x-transformers 2.9.0__py3-none-any.whl → 2.9.2__py3-none-any.whl

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
@@ -171,6 +180,9 @@ class Attend(Module):
         qk_norm = False,
         l2_distance = False,
         sigmoid = False,
+        gumbel_softmax = False,
+        gumbel_softmax_temp = 1.,
+        gumbel_softmax_hard = True,
         custom_attn_fn: Callable | None = None,
         flash = False,
         softclamp_logits = False,
@@ -203,7 +215,7 @@ class Attend(Module):
         assert not (flash and hard), 'hard attention not available for flash'
         assert not (flash and is_sparse_topk_attn), 'topk attention not available for flash'
 
-        assert at_most_one_of(sigmoid, hard, l2_distance, is_sparse_topk_attn)
+        assert at_most_one_of(sigmoid, hard, l2_distance, gumbel_softmax, is_sparse_topk_attn)
 
         if exists(custom_attn_fn):
             self.attn_fn = custom_attn_fn
@@ -213,6 +225,8 @@ class Attend(Module):
             self.attn_fn = one_hot_straight_through
         elif is_sparse_topk_attn:
             self.attn_fn = partial(sparse_topk_attn, sparse_topk = sparse_topk, straight_through = sparse_topk_straight_through)
+        elif gumbel_softmax:
+            self.attn_fn = partial(F.gumbel_softmax, dim = -1, tau = gumbel_softmax_temp, hard = gumbel_softmax_hard)
         else:
             softmax_fn = partial(F.softmax, dim = -1)
             self.attn_fn = partial(softmax_fn, dtype = torch.float32) if not qk_norm else softmax_fn

x_transformers/x_transformers.py

@@ -1336,6 +1336,9 @@ class Attention(Module):
         value_rmsnorm = False,      # used in alphagenome and bytedance's GR3 for further stability
         l2_distance = False,
         sigmoid = False,
+        gumbel_softmax = False,
+        gumbel_softmax_temp = 1.,
+        gumbel_softmax_hard = True,
         selective = False,
         custom_attn_fn: Callable | None = None,
         hybrid_module: Module | None = None,
@@ -1541,6 +1544,9 @@ class Attention(Module):
             scale = qk_norm_scale if qk_norm else self.scale,
             l2_distance = l2_distance,
             sigmoid = sigmoid,
+            gumbel_softmax = gumbel_softmax,
+            gumbel_softmax_temp = gumbel_softmax_temp,
+            gumbel_softmax_hard = gumbel_softmax_hard,
             selective = selective,
             custom_attn_fn = custom_attn_fn,
             add_zero_kv = add_zero_kv,

{x_transformers-2.9.0.dist-info → x_transformers-2.9.2.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: x-transformers
-Version: 2.9.0
+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
@@ -2574,4 +2574,16 @@ ids_out, num_out, is_number_mask = model.generate(start_ids, start_nums, 17)
 }
 ```
 
+```bibtex
+@misc{yan2017hierarchicalmultiscaleattentionnetworks,
+    title   = {Hierarchical Multi-scale Attention Networks for Action Recognition}, 
+    author  = {Shiyang Yan and Jeremy S. Smith and Wenjin Lu and Bailing Zhang},
+    year    = {2017},
+    eprint  = {1708.07590},
+    archivePrefix = {arXiv},
+    primaryClass = {cs.CV},
+    url     = {https://arxiv.org/abs/1708.07590}, 
+}
+```
+
 *solve intelligence... then use that to solve everything else.* - Demis Hassabis

{x_transformers-2.9.0.dist-info → x_transformers-2.9.2.dist-info}/RECORD

@@ -1,5 +1,5 @@
 x_transformers/__init__.py,sha256=aVuhUU0572TJHW88BVc4yA2tla0Zb8l3NH7W4RZ1AEs,1005
-x_transformers/attend.py,sha256=jzOwrtCIdAt1dRQBO68htDsgtjdTx6TAQQVB2xflS1w,18091
+x_transformers/attend.py,sha256=RZJT9pPlpqSG3nOUqQHNRR6jOeJ2r-Fvvar2wdu9HLw,18687
 x_transformers/autoregressive_wrapper.py,sha256=BsGO9xfVYkvynqbU1__tu_S_cxl7gss0YwnkhIa2baY,18401
 x_transformers/belief_state_wrapper.py,sha256=YLUMk6t2MhFBEw5lHDDHJHcoCxTIkHvxTNY__GGZEKU,13374
 x_transformers/continuous.py,sha256=WwpQCjyVY4PtuEAOFY68zqgklbF9I7AL5w6874YlDe8,13249
@@ -10,10 +10,10 @@ x_transformers/multi_input.py,sha256=tCh-fTJDj2ib4SMGtsa-AM8MxKzJAQSwqAXOu3HU2mg
 x_transformers/neo_mlp.py,sha256=XCNnnop9WLarcxap1kGuYc1x8GHvwkZiDRnXOxSl3Po,3452
 x_transformers/nonautoregressive_wrapper.py,sha256=hMQqNimGtchNIe13cR5LZule1V7I1qM5LmY8VQfVdnA,11698
 x_transformers/up_wrapper.py,sha256=YC2LN14_7Xx9Wtiek2rtEJ_qHqdfSmKlh3d7Cgxwd80,7073
-x_transformers/x_transformers.py,sha256=xIaM5LGimNmDThr9WK0ZF5vfm3sSp0OMll0aRIqjymQ,125443
+x_transformers/x_transformers.py,sha256=o6B10urcC7MRUrmoHOgYJgkrVDzHhX-jt6zZY3pZEgA,125700
 x_transformers/xl_autoregressive_wrapper.py,sha256=CvZMJ6A6PA-Y_bQAhnORwjJBSl6Vjq2IdW5KTdk8NI8,4195
 x_transformers/xval.py,sha256=AwwYUm8yDAtKQyKJDIhYMsiLTJ_skh3scUFMjp5sda8,8597
-x_transformers-2.9.0.dist-info/METADATA,sha256=tQ93i_TSfQIf9OViWNJj57ezY0Slred1NblIyf4UIs0,94986
-x_transformers-2.9.0.dist-info/WHEEL,sha256=qtCwoSJWgHk21S1Kb4ihdzI2rlJ1ZKaIurTj_ngOhyQ,87
-x_transformers-2.9.0.dist-info/licenses/LICENSE,sha256=As9u198X-U-vph5noInuUfqsAG2zX_oXPHDmdjwlPPY,1066
-x_transformers-2.9.0.dist-info/RECORD,,
+x_transformers-2.9.2.dist-info/METADATA,sha256=3JsbSIp9fsGpuXopeIaIq4ffjYTJIHyqdRLxM21cfUM,95381
+x_transformers-2.9.2.dist-info/WHEEL,sha256=qtCwoSJWgHk21S1Kb4ihdzI2rlJ1ZKaIurTj_ngOhyQ,87
+x_transformers-2.9.2.dist-info/licenses/LICENSE,sha256=As9u198X-U-vph5noInuUfqsAG2zX_oXPHDmdjwlPPY,1066
+x_transformers-2.9.2.dist-info/RECORD,,