Validation accuracy going down through epoch

1

I think I’m misunderstanding the use of metric reset/get or something, because as I progress through batches in an epoch on my validation set, the accuracy starts high and then steadily decreases.

valid_acc.reset()
cumulative_valid_loss = mx.nd.zeros(1, ctx)
valid_samples = 0
tbar = tqdm(valid_dl)
for batch_idx, (data, label) in enumerate(tbar):
  data = data.as_in_context(ctx)
  label = label.as_in_context(ctx)      
  output = self.net(data)
  loss = loss_fn(output, label)
  cumulative_valid_loss += loss.sum()
  valid_samples += data.shape[0]
  valid_loss = cumulative_valid_loss.asscalar()/valid_samples
  valid_acc.update(label, output)
  metric_name, metric_val = valid_acc.get()
  tbar.set_description(f'validation loss {valid_loss:.3f}, '
                       f'validation {metric_name}: {metric_val:.3f}')

Does anybody know why this is? All the examples I’ve seen only call metric.get() at the end of an epoch. Is there a reason for that?