able to return per token logit entropy

setup.py

@@ -3,7 +3,7 @@
 setup(
   name = 'x-transformers',
   packages = find_packages(exclude=['examples']),
-  version = '1.43.4',
+  version = '1.43.5',
   license='MIT',
   description = 'X-Transformers - Pytorch',
   author = 'Phil Wang',

x_transformers/x_transformers.py

@@ -38,6 +38,7 @@ class LayerIntermediates:
     attn_z_loss:        Tensor | None = None
     mems:               Tensor | None = None
     memory_tokens:      Tensor | None = None
+    logit_entropies:    Tensor | None = None
 
 LinearNoBias = partial(nn.Linear, bias = False)
 
@@ -136,6 +137,15 @@ def or_reduce(masks):
         head = head | rest
     return head
 
+# entropy
+
+def calc_entropy(
+    t: Tensor,
+    is_prob = False
+):
+    prob = t.softmax(dim = -1) if not is_prob else t
+    return -(prob * log(prob)).sum(dim = -1)
+
 # auxiliary loss helpers
 
 def calc_z_loss(
@@ -2592,6 +2602,7 @@ def forward(
         return_embeddings = False,
         return_logits_and_embeddings = False,
         return_intermediates = False,
+        return_logit_entropies = False,
         mask = None,
         return_mems = False,
         return_attn = False,
@@ -2809,6 +2820,12 @@ def forward(
         else:
             out = logits
 
+        # logit entropies
+
+        if return_logit_entropies:
+            intermediates.logit_entropies = calc_entropy(logits)
+            return_intermediates = True
+
         # aux loss
 
         if return_attn_z_loss: