Basic implementation of WeightedCrossEntropy torchmetric.

models/dimamba.py

@@ -1,6 +1,6 @@
 import math
 from functools import partial
-from typing import Optional, Tuple, Union
+from typing import Optional, Tuple, Union, Mapping
 
 import huggingface_hub
 import numpy as np
@@ -23,6 +23,7 @@
     BaseModelOutputWithNoAttention,
     MaskedLMOutput,
 )
+from composer.metrics.nlp import LanguageCrossEntropy
 
 try:
     from mamba_ssm.ops.triton.layernorm import (
@@ -835,13 +836,47 @@ def forward(
         return hidden_states, all_hidden_states
 
 
+class WeightedCrossEntropy(LanguageCrossEntropy):
+    """A weighted variant of cross entropy loss, which will discount certain tokens."""
+
+    def __init__(self, loss_weights: Tensor, dist_sync_on_step: bool = False, ignore_index: int = -100):
+        super().__init__(dist_sync_on_step, ignore_index)
+        self.name = "weighted_cross_entropy"
+        self.loss_weights = loss_weights
+        # Note: this differs from the inherited implementation, whose `reduction='sum'`.
+        self.loss_fn = torch.nn.CrossEntropyLoss(ignore_index=ignore_index, reduction='none')
+
+    def update(self, output: Union[Mapping, Tensor], target: Tensor) -> None:
+        """An adaptation of `LangaugeCrossEntropy` that takes a weighted loss."""
+        if isinstance(output, Mapping):
+            logits = output['logits']
+        elif isinstance(output, Tensor):
+            logits = output
+        else:
+            raise Exception(f'Type {type(output)} for the output is unsupported.')
+
+        target = target.view(-1)
+        logits = logits.view(target.shape[0], -1)
+        losses = self.loss_fn(logits, target)
+
+        # Note: this is an addition to the original implementation.
+        loss_weights = self.loss_weights.view(-1)
+        loss_weights[target == self.ignore_index] = 0.0
+
+        total_items = (target != self.ignore_index).sum()
+        self.total_items += total_items  # type: ignore (third-party)
+
+        # accumulate **weighted** loss over all batches
+        self.sum_loss += losses * (loss_weights / loss_weights.sum())
+
+# Deprecated in favor of standard metric from Composer.
 def cross_entropy(logits, y, ignore_index=-100):
     """Cross entropy loss."""
     logits = logits.view(-1, logits.shape[-1])
     y = y.view(-1)
     return F.cross_entropy(logits, y, ignore_index=ignore_index)
 
-
+# Deprecated in favor of above implementation
 def weighted_cross_entropy(logits, y, loss_weights, ignore_index=-100):
     """Weighted cross entropy loss (discounts certain tokens)."""
     logits = logits.view(-1, logits.shape[-1])

requirements.yaml

@@ -12,6 +12,7 @@ dependencies:
   - pytorch-cuda=12.1
   - pip:
       - causal-conv1d==1.1.3.post1
+      - mosaicml
       - datasets==2.18.0
       - einops==0.7.0
       - fsspec==2024.2.0