Remove torch dependency

src/deepsparse/transformers/metrics.py

@@ -20,8 +20,8 @@
 
 import numpy
 
-import torch
 from sklearn.metrics import precision_recall_fscore_support
+from scipy.special import log_softmax
 
 
 __all__ = [
@@ -71,11 +71,7 @@ def add_batch(self, predictions: numpy.ndarray, targets: numpy.ndarray):
             targets = targets.flatten()
 
             # Compute negative log-likelihood and accumulate
-            self._neg_log_likelihood += torch.nn.functional.cross_entropy(
-                torch.tensor(predictions),
-                torch.tensor(targets),
-                reduction="sum",
-            ).item()
+            self._neg_log_likelihood += _cross_entropy(predictions, targets, reduction="sum").sum()
 
             # Track number of tokens processed
             self._number_tokens += predictions.shape[0]
@@ -90,11 +86,7 @@ def add_batch(self, predictions: numpy.ndarray, targets: numpy.ndarray):
                 targets = numpy.expand_dims(targets, axis=0)
 
             # Compute negative log-likelihoods for batch
-            neg_log_likelihoods = torch.nn.functional.cross_entropy(
-                torch.tensor(predictions.transpose(0, 2, 1)),
-                torch.tensor(targets),
-                reduction="none",
-            ).numpy().mean(-1)
+            neg_log_likelihoods = _cross_entropy(predictions, targets)
 
             # Compute perplexities for batch
             perplexities = numpy.exp(neg_log_likelihoods)
@@ -181,3 +173,15 @@ def compute(self) -> Dict[str, float]:
         results["f1_std"] = f1.std()
 
         return results
+
+
+def _cross_entropy(predictions, targets, reduction="mean"):
+    logp = log_softmax(predictions, axis=-1)
+    neg_log_likelihoods = -1. * numpy.take_along_axis(logp, numpy.expand_dims(targets, axis=-1), axis=-1)
+    neg_log_likelihoods = numpy.squeeze(neg_log_likelihoods, axis=-1)
+    if reduction == "mean":
+        neg_log_likelihoods = neg_log_likelihoods.mean(axis=-1)
+    elif reduction == "sum":
+        neg_log_likelihoods = neg_log_likelihoods.sum(axis=-1)
+
+    return neg_log_likelihoods