x-transformers 1.28.5__tar.gz → 1.29.0__tar.gz

{x_transformers-1.28.5/x_transformers.egg-info → x_transformers-1.29.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: x-transformers
-Version: 1.28.5
+Version: 1.29.0
 Summary: X-Transformers - Pytorch
 Home-page: https://github.com/lucidrains/x-transformers
 Author: Phil Wang

{x_transformers-1.28.5 → x_transformers-1.29.0}/README.md

@@ -674,6 +674,55 @@ model = TransformerWrapper(
 )
 ```
 
+### Weight-tied Layers
+
+In the early days of the cambrian explosion of BERT, a paper explored weight tying all the layers, the model named <a href="https://arxiv.org/abs/1909.11942">ALBERT</a>. You can use it by setting `weight_tie_layers = True`
+
+```python
+import torch
+from x_transformers import TransformerWrapper, Encoder
+
+model = TransformerWrapper(
+    num_tokens = 20000,
+    max_seq_len = 1024,
+    attn_layers = Encoder(
+        dim = 512,
+        depth = 12,
+        weight_tie_layers = True   # set this to True to weight tie all the layers
+    )
+)
+```
+
+If you wish to do something more sophisticated, say 3 layers, with each layer recurrent 4 times before onto the next, that is possible as well.
+
+```python
+import torch
+from x_transformers import TransformerWrapper, Decoder
+
+model = TransformerWrapper(
+    num_tokens = 20000,
+    max_seq_len = 1024,
+    attn_layers = Decoder(
+        dim = 512,
+        custom_layers = (
+            'a', 'f',        # 3 sets of attention and feedforward
+            'a', 'f',
+            'a', 'f'
+        ),
+        layers_execute_order = (
+            *((0, 1) * 4),   # each done 4 times before sequentially passed forward, but you can probably imagine some more interesting configurations...
+            *((2, 3) * 4),
+            *((4, 5) * 4),
+        )
+    )
+)
+
+x = torch.randint(0, 256, (1, 1024))
+
+model(x) # (1, 1024, 20000)
+
+```
+
 ### Understanding and Improving Transformer From a Multi-Particle Dynamic System Point of View
 
 <img src="./images/macaron-1.png"></img>

{x_transformers-1.28.5 → x_transformers-1.29.0}/setup.py

@@ -3,7 +3,7 @@ from setuptools import setup, find_packages
 setup(
   name = 'x-transformers',
   packages = find_packages(exclude=['examples']),
-  version = '1.28.5',
+  version = '1.29.0',
   license='MIT',
   description = 'X-Transformers - Pytorch',
   author = 'Phil Wang',

{x_transformers-1.28.5 → x_transformers-1.29.0}/x_transformers/x_transformers.py

@@ -1001,7 +1001,7 @@ class AttentionLayers(Module):
     def __init__(
         self,
         dim,
-        depth,
+        depth = None,
         heads = 8,
         causal = False,
         cross_attend = False,
@@ -1054,6 +1054,8 @@ class AttentionLayers(Module):
         attn_kwargs, kwargs = groupby_prefix_and_trim('attn_', kwargs)
         cross_attn_kwargs, kwargs = groupby_prefix_and_trim('cross_attn_', kwargs)
 
+        assert len(kwargs) == 0, f'unrecognized kwargs passed in {kwargs.keys()}'
+
         dim_head = attn_kwargs.get('dim_head', DEFAULT_DIM_HEAD)
 
         self.dim = dim
@@ -1138,9 +1140,12 @@ class AttentionLayers(Module):
 
         # setup weight tying, which is a special case of `layer_execute_order`
 
+        assert not (exists(layers_execute_order) and exists(custom_layers) and exists(depth)), 'depth should not be passed in if using custom layers and custom layer execution order'
+
         assert not (weight_tie_layers and any([*map(exists, (custom_layers, par_ratio, sandwich_coef))]))
 
         if weight_tie_layers:
+            assert exists(depth), 'depth must be passed in with `weight_tie_layers` = True'
             assert not exists(layers_execute_order)
             layers_execute_order = tuple(range(len(default_block))) * depth
             depth = 1
@@ -1164,6 +1169,7 @@ class AttentionLayers(Module):
             assert sandwich_coef > 0 and sandwich_coef <= depth, 'sandwich coefficient should be less than the depth'
             layer_types = ('a',) * sandwich_coef + default_block * (depth - sandwich_coef) + ('f',) * sandwich_coef
         else:
+            assert exists(depth), '`depth` must be passed in for `Decoder` or `Encoder`'
             layer_types = default_block * depth
 
         self.layer_types = layer_types

{x_transformers-1.28.5 → x_transformers-1.29.0/x_transformers.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: x-transformers
-Version: 1.28.5
+Version: 1.29.0
 Summary: X-Transformers - Pytorch
 Home-page: https://github.com/lucidrains/x-transformers
 Author: Phil Wang