
self._report_settings()
self._model_loader()
self._data_loader()
self._main_loop()
if __name__ == "__main__":
faceu = MetricFace(training_args())
faceu.train_runner()
shutil.copy(save_name, normal_name)
def lgsc_training(self):
self.result['max_acc'] = -1.0
self.result['min_loss'] = 100
for epoch in range(self.args.start_epoch, self.args.n_epoches + 1):
start_time = time.time()
self.result['epoch'] = epoch
train_loss = self.train_one_epoch()
test_info = self.test_one_epoch()
finish_time = time.time()
print('single epoch costs %.4f mins' %
((finish_time - start_time) / 60))
self.save_weights(test_info)
if self.args.is_debug:
break
def main_runner(self):
self._model_loader()
self._data_loader()
self.lgsc_training()
if __name__ == "__main__":
lgsc = LGSCTrainer(args=training_args())
lgsc.main_runner()
self.result['opt_thresh'] = -1.0
self.result['sota_acc'] = 0.0
for epoch in range(self.args.start_epoch, self.args.end_epoch + 1):
start_time = time.time()
self.result['epoch'] = epoch
train_loss = self._train_one_epoch()
self.model['scheduler'].step()
eval_info = self._evaluate_one_epoch(loader='lfw')
end_time = time.time()
print('Single epoch cost time : %.2f mins' %
((end_time - start_time) / 60))
if self.args.is_debug:
break
def train_runner(self):
self._report_settings()
self._model_loader()
self._data_loader()
self._point_estimate_training()
if __name__ == "__main__":
faceu = PointEstimateTrainer(training_args())
faceu.train_runner()
for epoch in range(self.args.start_epoch, self.args.end_epoch + 1):
start_time = time.time()
self.result['epoch'] = epoch
train_loss = self._train_one_epoch(epoch)
self.model['scheduler'].step()
eval_info = self._evaluate_one_epoch(loader='lfw')
end_time = time.time()
print('Single epoch cost time : %.2f mins' %
((end_time - start_time) / 60))
self._save_weights(eval_info)
if self.args.is_debug:
break
def train_runner(self):
self._report_settings()
self._model_loader()
self._data_loader()
self._dul_training()
if __name__ == "__main__":
dul_cls = DulClsTrainer(training_args())
dul_cls.train_runner()
val_loss = self._model_eval()
if val_loss < min_loss:
min_loss = val_loss
filename = os.path.join(str(self.args.snapshot),
'sota.pth.tar')
torch.save(
{
'epoch': epoch,
'backbone': self.model['backbone'].state_dict(),
'auxilnet': self.model['auxilnet'].state_dict()
}, filename)
print('sota performance was updated ...')
self.model['scheduler'].step()
def train_runner(self):
self._report_settings()
self._model_loader()
self._data_loader()
self._main_loop()
if __name__ == "__main__":
pfld_trainer = PFLDTrain(training_args())
pfld_trainer.train_runner()