titans-pytorch 0.0.3__py3-none-any.whl → 0.0.5__py3-none-any.whl

titans_pytorch/titans.py

@@ -1,4 +1,5 @@
 from __future__ import annotations
+import math
 from functools import partial
 
 import torch
@@ -11,12 +12,13 @@ from tensordict import TensorDict
 
 from titans_pytorch.associative_scan import (
     associative_scan,
-    binary_operator
+    binary_operator,
+    pad_at_dim
 )
 
 import einx
 from einops import rearrange, pack, unpack
-from einops.layers.torch import Rearrange
+from einops.layers.torch import Rearrange, Reduce
 
 """
 ein notation:
@@ -41,6 +43,9 @@ def default(v, d):
 def round_down_multiple(seq, mult):
     return seq // mult * mult
 
+def round_up_multiple(seq, mult):
+    return math.ceil(seq / mult) * mult
+
 def pack_one_with_inverse(t, pattern):
     packed, packed_shape = pack([t], pattern)
 
@@ -84,6 +89,7 @@ class NeuralMemory(Module):
     def __init__(
         self,
         dim,
+        chunk_size = 1,
         model: Module | None = None,
         store_memory_loss_fn: Callable = default_loss_fn
     ):
@@ -98,11 +104,15 @@ class NeuralMemory(Module):
 
         self.memory_model = model
 
+        # the chunk size within the paper where adaptive step, momentum, weight decay are shared
+
+        self.chunk_size = chunk_size
+
         # prepare function for per sample gradients from model above, using torch.func
 
         def forward_and_loss(params, inputs, target):
             pred = functional_call(self.memory_model, params, inputs)
-            loss = self.store_memory_loss_fn(pred, target) # simple mse loss in paper - eq (12) - |M(k) == v|²
+            loss = self.store_memory_loss_fn(pred, target) # simple mse loss in paper - eq (12) - |M(k) - v|²
             return loss
 
         self.per_sample_grad_and_value_fn = vmap(grad_and_value(forward_and_loss), in_dims = (None, 0, 0))
@@ -119,9 +129,23 @@ class NeuralMemory(Module):
         # learned adaptive learning rate and momentum
         # todo - explore mlp layerwise learned lr / momentum
 
-        self.to_momentum = LinearNoBias(dim, 1)
-        self.to_adaptive_step = nn.Sequential(LinearNoBias(dim, 1), Rearrange('... 1 -> ...'))
-        self.to_decay_factor = LinearNoBias(dim, 1) # weight decay factor
+        self.to_momentum = nn.Sequential(
+            Reduce('b (n c) ... -> b n ...', 'mean', c = chunk_size),
+            LinearNoBias(dim, 1)
+        )
+
+        self.to_adaptive_step = nn.Sequential(
+            Reduce('b (n c) ... -> b n ...', 'mean', c = chunk_size),
+            LinearNoBias(dim, 1),
+            Rearrange('... 1 -> ...')
+        )
+
+        # weight decay factor
+
+        self.to_decay_factor = nn.Sequential(
+            Reduce('b (n c) ... -> b n ...', 'mean', c = chunk_size),
+            LinearNoBias(dim, 1)
+        )
 
     def init_weights_and_momentum(self):
         params = TensorDict(dict(self.memory_model.named_parameters()))
@@ -137,6 +161,16 @@ class NeuralMemory(Module):
         past_state: tuple[dict[str, Tensor], dict[str, Tensor]]
     ):
 
+        # curtail sequence by multiple of the chunk size
+        # only a complete chunk of the sequence provides the memory for the next chunk
+
+        seq_len = seq.shape[-2]
+        round_down_seq_len = round_down_multiple(seq_len, self.chunk_size)
+
+        seq = seq[:, :round_down_seq_len]
+
+        # curr weights + past weights, in the case that the initial weights are learned
+
         curr_weights = TensorDict(dict(self.memory_model.named_parameters()))
 
         past_state = tuple(TensorDict(d) for d in past_state)
@@ -148,16 +182,19 @@ class NeuralMemory(Module):
 
         batch = seq.shape[0]
 
