titans-pytorch 0.0.30__py3-none-any.whl → 0.0.32__py3-none-any.whl

Sign up to get free protection for your applications and to get access to all the features.

Potentially problematic release.


This version of titans-pytorch might be problematic. Click here for more details.

@@ -17,6 +17,10 @@ from hyper_connections import get_init_and_expand_reduce_stream_functions
17
17
  from axial_positional_embedding import ContinuousAxialPositionalEmbedding
18
18
  from rotary_embedding_torch import RotaryEmbedding
19
19
 
20
+ # proposed neural memory
21
+
22
+ from titans_pytorch.titans import NeuralMemory
23
+
20
24
  # constants
21
25
 
22
26
  LinearNoBias = partial(Linear, bias = False)
@@ -29,9 +33,33 @@ def exists(v):
29
33
  def default(v, d):
30
34
  return v if exists(v) else d
31
35
 
36
+ def identity(t):
37
+ return t
38
+
32
39
  def round_up_multiple(seq, mult):
33
40
  return ceil(seq / mult) * mult
34
41
 
42
+ def pad_and_segment_with_inverse(seq, segment_len):
43
+ batch, seq_len = seq.shape[:2]
44
+
45
+ need_segment = seq_len >= segment_len
46
+
47
+ if not need_segment:
48
+ return seq, identity
49
+
50
+ next_seq_len_mult = round_up_multiple(seq_len, segment_len)
51
+
52
+ padding = next_seq_len_mult - seq_len
53
+ seq = F.pad(seq, (0, 0, 0, padding))
54
+
55
+ seq = rearrange(seq, 'b (w n) d -> (b w) n d', n = segment_len)
56
+
57
+ def inverse(out):
58
+ out = rearrange(out, '(b w) n d -> b (w n) d', b = batch)
59
+ return out[:, :-padding]
60
+
61
+ return seq, inverse
62
+
35
63
  # feedforward and attention
36
64
 
37
65
  class GEGLU(Module):
@@ -77,9 +105,6 @@ class SegmentedAttention(Module):
77
105
  self.split_heads = Rearrange('b n (h d) -> b h n d', h = heads)
78
106
  self.merge_heads = Rearrange('b h n d -> b n (h d)')
79
107
 
80
- self.segment_seq = Rearrange('b (n w) d -> (b n) w d', n = total_segment_len)
81
- self.merge_seq_back = Rearrange('(b n) w d -> b (n w) d', n = total_segment_len)
82
-
83
108
  self.persistent_memory = nn.Parameter(torch.zeros(2, heads, num_persist_mem_tokens, dim_head))
84
109
 
85
110
  def forward(self, seq):
@@ -91,16 +116,7 @@ class SegmentedAttention(Module):
91
116
  # auto pad to multiple
92
117
  # todo - get rid of logic with flex attention
93
118
 
94
- need_segment = seq_len >= total_segment_len
95
-
96
- if need_segment:
97
- next_seq_len = round_up_multiple(seq_len, total_segment_len)
98
- padding = next_seq_len - seq_len
99
-
100
- if padding > 0:
101
- seq = F.pad(seq, (0, 0, 0, padding))
102
-
103
- seq = self.segment_seq(seq)
119
+ seq, inverse_segment = pad_and_segment_with_inverse(seq, total_segment_len)
104
120
 
105
121
  # attention
106
122
 
@@ -130,10 +146,9 @@ class SegmentedAttention(Module):
130
146
 
131
147
  out = self.to_out(out)
132
148
 
133
- if need_segment:
134
- out = self.merge_seq_back(out)
149
+ out = inverse_segment(out)
135
150
 
136
- return out[:, :seq_len]
151
+ return out
137
152
 
138
153
  # MAC transformer
139
154
 
@@ -150,7 +165,9 @@ class MemoryAsContextTransformer(Module):
150
165
  dim_head = 64,
151
166
  heads = 8,
152
167
  ff_mult = 4,
153
- num_residual_streams = 4
168
+ num_residual_streams = 4,
169
+ neural_memory_kwargs: dict = dict(),
170
+ neural_memory_layers: tuple[int, ...] | None = None,
154
171
  ):
155
172
  super().__init__()
156
173
 
@@ -170,8 +187,25 @@ class MemoryAsContextTransformer(Module):
170
187
  init_hyper_conn, self.expand_streams, self.reduce_streams = get_init_and_expand_reduce_stream_functions(num_residual_streams, disable = num_residual_streams == 1)
171
188
 
172
189
  self.layers = ModuleList([])
190
+ self.neural_mem_layers = ModuleList([])
191
+
192
+ layers = tuple(range(1, depth + 1))
193
+ neural_memory_layers = set(default(neural_memory_layers, layers))
194
+
195
+ for layer in layers:
196
+
197
+ # neural memory
198
+
199
+ mem = None
200
+
201
+ if num_longterm_mem_tokens > 0 and layer in neural_memory_layers:
202
+ mem = NeuralMemory(dim = dim, chunk_size = num_longterm_mem_tokens)
203
+ mem = init_hyper_conn(dim = dim, branch = mem)
204
+
205
+ self.neural_mem_layers.append(mem)
206
+
207
+ # attention and feedforward
173
208
 
174
- for _ in range(depth):
175
209
  attn = SegmentedAttention(
176
210
  dim = dim,
177
211
  dim_head = dim_head,
@@ -207,40 +241,54 @@ class MemoryAsContextTransformer(Module):
207
241
 
