titans-pytorch 0.0.35__tar.gz → 0.0.37__tar.gz

{titans_pytorch-0.0.35 → titans_pytorch-0.0.37}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: titans-pytorch
-Version: 0.0.35
+Version: 0.0.37
 Summary: Titans
 Project-URL: Homepage, https://pypi.org/project/titans-pytorch/
 Project-URL: Repository, https://github.com/lucidrains/titans-pytorch

{titans_pytorch-0.0.35 → titans_pytorch-0.0.37}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "titans-pytorch"
-version = "0.0.35"
+version = "0.0.37"
 description = "Titans"
 authors = [
     { name = "Phil Wang", email = "lucidrains@gmail.com" }

{titans_pytorch-0.0.35 → titans_pytorch-0.0.37}/titans_pytorch/mac_transformer.py

@@ -7,7 +7,7 @@ from torch import nn, cat
 import torch.nn.functional as F
 from torch.nn import Module, ModuleList, Linear
 
-from einops import repeat, rearrange
+from einops import repeat, rearrange, pack, unpack
 from einops.layers.torch import Rearrange
 
 from hyper_connections import get_init_and_expand_reduce_stream_functions
@@ -185,7 +185,9 @@ class MemoryAsContextTransformer(Module):
         # long term mem tokens
 
         self.segment_len = segment_len
+
         self.num_longterm_mem_tokens = num_longterm_mem_tokens
+        has_longterm_mems = num_longterm_mem_tokens > 0
 
         self.longterm_mems = nn.Parameter(torch.randn(num_longterm_mem_tokens, dim) * 0.02)
 
@@ -197,7 +199,11 @@ class MemoryAsContextTransformer(Module):
         self.neural_mem_layers = ModuleList([])
 
         layers = tuple(range(1, depth + 1))
-        neural_memory_layers = set(default(neural_memory_layers, layers))
+
+        if not exists(neural_memory_layers):
+            neural_memory_layers = layers if has_longterm_mems else ()
+
+        assert not (num_longterm_mem_tokens > 0 and len(neural_memory_layers) == 0), 'empty `neural_memory_layers` when longterm memory tokens are present'
 
         for layer in layers:
 
@@ -205,8 +211,15 @@ class MemoryAsContextTransformer(Module):
 
             mem = None
 
-            if num_longterm_mem_tokens > 0 and layer in neural_memory_layers:
-                mem = NeuralMemory(dim = dim, chunk_size = num_longterm_mem_tokens)
+            if layer in neural_memory_layers:
+                assert has_longterm_mems, '`num_longterm_mem_tokens` must be greater than 0'
+
+                mem = NeuralMemory(
+                    dim = dim,
+                    chunk_size = num_longterm_mem_tokens + segment_len,
+                    **neural_memory_kwargs
+                )
+
                 mem = init_hyper_conn(dim = dim, branch = mem)
 
             self.neural_mem_layers.append(mem)
@@ -258,7 +271,7 @@ class MemoryAsContextTransformer(Module):
         x, inverse_segment = pad_and_segment_with_inverse(x, segment_len)
 
         mems = repeat(self.longterm_mems, 'n d -> b n d', b = x.shape[0])
-        x = cat((mems, x), dim = -2)
+        x, mem_ps = pack((x, mems), 'b * d')
 
         x = inverse_segment(x)
 
@@ -275,21 +288,7 @@ class MemoryAsContextTransformer(Module):
         for (attn, ff), maybe_neural_mem in zip(self.layers, self.neural_mem_layers):
 
             if exists(maybe_neural_mem):
-                batch_streams = x.shape[0]
-
-                x, inverse_segment = pad_and_segment_with_inverse(x, total_segment_len)
-
-                longterm_mems, x = x[:, :num_longterm_mem_tokens], x[:, num_longterm_mem_tokens:]
-
-                longterm_mems = rearrange(longterm_mems, '(b w) n d -> b (w n) d', b = batch_streams)
-
-                longterm_mems = maybe_neural_mem(longterm_mems)
-
-                longterm_mems = rearrange(longterm_mems, 'b (w n) d -> (b w) n d', n = num_longterm_mem_tokens)
-
-                x = cat((longterm_mems, x), dim = -2)
-
-                x = inverse_segment(x)
+                mems = maybe_neural_mem(mems)
 
             x = attn(x)
 
@@ -301,7 +300,7 @@ class MemoryAsContextTransformer(Module):
 
         x, inverse_segment = pad_and_segment_with_inverse(x, total_segment_len)
 
-        x = x[:, num_longterm_mem_tokens:]
+        x, mem = unpack(x, mem_ps, 'b * d')
 
         x = inverse_segment(x)
 

{titans_pytorch-0.0.35 → titans_pytorch-0.0.37}/titans_pytorch/titans.py

@@ -132,7 +132,7 @@ class NeuralMemory(Module):
         max_grad_norm: float | None = None,
         use_accelerated_scan = False,
         default_mlp_kwargs: dict = dict(
-            depth = 4
+            depth = 2
         )
     ):
         super().__init__()

{titans_pytorch-0.0.35 → titans_pytorch-0.0.37}/train_mac.py

@@ -24,11 +24,13 @@ SHOULD_GENERATE = False
 SEQ_LEN = 512
 
 PROJECT_NAME = 'titans-mac-transformer'
-WANDB_ONLINE = False # turn this on to pipe experiment to cloud
-GLOBAL_LAYERS = (2, 4)
+WANDB_ONLINE = True # turn this on to pipe experiment to cloud
 NEURAL_MEMORY_DEPTH = 2
-WINDOW_SIZE = 64
-RUN_NAME = 'mac'
+NUM_PERSIST_MEM = 4
+NUM_LONGTERM_MEM = 4
+NEURAL_MEM_LAYERS = (4,)
+WINDOW_SIZE = 32
+RUN_NAME = 'mac - 4 longterm mems, layers (4,)'
 
 # wandb experiment tracker
 
@@ -57,12 +59,14 @@ model = MemoryAsContextTransformer(
     dim = 384,
     depth = 8,
     segment_len = WINDOW_SIZE,
-    num_persist_mem_tokens = 16,
-    num_longterm_mem_tokens = 16,
-    neural_memory_layers = (3, 4),
+    num_persist_mem_tokens = NUM_PERSIST_MEM,
+    num_longterm_mem_tokens = NUM_LONGTERM_MEM,
+    neural_memory_layers = NEURAL_MEM_LAYERS,
     neural_memory_kwargs = dict(
+        dim_head = 64,
+        heads = 4,
         default_mlp_kwargs = dict(
-            depth = NEURAL_MEMORY_DEPTH
+            depth = NEURAL_MEMORY_DEPTH,
         )
     )
 ).cuda()