x-transformers 2.9.1__tar.gz → 2.9.2__tar.gz

{x_transformers-2.9.1 → x_transformers-2.9.2}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: x-transformers
-Version: 2.9.1
+Version: 2.9.2
 Summary: X-Transformers
 Project-URL: Homepage, https://pypi.org/project/x-transformers/
 Project-URL: Repository, https://github.com/lucidrains/x-transformers

{x_transformers-2.9.1 → x_transformers-2.9.2}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "x-transformers"
-version = "2.9.1"
+version = "2.9.2"
 description = "X-Transformers"
 authors = [
     { name = "Phil Wang", email = "lucidrains@gmail.com" }

{x_transformers-2.9.1 → x_transformers-2.9.2}/tests/test_x_transformers.py

@@ -1378,6 +1378,10 @@ def test_stochastic_attn():
     from x_transformers import Attention
 
     attn = Attention(dim = 512, gumbel_softmax = True)
-    out = attn(torch.randn(1, 1024, 512))
+    out, intermediate = attn(torch.randn(1, 1024, 512), return_intermediates = True)
 
     assert out.shape == (1, 1024, 512)
+
+    from x_transformers.attend import log_prob_from_hard_attend
+    log_probs = log_prob_from_hard_attend(intermediate)
+    assert log_probs.shape == (1, 8, 1024)

{x_transformers-2.9.1 → x_transformers-2.9.2}/x_transformers/attend.py

@@ -67,6 +67,15 @@ def once(fn):
 
 print_once = once(print)
 
+# gumbel softmax attention related
+
+def log_prob_from_hard_attend(intermeds: Intermediates):
+    log_probs = intermeds.pre_softmax_attn.log_softmax(dim = -1)
+
+    one_hot = intermeds.post_softmax_attn.argmax(dim = -1, keepdim = True)
+    log_prob = log_probs.gather(-1, one_hot)
+    return rearrange(log_prob, 'b h i 1 -> b h i')
+
 # selective attention
 # https://arxiv.org/abs/2410.02703 - section 3.3
 # it is a technique to allow each token to prevent itself from being attended to by future tokens