titans-pytorch 0.0.27__py3-none-any.whl → 0.0.29__py3-none-any.whl

titans_pytorch/mac_transformer.py

@@ -1,5 +1,5 @@
 from __future__ import annotations
-import math
+from math import ceil
 from functools import partial
 
 import torch
@@ -10,8 +10,14 @@ from torch.nn import Module, ModuleList, Linear
 from einops import repeat
 from einops.layers.torch import Rearrange
 
+
 from hyper_connections import get_init_and_expand_reduce_stream_functions
 
+# absolute and relative positions
+
+from axial_positional_embedding import ContinuousAxialPositionalEmbedding
+from rotary_embedding_torch import RotaryEmbedding
+
 # constants
 
 LinearNoBias = partial(Linear, bias = False)
@@ -25,7 +31,7 @@ def default(v, d):
     return v if exists(v) else d
 
 def round_up_multiple(seq, mult):
-    return math.ceil(seq / mult) * mult
+    return ceil(seq / mult) * mult
 
 # feedforward and attention
 
@@ -58,6 +64,8 @@ class SegmentedAttention(Module):
 
         dim_inner = dim_head * heads
 
+        self.rotary_emb = RotaryEmbedding(dim_head)
+
         self.to_qkv = LinearNoBias(dim, dim_inner * 3)
         self.to_out = LinearNoBias(dim_inner, dim)
 
@@ -99,6 +107,12 @@ class SegmentedAttention(Module):
 
         pmk, pmv = tuple(repeat(t, 'h n d -> b h n d', b = seq.shape[0]) for t in self.persistent_memory)
 
+        # relative positions
+
+        q, k = self.rotary_emb.rotate_queries_with_cached_keys(q, k)
+
+        # persistent memory
+
         k = cat((pmk, k), dim = -2)
         v = cat((pmv, v), dim = -2)
 
@@ -133,6 +147,9 @@ class MemoryAsContextTransformer(Module):
     ):
         super().__init__()
 
+        self.segment_len = segment_len
+        self.axial_pos_emb = ContinuousAxialPositionalEmbedding(dim = dim, num_axial_dims = 2)
+
         self.token_emb = nn.Embedding(num_tokens, dim)
 
         init_hyper_conn, self.expand_streams, self.reduce_streams = get_init_and_expand_reduce_stream_functions(num_residual_streams, disable = num_residual_streams == 1)
@@ -160,9 +177,19 @@ class MemoryAsContextTransformer(Module):
         self.to_logits = LinearNoBias(dim, num_tokens)
 
     def forward(self, x):
+        seq_len, segment_len = x.shape[-1], self.segment_len
+        windows = ceil(seq_len / segment_len)
 
         x = self.token_emb(x)
 
+        # apply axial positional embedding
+        # so intra and inter segment can be more easily discerned by the network
+
+        pos_emb = self.axial_pos_emb((windows, segment_len), flatten = True)
+        x = x + pos_emb[:seq_len]
+
+        # expand and reduce streams for hyper connections
+
         x = self.expand_streams(x)
 
         for attn, ff in self.layers:

{titans_pytorch-0.0.27.dist-info → titans_pytorch-0.0.29.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: titans-pytorch
-Version: 0.0.27
+Version: 0.0.29
 Summary: Titans
 Project-URL: Homepage, https://pypi.org/project/titans-pytorch/
 Project-URL: Repository, https://github.com/lucidrains/titans-pytorch
@@ -35,10 +35,12 @@ Classifier: Programming Language :: Python :: 3.9
 Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
 Requires-Python: >=3.9
 Requires-Dist: accelerated-scan>=0.2.0
+Requires-Dist: axial-positional-embedding>=0.3.5
 Requires-Dist: einops>=0.8.0
 Requires-Dist: einx>=0.3.0
 Requires-Dist: hyper-connections>=0.1.8
 Requires-Dist: ninja
+Requires-Dist: rotary-embedding-torch
 Requires-Dist: tensordict
 Requires-Dist: torch>=2.2
 Provides-Extra: examples

{titans_pytorch-0.0.27.dist-info → titans_pytorch-0.0.29.dist-info}/RECORD

@@ -1,9 +1,9 @@
 titans_pytorch/__init__.py,sha256=nB0873FZ_OyCda3qFeWTdpO4LbbrXDEvtAefVLzh6g0,71
 titans_pytorch/associative_scan.py,sha256=Y-iYqmFuG-NoCKu6kgql1mhowXTeJfyawi3eUIXamp0,2650
-titans_pytorch/mac_transformer.py,sha256=4xUSfGDdVsR-WmeXX7yXoFfybROvNCjOxL_EHDJ_Wlk,4681
+titans_pytorch/mac_transformer.py,sha256=JBrJah7gfQDPizYRcBvpUKinrd2I9KMB997f3RIR8TA,5568
 titans_pytorch/titans.py,sha256=Kx_tl_QkeDccvkMwPZ0xQ_saYjZfbKzDNPTTSHNWYcA,14304
 titans_pytorch/titans_attn_memory.py,sha256=Rwx_-riGeISBefZg5Kjic8jzmmRRys-u93D2Kgb7Mos,12691
-titans_pytorch-0.0.27.dist-info/METADATA,sha256=qJp1IDbphEUfW7EyNvQ7RfmHuvB7SH5h_tlnCVwV4EY,3851
-titans_pytorch-0.0.27.dist-info/WHEEL,sha256=qtCwoSJWgHk21S1Kb4ihdzI2rlJ1ZKaIurTj_ngOhyQ,87
-titans_pytorch-0.0.27.dist-info/licenses/LICENSE,sha256=1yCiA9b5nhslTavxPjsQAO-wpOnwJR9-l8LTVi7GJuk,1066
-titans_pytorch-0.0.27.dist-info/RECORD,,
+titans_pytorch-0.0.29.dist-info/METADATA,sha256=EhS4E9SAoqzDa0PIjZpQmUSYAo5IS-XePofWlZZnIS0,3938
+titans_pytorch-0.0.29.dist-info/WHEEL,sha256=qtCwoSJWgHk21S1Kb4ihdzI2rlJ1ZKaIurTj_ngOhyQ,87
+titans_pytorch-0.0.29.dist-info/licenses/LICENSE,sha256=1yCiA9b5nhslTavxPjsQAO-wpOnwJR9-l8LTVi7GJuk,1066
+titans_pytorch-0.0.29.dist-info/RECORD,,