-        adaptive_lr = self.to_adaptive_step(seq).tanh() * 0.5 + 0.5.
+        adaptive_lr = self.to_adaptive_step(seq).tanh() * 0.5 + 0.5
 
         adaptive_momentum = self.to_momentum(seq).sigmoid()
         decay_factor = self.to_decay_factor(seq).sigmoid()
 
         # keys and values
 
-        seq = rearrange(seq, 'b n d -> (b n) d')
         keys, values = self.to_keys_values(seq).chunk(2, dim = -1)
 
+        # take care of chunking
+
+        keys, values = tuple(rearrange(t, 'b (n c) d -> (b n) c d', c = self.chunk_size) for t in (keys, values))
+
         # get grads and extra auxiliary loss (for backwarding through qkv projection in base neural memory module)
 
         grads, aux_store_loss = self.per_sample_grad_and_value_fn(dict(curr_weights), keys, values)
@@ -172,31 +209,24 @@ class NeuralMemory(Module):
 
         surprises = grads.apply(lambda t: einx.multiply('b n ..., b n -> b n ...', t, -adaptive_lr))
 
-        # derive momentum with associative scan - eq (10)
+        # momentum + weight decay - momentum is the new contribution, as most linear RNNs have learned forgetting gates
 
         next_momentum = TensorDict()
+        updates = TensorDict()
 
         for param_name, surprise in surprises.items():
             surprise, inverse_pack = pack_one_with_inverse(surprise, 'b n *')
 
-            _, momentum = associative_scan(binary_operator, (adaptive_momentum, surprise)) # momentum is S / surprise in the paper
-
-            momentum = inverse_pack(momentum)
-
-            next_momentum[param_name] = momentum
+            # derive momentum with associative scan - eq (10)
 
-        # use associative scan again for learned forgetting (weight decay) - eq (13)
-
-        updates = TensorDict()
+            _, momentum = associative_scan(binary_operator, (adaptive_momentum, surprise)) # momentum is S / surprise in the paper
 
-        for param_name, momentum in next_momentum.items():
-            momentum, inverse_pack = pack_one_with_inverse(momentum, 'b n *')
+            # use associative scan again for learned forgetting (weight decay) - eq (13)
 
             _, update = associative_scan(binary_operator, (1. - decay_factor, momentum)) # momentum is S / surprise in the paper
 
-            update = inverse_pack(update)
-
-            updates[param_name] = update
+            updates[param_name] = inverse_pack(update)
+            next_momentum[param_name] = inverse_pack(momentum)
 
         # compute the next weight per batch
 
@@ -211,7 +241,19 @@ class NeuralMemory(Module):
         seq,
         past_weights: dict[str, Tensor] | None = None,
     ):
-        batch = seq.shape[0]
+        chunk_size = self.chunk_size
+        batch, seq_len = seq.shape[:2]
+
+        assert seq_len >= chunk_size
+
+        seq = seq[:, (chunk_size - 1):]
+        curtailed_seq_len = seq.shape[-2]
+
+        next_seq_len = round_up_multiple(curtailed_seq_len, chunk_size)
+
+        padding = next_seq_len - curtailed_seq_len
+
+        seq = pad_at_dim(seq, (0, padding), dim = 1)
 
         # the parameters of the memory model stores the memories of the key / values
         # when the MLP has only 1 weight matrix, it is equivalent to `kv` fast weight memories from linear attention literature (recall fetching of memories is q @ (kv)) / schmidhuber's paper
@@ -231,7 +273,7 @@ class NeuralMemory(Module):
         # fetch values from memory model
 
         curr_weights = curr_weights.apply(lambda t: rearrange(t, 'b n ... -> (b n) ...'))
-        queries = rearrange(queries, 'b n d -> (b n) 1 d')
+        queries = rearrange(queries, 'b (n c) d -> (b n) c d', c = chunk_size)
 
         # forward functional call
 
@@ -239,7 +281,12 @@ class NeuralMemory(Module):
 
