titans-pytorch 0.0.30__tar.gz → 0.0.32__tar.gz

{titans_pytorch-0.0.30 → titans_pytorch-0.0.32}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: titans-pytorch
-Version: 0.0.30
+Version: 0.0.32
 Summary: Titans
 Project-URL: Homepage, https://pypi.org/project/titans-pytorch/
 Project-URL: Repository, https://github.com/lucidrains/titans-pytorch

{titans_pytorch-0.0.30 → titans_pytorch-0.0.32}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "titans-pytorch"
-version = "0.0.30"
+version = "0.0.32"
 description = "Titans"
 authors = [
     { name = "Phil Wang", email = "lucidrains@gmail.com" }

{titans_pytorch-0.0.30 → titans_pytorch-0.0.32}/titans_pytorch/mac_transformer.py

@@ -17,6 +17,10 @@ from hyper_connections import get_init_and_expand_reduce_stream_functions
 from axial_positional_embedding import ContinuousAxialPositionalEmbedding
 from rotary_embedding_torch import RotaryEmbedding
 
+# proposed neural memory
+
+from titans_pytorch.titans import NeuralMemory
+
 # constants
 
 LinearNoBias = partial(Linear, bias = False)
@@ -29,9 +33,33 @@ def exists(v):
 def default(v, d):
     return v if exists(v) else d
 
+def identity(t):
+    return t
+
 def round_up_multiple(seq, mult):
     return ceil(seq / mult) * mult
 
+def pad_and_segment_with_inverse(seq, segment_len):
+    batch, seq_len = seq.shape[:2]
+
+    need_segment = seq_len >= segment_len
+
+    if not need_segment:
+        return seq, identity
+
+    next_seq_len_mult = round_up_multiple(seq_len, segment_len)
+
+    padding = next_seq_len_mult - seq_len
+    seq = F.pad(seq, (0, 0, 0, padding))
+
+    seq = rearrange(seq, 'b (w n) d -> (b w) n d', n = segment_len)
+
+    def inverse(out):
+        out = rearrange(out, '(b w) n d -> b (w n) d', b = batch)
+        return out[:, :-padding]
+
+    return seq, inverse
+
 # feedforward and attention
 
 class GEGLU(Module):
@@ -77,9 +105,6 @@ class SegmentedAttention(Module):
         self.split_heads = Rearrange('b n (h d) -> b h n d', h = heads)
         self.merge_heads = Rearrange('b h n d -> b n (h d)')
 
-        self.segment_seq = Rearrange('b (n w) d -> (b n) w d', n = total_segment_len)
-        self.merge_seq_back = Rearrange('(b n) w d -> b (n w) d', n = total_segment_len)
-
         self.persistent_memory = nn.Parameter(torch.zeros(2, heads, num_persist_mem_tokens, dim_head))
 
     def forward(self, seq):
@@ -91,16 +116,7 @@ class SegmentedAttention(Module):
         # auto pad to multiple
         # todo - get rid of logic with flex attention
 
-        need_segment = seq_len >= total_segment_len
-
-        if need_segment:
-            next_seq_len = round_up_multiple(seq_len, total_segment_len)
-            padding = next_seq_len - seq_len
-
-            if padding > 0:
-                seq = F.pad(seq, (0, 0, 0, padding))
-
-            seq = self.segment_seq(seq)
+        seq, inverse_segment = pad_and_segment_with_inverse(seq, total_segment_len)
 
         # attention
 
@@ -130,10 +146,9 @@ class SegmentedAttention(Module):
 
         out = self.to_out(out)
 
-        if need_segment:
-            out = self.merge_seq_back(out)
+        out = inverse_segment(out)
 
-        return out[:, :seq_len]
+        return out
 
 # MAC transformer
 
@@ -150,7 +165,9 @@ class MemoryAsContextTransformer(Module):
         dim_head = 64,
         heads = 8,
         ff_mult = 4,
-        num_residual_streams = 4
+        num_residual_streams = 4,
+        neural_memory_kwargs: dict = dict(),
+        neural_memory_layers: tuple[int, ...] | None = None,
     ):
         super().__init__()
 
@@ -170,8 +187,25 @@ class MemoryAsContextTransformer(Module):
         init_hyper_conn, self.expand_streams, self.reduce_streams = get_init_and_expand_reduce_stream_functions(num_residual_streams, disable = num_residual_streams == 1)
 
         self.layers = ModuleList([])
+        self.neural_mem_layers = ModuleList([])
+
+        layers = tuple(range(1, depth + 1))
+        neural_memory_layers = set(default(neural_memory_layers, layers))
+
+        for layer in layers:
+
+            # neural memory
+
+            mem = None
+
+            if num_longterm_mem_tokens > 0 and layer in neural_memory_layers:
+                mem = NeuralMemory(dim = dim, chunk_size = num_longterm_mem_tokens)
+                mem = init_hyper_conn(dim = dim, branch = mem)
+
+            self.neural_mem_layers.append(mem)
+
+            # attention and feedforward
 
-        for _ in range(depth):
             attn = SegmentedAttention(
                 dim = dim,
                 dim_head = dim_head,
@@ -207,40 +241,54 @@ class MemoryAsContextTransformer(Module):
 
         # intersperse longterm memory
 
-        need_segment = seq_len >= segment_len
-
-        if need_segment:
-            next_seq_len = round_up_multiple(seq_len, segment_len)
-            padding = next_seq_len - seq_len
-
-            if padding > 0:
-                x = F.pad(x, (0, 0, 0, padding))
-
-            x = rearrange(x, 'b (w n) d -> (b w) n d', n = segment_len)
+        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 = torch.cat((mems, x), dim = -2)
+        x = cat((mems, x), dim = -2)
 
-        if need_segment:
-            x = rearrange(x, '(b w) n d -> b (w n) d', b = batch)
-            x = x[:, :seq_len]
+        x = inverse_segment(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, total_segment_len), flatten = True)
-        x = x + pos_emb[:seq_len]
+        x = x + pos_emb[:x.shape[-2]]
 
         # expand and reduce streams for hyper connections
 
         x = self.expand_streams(x)
 
-        for attn, ff in self.layers:
+        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)
+
             x = attn(x)
             x = ff(x)
 
         x = self.reduce_streams(x)
 
+        # excise out the memories
+
+        x, inverse_segment = pad_and_segment_with_inverse(x, total_segment_len)
+
+        x = x[:, num_longterm_mem_tokens:]
+
+        x = inverse_segment(x)
+
         # to logits
 
         x = self.norm(x)