def __call__(self,
outputs: torch.Tensor,
targets: torch.LongTensor,
masks: torch.LongTensor = None,
relevant_ignored: torch.LongTensor = None,
irrelevant_ignored: torch.LongTensor = None) -> None:
if not len(outputs.shape) == len(targets.shape):
targets = torch.unsqueeze(targets, 1)
targets = torch.zeros_like(outputs).scatter_(1, targets, 1)
if relevant_ignored is None:
relevant_ignored = np.zeros(outputs.shape[0])
else:
relevant_ignored = relevant_ignored.cpu().numpy()
if irrelevant_ignored is None:
irrelevant_ignored = np.zeros(outputs.shape[0])
else:
irrelevant_ignored = irrelevant_ignored.cpu().numpy()
if masks is None:
masks = torch.ones_like(targets).int()
else:
# try converting to int
masks = masks.int()
for output, target, mask, rel_ignored, irrel_ignored in zip(
outputs, targets, masks, relevant_ignored, irrelevant_ignored):
valid_docs = torch.sum(mask)
output, target = output[:valid_docs], target[:valid_docs]
output, target = paired_sort(output, target)
output = self._cutoff(output)
confusion = self.confusion_matrix(output, target)
if torch.sum(target).cpu().numpy() + rel_ignored == 0:
if self.version == 'tuning':
# ignore both miss and false alarm when tuning
pass
elif self.version == 'program':
# ignore miss when calculating program target
false_alarm = confusion[0, 1] / float(
sum(confusion[0, :]) + irrel_ignored)
self.false_alarm.append(false_alarm)
else:
miss = (confusion[1, 0] + rel_ignored
) / float(sum(confusion[1, :]) + rel_ignored)
false_alarm = confusion[0, 1] / float(
sum(confusion[0, :]) + irrel_ignored)
self.false_alarm.append(false_alarm)
self.miss.append(miss)
def _get_loss(self, log_probs: torch.FloatTensor,
target_tokens: Dict[str, torch.LongTensor],
source_lengths: torch.LongTensor) -> torch.FloatTensor:
targets = target_tokens[self._target_namespace]
mask = (targets != self._pad_index).bool()
relevant_mask = mask[:, 1:]
relevant_targets = targets[:, 1:]
target_lengths = util.get_lengths_from_binary_sequence_mask(
relevant_mask)
loss = self._loss(log_probs, relevant_targets.int(),
source_lengths.int(), target_lengths.int())
return loss
