titans-pytorch 0.0.31__py3-none-any.whl → 0.0.34__py3-none-any.whl

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)
@@ -46,13 +50,20 @@ def pad_and_segment_with_inverse(seq, segment_len):
     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))
+    needs_pad = padding > 0
+
+    if needs_pad:
+        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]
+
+        if needs_pad:
+            out = out[:, :-padding]
+
+        return out
 
     return seq, inverse
 
@@ -161,7 +172,9 @@ 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__()
 
@@ -181,8 +194,25 @@ 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,
@@ -203,7 +233,14 @@ class MemoryAsContextTransformer(Module):
 
         self.to_logits = LinearNoBias(dim, num_tokens)
 
-    def forward(self, x):
+    def forward(
+        self,
+        x,
+        return_loss = False
+    ):
+
+        if return_loss:
+            x, labels = x[:, :-1], x[:, 1:]
 
         # math
 
@@ -221,7 +258,7 @@ 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 = torch.cat((mems, x), dim = -2)
+        x = cat((mems, x), dim = -2)
 
         x = inverse_segment(x)
 
@@ -235,8 +272,27 @@ class MemoryAsContextTransformer(Module):
 
         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)
@@ -245,7 +301,7 @@ class MemoryAsContextTransformer(Module):
 
         x, inverse_segment = pad_and_segment_with_inverse(x, total_segment_len)
 
-        x = x[:, self.num_longterm_mem_tokens:]
+        x = x[:, num_longterm_mem_tokens:]
 
         x = inverse_segment(x)
 
@@ -253,4 +309,9 @@ class MemoryAsContextTransformer(Module):
 
         x = self.norm(x)
 
-        return self.to_logits(x)
+        logits = self.to_logits(x)
+
+        if not return_loss:
+            return logits
+
+        return F.cross_entropy(rearrange(logits, 'b n l -> b l n'), labels)

{titans_pytorch-0.0.31.dist-info → titans_pytorch-0.0.34.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: titans-pytorch
-Version: 0.0.31
+Version: 0.0.34
 Summary: Titans
 Project-URL: Homepage, https://pypi.org/project/titans-pytorch/
 Project-URL: Repository, https://github.com/lucidrains/titans-pytorch

{titans_pytorch-0.0.31.dist-info → titans_pytorch-0.0.34.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=bq5RbCgA0GWLFHTrDTIKUSQhkkuCkdjEykOwjfHDs0M,6747
+titans_pytorch/mac_transformer.py,sha256=FGShQHD-dQQdQKKzvNS_jTC_FcikdqO_s3ZKOKfr_9E,8502
 titans_pytorch/titans.py,sha256=Kx_tl_QkeDccvkMwPZ0xQ_saYjZfbKzDNPTTSHNWYcA,14304
 titans_pytorch/titans_attn_memory.py,sha256=Rwx_-riGeISBefZg5Kjic8jzmmRRys-u93D2Kgb7Mos,12691
-titans_pytorch-0.0.31.dist-info/METADATA,sha256=bN1fVL2S_vML1oqLIA92tvBhkVvnpQN11fU4e1QVI4s,3938
-titans_pytorch-0.0.31.dist-info/WHEEL,sha256=qtCwoSJWgHk21S1Kb4ihdzI2rlJ1ZKaIurTj_ngOhyQ,87
-titans_pytorch-0.0.31.dist-info/licenses/LICENSE,sha256=1yCiA9b5nhslTavxPjsQAO-wpOnwJR9-l8LTVi7GJuk,1066
-titans_pytorch-0.0.31.dist-info/RECORD,,
+titans_pytorch-0.0.34.dist-info/METADATA,sha256=CNqv_jMqk7yj15IpDn2O3jBdVe4wtrSVkht7mk0wW_E,3938
+titans_pytorch-0.0.34.dist-info/WHEEL,sha256=qtCwoSJWgHk21S1Kb4ihdzI2rlJ1ZKaIurTj_ngOhyQ,87
+titans_pytorch-0.0.34.dist-info/licenses/LICENSE,sha256=1yCiA9b5nhslTavxPjsQAO-wpOnwJR9-l8LTVi7GJuk,1066
+titans_pytorch-0.0.34.dist-info/RECORD,,