         # reconstitute batch dimension
 
-        values = rearrange(values, '(b n) 1 d -> b n d', b = batch)
+        values = rearrange(values, '(b n) c d -> b (n c) d', b = batch)
+
+        # restore
+
+        values = pad_at_dim(values, (chunk_size - 1, 0), dim = 1, value = 0.) # todo, used a learned null memory embedding instead of 0s for retrieving from empty neural memory
+        values = values[:, :-padding]
 
         return values
 

{titans_pytorch-0.0.3.dist-info → titans_pytorch-0.0.5.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: titans-pytorch
-Version: 0.0.3
+Version: 0.0.5
 Summary: Titans
 Project-URL: Homepage, https://pypi.org/project/titans-pytorch/
 Project-URL: Repository, https://github.com/lucidrains/titans-pytorch
@@ -49,7 +49,7 @@ Description-Content-Type: text/markdown
 
 ## Titans - Pytorch (wip)
 
-Unofficial implementation of [Titans](https://arxiv.org/abs/2501.00663) in Pytorch. Will also contain some explorations into architectures beyond their simple 1-4 layer MLP for the neural memory module.
+Unofficial implementation of [Titans](https://arxiv.org/abs/2501.00663) in Pytorch. Will also contain some explorations into architectures beyond their simple 1-4 layer MLP for the neural memory module, if it works well to any degree.
 
 ## Install
 
@@ -65,7 +65,10 @@ from titans_pytorch import NeuralMemory
 
 x = torch.randn(2, 64, 32)
 
-mem = NeuralMemory(32)
+mem = NeuralMemory(
+    dim = 32,
+    chunk_size = 2
+)
 
 out = mem(x)
 

titans_pytorch-0.0.5.dist-info/RECORD

@@ -0,0 +1,7 @@
+titans_pytorch/__init__.py,sha256=QKuJPCOJCdgtaPeKoHEkYkiQe65_LV9_8-cIMbBPU30,55
+titans_pytorch/associative_scan.py,sha256=Y-iYqmFuG-NoCKu6kgql1mhowXTeJfyawi3eUIXamp0,2650
+titans_pytorch/titans.py,sha256=3Mewuysj0g7iAlfjdqMlJhn9-pKJuOerB1frQmQYXuc,9428
+titans_pytorch-0.0.5.dist-info/METADATA,sha256=f1DgCKZz9nqNfZOrqbOpyn-yEx2v5M5zgGIW0Zeu84I,3032
+titans_pytorch-0.0.5.dist-info/WHEEL,sha256=qtCwoSJWgHk21S1Kb4ihdzI2rlJ1ZKaIurTj_ngOhyQ,87
+titans_pytorch-0.0.5.dist-info/licenses/LICENSE,sha256=1yCiA9b5nhslTavxPjsQAO-wpOnwJR9-l8LTVi7GJuk,1066
+titans_pytorch-0.0.5.dist-info/RECORD,,

titans_pytorch-0.0.3.dist-info/RECORD

@@ -1,7 +0,0 @@
-titans_pytorch/__init__.py,sha256=QKuJPCOJCdgtaPeKoHEkYkiQe65_LV9_8-cIMbBPU30,55
-titans_pytorch/associative_scan.py,sha256=Y-iYqmFuG-NoCKu6kgql1mhowXTeJfyawi3eUIXamp0,2650
-titans_pytorch/titans.py,sha256=0Mh9LJv5hLVbB2MvRJX5QanAeTtU9LAuj6YOQUwsyUQ,7813
-titans_pytorch-0.0.3.dist-info/METADATA,sha256=AXfDl_MTIu24VRagi_rgiH8rHXFBU5euwSD6DMwLgsg,2968
-titans_pytorch-0.0.3.dist-info/WHEEL,sha256=qtCwoSJWgHk21S1Kb4ihdzI2rlJ1ZKaIurTj_ngOhyQ,87
-titans_pytorch-0.0.3.dist-info/licenses/LICENSE,sha256=1yCiA9b5nhslTavxPjsQAO-wpOnwJR9-l8LTVi7GJuk,1066
-titans_pytorch-0.0.3.dist-info/RECORD,,