def __init__(self, num_classes, class_weights=None):
LossNM.__init__(self)
if class_weights:
class_weights = torch.FloatTensor(class_weights).to(self._device)
self._criterion = nn.CrossEntropyLoss(weight=class_weights)
self.num_classes = num_classes
def __init__(
self,
num_slots,
slot_classes_loss_weights=None,
intent_classes_loss_weights=None,
intent_loss_weight=0.6,
):
LossNM.__init__(self)
self.num_slots = num_slots
self.intent_loss_weight = intent_loss_weight
self.slot_classes_loss_weights = slot_classes_loss_weights
self.intent_classes_loss_weights = intent_classes_loss_weights
# For weighted loss to tackle class imbalance
if slot_classes_loss_weights:
self.slot_classes_loss_weights = torch.FloatTensor(
slot_classes_loss_weights).to(self._device)
if intent_classes_loss_weights:
self.intent_classes_loss_weights = torch.FloatTensor(
intent_classes_loss_weights).to(self._device)
self._criterion_intent = nn.CrossEntropyLoss(
weight=self.intent_classes_loss_weights)
self._criterion_slot = nn.CrossEntropyLoss(
weight=self.slot_classes_loss_weights)
def __init__(self, pad_id=None, label_smoothing=0, predict_last_k=0):
LossNM.__init__(self)
self._loss_fn = SmoothedCrossEntropy(label_smoothing, predict_last_k)
self._pad_id = pad_id
def __init__(self, num_inputs=2):
# Store number of inputs/losses.
self.num_losses = num_inputs
LossNM.__init__(self)
def __init__(self):
LossNM.__init__(self)
def __init__(self, label_smoothing=0.0):
LossNM.__init__(self)
self._criterion = SmoothedCrossEntropyLoss(label_smoothing)