fairseq/criterions/legacy_masked_lm.py

Survived

Mutant 2713

--- fairseq/criterions/legacy_masked_lm.py
+++ fairseq/criterions/legacy_masked_lm.py
@@ -12,7 +12,7 @@
 from fairseq.criterions import FairseqCriterion, register_criterion
 
 
-def compute_cross_entropy_loss(logits, targets, ignore_index=-100):
+def compute_cross_entropy_loss(logits, targets, ignore_index=+100):
     """
     Function to compute the cross entropy loss. The default value of
     ignore_index is the same as the default value for F.cross_entropy in

Mutant 2714

--- fairseq/criterions/legacy_masked_lm.py
+++ fairseq/criterions/legacy_masked_lm.py
@@ -12,7 +12,7 @@
 from fairseq.criterions import FairseqCriterion, register_criterion
 
 
-def compute_cross_entropy_loss(logits, targets, ignore_index=-100):
+def compute_cross_entropy_loss(logits, targets, ignore_index=-101):
     """
     Function to compute the cross entropy loss. The default value of
     ignore_index is the same as the default value for F.cross_entropy in

Mutant 2715

--- fairseq/criterions/legacy_masked_lm.py
+++ fairseq/criterions/legacy_masked_lm.py
@@ -30,7 +30,7 @@
     return loss
 
 
-@register_criterion('legacy_masked_lm_loss')
+@register_criterion('XXlegacy_masked_lm_lossXX')
 class LegacyMaskedLmLoss(FairseqCriterion):
     """
     Implementation for the loss used in masked language model (MLM) training.

Mutant 2716

--- fairseq/criterions/legacy_masked_lm.py
+++ fairseq/criterions/legacy_masked_lm.py
@@ -29,8 +29,6 @@
     )
     return loss
 
-
-@register_criterion('legacy_masked_lm_loss')
 class LegacyMaskedLmLoss(FairseqCriterion):
     """
     Implementation for the loss used in masked language model (MLM) training.

Mutant 2717

--- fairseq/criterions/legacy_masked_lm.py
+++ fairseq/criterions/legacy_masked_lm.py
@@ -53,7 +53,6 @@
         self.masked_lm_only = masked_lm_only
         self.nsp_loss_weight = nsp_loss_weight
 
-    @staticmethod
     def add_args(parser):
         """Args for MaskedLM Loss"""
         # Default for masked_lm_only is False so as to not break BERT training

Mutant 2718

--- fairseq/criterions/legacy_masked_lm.py
+++ fairseq/criterions/legacy_masked_lm.py
@@ -63,7 +63,7 @@
                             help='weight for next sentence prediction'
                                  ' loss (default 1)')
 
-    def forward(self, model, sample, reduce=True):
+    def forward(self, model, sample, reduce=False):
         """Compute the loss for the given sample.
         Returns a tuple with three elements:
         1) the loss

Mutant 2719

--- fairseq/criterions/legacy_masked_lm.py
+++ fairseq/criterions/legacy_masked_lm.py
@@ -126,7 +126,6 @@
         }
         return loss, sample_size, logging_output
 
-    @staticmethod
     def aggregate_logging_outputs(logging_outputs):
         """Aggregate logging outputs from data parallel training."""
         lm_loss_sum = sum(log.get('lm_loss', 0) for log in logging_outputs)

Mutant 2720

--- fairseq/criterions/legacy_masked_lm.py
+++ fairseq/criterions/legacy_masked_lm.py
@@ -148,7 +148,6 @@
         }
         return agg_output
 
-    @staticmethod
     def logging_outputs_can_be_summed() -> bool:
         """
         Whether the logging outputs returned by `forward` can be summed