def _eval_steps(self):
eval_tqdm = tqdm(initial=0,
total=len(self.dataloader["eval"]),
desc="eval")
for batch in self.dataloader["eval"]:
batch = to_device(batch, self.device)
self.eval(batch)
eval_tqdm.update(1)
eval_tqdm.close()
def _dev_step(self):
if (self.steps % self.conf["dev_steps"] == 0
and self.steps > self.conf["dev_steps"] - 1
and self.steps != self.resume_steps):
dev_loss_values = self._get_loss_dict()
for dev_idx, batch in enumerate(self.dataloader["dev"]):
batch = to_device(batch, self.device)
dev_loss_values = self.dev(batch)
if dev_idx > 0:
break
self._print_loss_values(dev_loss_values, phase="dev")
def _reconstruction_steps(self, tdir=False):
for dkey in ["train", "dev"]:
recon_tqdm = tqdm(
initial=0,
total=len(self.dataloader[dkey]),
desc="reconstruction ({})".format(dkey),
)
for batch in self.dataloader[dkey]:
batch = to_device(batch, self.device)
self.reconstruction(batch, tdir="reconstruction")
recon_tqdm.update(1)
recon_tqdm.close()
def _tr_step(self):
for batch in self.dataloader["train"]:
batch = to_device(batch, self.device)
loss_values = self.train(batch, phase="train")
if self.steps % self.conf["n_steps_print_loss"] == 0:
self._print_loss_values(loss_values, phase="train")
self._dev_step()
# check step-by-step
self._check_save_model()
self._step_update()
self._check_finish()
# check custum func in each child
self.check_custom_start()