208
242
  # intersperse longterm memory
209
243
 
210
- need_segment = seq_len >= segment_len
211
-
212
- if need_segment:
213
- next_seq_len = round_up_multiple(seq_len, segment_len)
214
- padding = next_seq_len - seq_len
215
-
216
- if padding > 0:
217
- x = F.pad(x, (0, 0, 0, padding))
218
-
219
- x = rearrange(x, 'b (w n) d -> (b w) n d', n = segment_len)
244
+ x, inverse_segment = pad_and_segment_with_inverse(x, segment_len)
220
245
 
221
246
  mems = repeat(self.longterm_mems, 'n d -> b n d', b = x.shape[0])
222
- x = torch.cat((mems, x), dim = -2)
247
+ x = cat((mems, x), dim = -2)
223
248
 
224
- if need_segment:
225
- x = rearrange(x, '(b w) n d -> b (w n) d', b = batch)
226
- x = x[:, :seq_len]
249
+ x = inverse_segment(x)
227
250
 
228
251
  # apply axial positional embedding
229
252
  # so intra and inter segment can be more easily discerned by the network
230
253
 
231
254
  pos_emb = self.axial_pos_emb((windows, total_segment_len), flatten = True)
232
- x = x + pos_emb[:seq_len]
255
+ x = x + pos_emb[:x.shape[-2]]
233
256
 
234
257
  # expand and reduce streams for hyper connections
235
258
 
236
259
  x = self.expand_streams(x)
237
260
 
238
- for attn, ff in self.layers:
261
+ for (attn, ff), maybe_neural_mem in zip(self.layers, self.neural_mem_layers):
262
+
263
+ if exists(maybe_neural_mem):
264
+ batch_streams = x.shape[0]
265
+ x, inverse_segment = pad_and_segment_with_inverse(x, total_segment_len)
266
+
267
+ longterm_mems, x = x[:, :num_longterm_mem_tokens], x[:, num_longterm_mem_tokens:]
268
+
269
+ longterm_mems = rearrange(longterm_mems, '(b w) n d -> b (w n) d', b = batch_streams)
270
+
271
+ longterm_mems = maybe_neural_mem(longterm_mems)
272
+
273
+ longterm_mems = rearrange(longterm_mems, 'b (w n) d -> (b w) n d', n = num_longterm_mem_tokens)
274
+
275
+ x = cat((longterm_mems, x), dim = -2)
276
+
277
+ x = inverse_segment(x)
278
+
239
279
  x = attn(x)
240
280
  x = ff(x)
241
281
 
242
282
  x = self.reduce_streams(x)
243
283
 
284
+ # excise out the memories
285
+
286
+ x, inverse_segment = pad_and_segment_with_inverse(x, total_segment_len)
287
+
288
+ x = x[:, num_longterm_mem_tokens:]
289
+
290
+ x = inverse_segment(x)
291
+
244
292
  # to logits
245
293
 
246
294
  x = self.norm(x)
@@ -1,6 +1,6 @@
1
1
  Metadata-Version: 2.4
2
2
  Name: titans-pytorch
3
- Version: 0.0.30
3
+ Version: 0.0.32
4
4
  Summary: Titans
5
5
  Project-URL: Homepage, https://pypi.org/project/titans-pytorch/
6
6
  Project-URL: Repository, https://github.com/lucidrains/titans-pytorch
@@ -1,9 +1,9 @@
1
1
  titans_pytorch/__init__.py,sha256=nB0873FZ_OyCda3qFeWTdpO4LbbrXDEvtAefVLzh6g0,71
2
2
  titans_pytorch/associative_scan.py,sha256=Y-iYqmFuG-NoCKu6kgql1mhowXTeJfyawi3eUIXamp0,2650
3
- titans_pytorch/mac_transformer.py,sha256=XDLc9NWXpVMza03XjU0lkw5lRvtJ25ReTPKNoGslOOk,6773
3
+ titans_pytorch/mac_transformer.py,sha256=ohZWVhMBtpm0Iz3w5g7pD3WXSXrvhwzZvfRplwhe1Qo,8149
4
4
  titans_pytorch/titans.py,sha256=Kx_tl_QkeDccvkMwPZ0xQ_saYjZfbKzDNPTTSHNWYcA,14304
5
5
  titans_pytorch/titans_attn_memory.py,sha256=Rwx_-riGeISBefZg5Kjic8jzmmRRys-u93D2Kgb7Mos,12691
6
- titans_pytorch-0.0.30.dist-info/METADATA,sha256=7H6WPsgfBE9ByEUP7r6C-cfBX6K13yNCaVhZMzwUvf8,3938
7
- titans_pytorch-0.0.30.dist-info/WHEEL,sha256=qtCwoSJWgHk21S1Kb4ihdzI2rlJ1ZKaIurTj_ngOhyQ,87
8
- titans_pytorch-0.0.30.dist-info/licenses/LICENSE,sha256=1yCiA9b5nhslTavxPjsQAO-wpOnwJR9-l8LTVi7GJuk,1066
9
- titans_pytorch-0.0.30.dist-info/RECORD,,
6
+ titans_pytorch-0.0.32.dist-info/METADATA,sha256=9X9nWfgIVS-9XIeLHQY53HXSMA6rMemPfyVC2bRrJOQ,3938
7
+ titans_pytorch-0.0.32.dist-info/WHEEL,sha256=qtCwoSJWgHk21S1Kb4ihdzI2rlJ1ZKaIurTj_ngOhyQ,87
8
+ titans_pytorch-0.0.32.dist-info/licenses/LICENSE,sha256=1yCiA9b5nhslTavxPjsQAO-wpOnwJR9-l8LTVi7GJuk,1066
9
+ titans_pytorch-0.0.32.dist-info/RECORD,,