x-transformers 1.40.3__tar.gz → 1.40.5__tar.gz

{x_transformers-1.40.3/x_transformers.egg-info → x_transformers-1.40.5}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: x-transformers
-Version: 1.40.3
+Version: 1.40.5
 Summary: X-Transformers - Pytorch
 Home-page: https://github.com/lucidrains/x-transformers
 Author: Phil Wang

{x_transformers-1.40.3 → x_transformers-1.40.5}/README.md

@@ -2287,4 +2287,27 @@ ids_out, num_out, is_number_mask = model.generate(start_ids, start_nums, 17)
 }
 ```
 
+```bibtex
+@article{Bai2019DeepEM,
+    title   = {Deep Equilibrium Models},
+    author  = {Shaojie Bai and J. Zico Kolter and Vladlen Koltun},
+    journal = {ArXiv},
+    year    = {2019},
+    volume  = {abs/1909.01377},
+    url     = {https://api.semanticscholar.org/CorpusID:202539738}
+}
+```
+
+```bibtex
+@article{Wu2021MuseMorphoseFA,
+    title   = {MuseMorphose: Full-Song and Fine-Grained Piano Music Style Transfer With One Transformer VAE},
+    author  = {Shih-Lun Wu and Yi-Hsuan Yang},
+    journal = {IEEE/ACM Transactions on Audio, Speech, and Language Processing},
+    year    = {2021},
+    volume  = {31},
+    pages   = {1953-1967},
+    url     = {https://api.semanticscholar.org/CorpusID:234338162}
+}
+```
+
 *solve intelligence... then use that to solve everything else.* - Demis Hassabis

{x_transformers-1.40.3 → x_transformers-1.40.5}/setup.py

@@ -3,7 +3,7 @@ from setuptools import setup, find_packages
 setup(
   name = 'x-transformers',
   packages = find_packages(exclude=['examples']),
-  version = '1.40.3',
+  version = '1.40.5',
   license='MIT',
   description = 'X-Transformers - Pytorch',
   author = 'Phil Wang',

{x_transformers-1.40.3 → x_transformers-1.40.5}/tests/test_x_transformers.py

@@ -298,3 +298,21 @@ def test_l2_distance(attn_one_kv_head):
     x = torch.randint(0, 256, (1, 1024))
 
     model(x)
+
+def test_reinject_input():
+
+    model = TransformerWrapper(
+        num_tokens = 20000,
+        max_seq_len = 1024,
+        recycling = True,
+        attn_layers = Decoder(
+            dim = 512,
+            depth = 12,
+            heads = 8,
+            reinject_input = True
+        )
+    )
+
+    x = torch.randint(0, 256, (1, 12))
+
+    model(x) # (1, 1024, 20000)

{x_transformers-1.40.3 → x_transformers-1.40.5}/x_transformers/x_transformers.py

@@ -1353,6 +1353,7 @@ class AttentionLayers(Module):
         use_layerscale = False,
         layerscale_init_value = 0.,
         unet_skips = False,
+        reinject_input = False, # seen first in DEQ paper https://arxiv.org/abs/1909.01377, but later used in a number of papers trying to achieve depthwise generalization https://arxiv.org/abs/2410.03020v1
         **kwargs
     ):
         super().__init__()
@@ -1582,6 +1583,11 @@ class AttentionLayers(Module):
 
         self.skip_combines = ModuleList([])
 
+        # whether there is reinjection of input at every layer
+
+        self.reinject_input = reinject_input
+        self.reinject_input_proj = nn.Linear(dim, dim, bias = False) if reinject_input else None
+
         # iterate and construct layers
 
         for ind, (layer_type, layer_shift_tokens) in enumerate(zip(self.layer_types, shift_tokens)):
@@ -1667,6 +1673,7 @@ class AttentionLayers(Module):
         rotary_pos_emb = None,
         attn_bias = None,
         condition = None,
+        in_attn_cond = None, # https://arxiv.org/abs/2105.04090
         layers_execute_order: tuple[int, ...] | None = None
     ):
         assert not (self.cross_attend ^ exists(context)), 'context must be passed in if cross_attend is set to True'
@@ -1766,6 +1773,16 @@ class AttentionLayers(Module):
 
         layer_variables = tuple(tuple(layer_variable[i] for i in layers_execute_order) for layer_variable in layer_variables)
 
+        # derived input for reinjection if needed
+
+        if self.reinject_input:
+            assert not exists(in_attn_cond)
+            inp_inject = self.reinject_input_proj(x)
+
+        elif exists(in_attn_cond):
+            # handle in-attention conditioning, which serves the same purpose of having the network learn the residual
+            inp_inject = in_attn_cond if in_attn_cond.ndim == 3 else rearrange(in_attn_cond, 'b d -> b 1 d')
+
         # store all hiddens for skips
 
         skip_hiddens = []
@@ -1810,6 +1827,9 @@ class AttentionLayers(Module):
                 post_branch_norm = maybe(partial)(post_branch_norm, **norm_kwargs)
                 post_main_norm = maybe(partial)(post_main_norm, **norm_kwargs)
 
+            if self.reinject_input:
+                x = x + inp_inject
+
             if exists(pre_norm):
                 x = pre_norm(x)
 

{x_transformers-1.40.3 → x_transformers-1.40.5/x_transformers.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: x-transformers
-Version: 1.40.3
+Version: 1.40.5
 Summary: X-Transformers - Pytorch
 Home-page: https://github.com/lucidrains/x-transformers
 Author: Phil Wang