kwargs = {'batch_size': 512, 'num_workers': 6, 'persistent_workers': True}
subgraph_loader = NeighborLoader(copy.copy(data), input_nodes=None,
                                     num_neighbors=[-1], shuffle=False, **kwargs)

for _ in range(n_epochs):
    model.train()

    total_loss = total_correct = total_examples = 0
    for batch in train_loader:
        optimizer.zero_grad()
        y = batch.y[:batch.batch_size]
        y_hat = model(batch.x, batch.edge_index)[:batch.batch_size]
        loss = F.cross_entropy(y_hat, y)
        loss.backward()
        optimizer.step()

        total_loss += float(loss) * batch.batch_size
        total_correct += int((y_hat.argmax(dim=-1) == y).sum())
        total_examples += batch.batch_size

    model.eval()
    val_acc = val_correct = val_examples = 0.
    with torch.no_grad():
        for batch in val_loader:
            y = batch.y[:batch.batch_size]
            y_hat = model(batch.x, batch.edge_index)[:batch.batch_size]

            val_correct += int((y_hat.argmax(dim=-1) == y).sum())
            val_examples += batch.batch_size

        for i in range(2):
            for batch in subgraph_loader:
                pass

If I keep this code below, the model is 83.55% accurate in the final test:

for i in range(2):
    for batch in subgraph_loader:
        pass

If I remove this code, the model is 77.27% accurate in the final test.

I would like to know what the cause of this phenomenon is.