def _general_epoch_end(self, outputs: List[Dict], group: str) -> Dict:
with torch.no_grad():
mean_loss = torch.stack([out["loss"]
for out in outputs]).mean().item()
logs = {f"{group}/loss": mean_loss}
logs.update(
Statistic.union_statistics([
out["statistic"] for out in outputs
]).calculate_metrics(group))
self.log_dict(logs)
self.log(f"{group}_loss", mean_loss)
def training_step(self, batch: TokensBatch, batch_idx: int) -> Dict:
# (batch size, output size)
logits = self(batch.tokens, batch.tokens_per_label)
loss = F.nll_loss(logits, batch.labels)
log: Dict[str, Union[float, torch.Tensor]] = {"train/loss": loss}
with torch.no_grad():
_, preds = logits.max(dim=1)
statistic = Statistic().calculate_statistic(
batch.labels,
preds,
2,
)
batch_matric = statistic.calculate_metrics(group="train")
log.update(batch_matric)
self.log_dict(log)
self.log("f1",
batch_matric["train/f1"],
prog_bar=True,
logger=False)
return {"loss": loss, "statistic": statistic}
def training_step(self, batch: PathContextBatch,
batch_idx: int) -> Dict: # type: ignore
# [batch size; num_classes]
logits = self(batch.contexts, batch.contexts_per_label)
loss = F.cross_entropy(logits, batch.labels)
log = {"train/loss": loss}
with torch.no_grad():
_, preds = logits.max(dim=1)
statistic = Statistic().calculate_statistic(
batch.labels,
preds,
2,
)
batch_matric = statistic.calculate_metrics(group="train")
log.update(batch_matric)
self.log_dict(log)
self.log("f1",
batch_matric["train/f1"],
prog_bar=True,
logger=False)
return {"loss": loss, "statistic": statistic}
def validation_step(self, batch: TokensBatch, batch_idx: int) -> Dict:
# (batch size, output size)
logits = self(batch.tokens, batch.tokens_per_label)
loss = F.nll_loss(logits, batch.labels)
with torch.no_grad():
_, preds = logits.max(dim=1)
statistic = Statistic().calculate_statistic(
batch.labels,
preds,
2,
)
return {"loss": loss, "statistic": statistic}
def validation_step(self, batch: PathContextBatch,
batch_idx: int) -> Dict: # type: ignore
# [batch size; num_classes]
logits = self(batch.contexts, batch.contexts_per_label)
loss = F.cross_entropy(logits, batch.labels)
with torch.no_grad():
_, preds = logits.max(dim=1)
statistic = Statistic().calculate_statistic(
batch.labels,
preds,
2,
)
return {"loss": loss, "statistic": statistic}
