def validate()

in detector/train.py [0:0]

• 30 lines of code
• 9 McCabe index (conditional complexity)

def validate(model: nn.Module, device: str, loader: DataLoader, votes=1, desc='Validation'):
    model.eval()

    validation_accuracy = 0
    validation_epoch_size = 0
    validation_loss = 0

    records = [record for v in range(votes) for record in tqdm(loader, desc=f'Preloading data ... {v}',
                                                               disable=dist.is_available() and dist.get_rank() > 0)]
    records = [[records[v * len(loader) + i] for v in range(votes)] for i in range(len(loader))]

    with tqdm(records, desc=desc, disable=distributed() and dist.get_rank() > 0) as loop, torch.no_grad():
        for example in loop:
            losses = []
            logit_votes = []

            for texts, masks, labels in example:
                texts, masks, labels = texts.to(device), masks.to(device), labels.to(device)
                batch_size = texts.shape[0]

                loss, logits = model(texts, attention_mask=masks, labels=labels)
                losses.append(loss)
                logit_votes.append(logits)

            loss = torch.stack(losses).mean(dim=0)
            logits = torch.stack(logit_votes).mean(dim=0)

            batch_accuracy = accuracy_sum(logits, labels)
            validation_accuracy += batch_accuracy
            validation_epoch_size += batch_size
            validation_loss += loss.item() * batch_size

            loop.set_postfix(loss=loss.item(), acc=validation_accuracy / validation_epoch_size)

    return {
        "validation/accuracy": validation_accuracy,
        "validation/epoch_size": validation_epoch_size,
        "validation/loss": validation_loss
    }