titans-pytorch 0.0.1__py3-none-any.whl

titans_pytorch/__init__.py

@@ -0,0 +1,3 @@
+from titans_pytorch.titans import (
+    NeuralMemory
+)

titans_pytorch/associative_scan.py

@@ -0,0 +1,90 @@
+from __future__ import annotations
+from typing import Callable
+
+import torch
+from torch import Tensor
+import torch.nn.functional as F
+
+# taken from S5-pytorch repository
+# https://github.com/i404788/s5-pytorch/blob/74e2fdae00b915a62c914bf3615c0b8a4279eb84/s5/jax_compat.py#L51-L134
+
+# helper functions
+
+def pad_at_dim(t, pad, dim = -1, value = 0.):
+    dims_from_right = (- dim - 1) if dim < 0 else (t.ndim - dim - 1)
+    zeros = ((0, 0) * dims_from_right)
+    return F.pad(t, (*zeros, *pad), value = value)
+
+# the operator that is needed
+
+@torch.jit.script
+def binary_operator(
+    a: tuple[Tensor, Tensor],
+    b: tuple[Tensor, Tensor]
+):
+    a_i, kv_i = a
+    a_j, kv_j = b
+    return a_j * a_i, torch.addcmul(kv_j, a_j, kv_i)
+
+# Pytorch impl. of jax.lax.associative_scan
+# made specifically for axis of 1 (sequence of tokens for autoregressive modeling)
+
+def associative_scan(
+    operator: Callable,
+    elems: tuple[Tensor, Tensor]
+):
+    num_elems = int(elems[0].shape[1])
+
+    if not all(int(elem.shape[1]) == num_elems for elem in elems[1:]):
+        raise ValueError('Array inputs to associative_scan must have the same '
+                         'first dimension. (saw: {})'
+                         .format([elem.shape for elem in elems]))
+
+    def _scan(elems):
+        """Perform scan on `elems`."""
+        num_elems = elems[0].shape[1]
+
+        if num_elems < 2:
+            return elems
+
+        # Combine adjacent pairs of elements.
+
+        reduced_elems = operator(
+          [elem[:, :-1:2] for elem in elems],
+          [elem[:, 1::2] for elem in elems])
+
+        # Recursively compute scan for partially reduced tensors.
+
+        odd_elems = _scan(reduced_elems)
+
+        if num_elems % 2 == 0:
+            even_elems = operator(
+                [e[:, :-1] for e in odd_elems],
+                [e[:, 2::2] for e in elems])
+        else:
+            even_elems = operator(
+                odd_elems,
+                [e[:, 2::2] for e in elems])
+
+        # The first element of a scan is the same as the first element
+        # of the original `elems`.
+
+        even_elems = [
+          torch.cat([elem[:, :1], result], dim=1)
+          for (elem, result) in zip(elems, even_elems)]
+
+        return list(map(_interleave, even_elems, odd_elems))
+
+    return _scan(elems)
+
+def _interleave(a, b):
+    a_axis_len, b_axis_len = a.shape[1], b.shape[1]
+    output_axis_len = a_axis_len + b_axis_len
+
+    if (a_axis_len == (b_axis_len + 1)):
+        b = pad_at_dim(b, (0, 1), dim = 1)
+
+    stacked = torch.stack([a, b], dim=2)
+    interleaved = torch.flatten(stacked, start_dim=1, end_dim=2)
+
+    return interleaved[:, :output_axis_len]

titans_pytorch/titans.py

@@ -0,0 +1,269 @@
+from __future__ import annotations
+from functools import partial
+
+import torch
+from torch import nn, Tensor
+import torch.nn.functional as F
+from torch.nn import Linear, Module
+from torch.func import functional_call, vmap, grad_and_value
+
+from tensordict import TensorDict
+
+from titans_pytorch.associative_scan import (
+    associative_scan,
+    binary_operator
+)
+
+import einx
+from einops import rearrange, pack, unpack
+from einops.layers.torch import Rearrange
+
+"""
+ein notation:
+b - batch
+n - sequence
+d - feature dimension
+c - intra-chunk
+"""
+
+# constants
+
+LinearNoBias = partial(Linear, bias = False)
+
+# functions
+
+def exists(v):
+    return v is not None
+
+def default(v, d):
+    return v if exists(v) else d
+
+def round_down_multiple(seq, mult):
+    return seq // mult * mult
+
+def pack_one_with_inverse(t, pattern):
+    packed, packed_shape = pack([t], pattern)
+
+    def inverse(out, inv_pattern = None):
+        inv_pattern = default(inv_pattern, pattern)
+        return unpack(out, packed_shape, inv_pattern)[0]
+
+    return packed, inverse
+
+# classes
+
+class MLP(Module):
+    def __init__(
+        self,
+        dim,
+        depth
+    ):
+        super().__init__()
+        self.weights = nn.ParameterList([nn.Parameter(torch.randn(dim, dim)) for _ in range(depth)])
+
+    def forward(
+        self,
+        x
+    ):
+        for ind, weight in enumerate(self.weights):
+            is_first = ind == 0
+
+            if not is_first:
+                x = F.silu(x)
+
+            x = x @ weight
+
+        return x
+
+# main neural memory
+
+def default_loss_fn(pred, target):
+    return (pred - target).pow(2).mean(dim = -1).sum()
+
+class NeuralMemory(Module):
+    def __init__(
+        self,
+        dim,
+        model: Module | None = None,
+        store_memory_loss_fn: Callable = default_loss_fn
+    ):
+        super().__init__()
+
+        if not exists(model):
+            model = MLP(dim, depth = 4)
+
+        assert not exists(next(model.buffers(), None)), 'model cannot have buffers for now'
+
+        # the memory is the weights of the model
+
+        self.memory_model = model
+
+        # 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
+            return loss
+
+        self.per_sample_grad_and_value_fn = vmap(grad_and_value(forward_and_loss), in_dims = (None, 0, 0))
+
+        # queries for retrieving from the model
+
+        self.to_queries = LinearNoBias(dim, dim)
+
+        # keys and values for storing to the model
+
+        self.to_keys_values = LinearNoBias(dim, dim * 2)
+        self.store_memory_loss_fn = store_memory_loss_fn
+
+        # 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 = nn.Sequential(LinearNoBias(dim, 1), nn.Sigmoid()) # weight decay factor
+
+    def init_weights_and_momentum(self):
+        params = TensorDict(dict(self.memory_model.named_parameters()))
+
+        init_weights = params.clone().zero_()
+        init_momentum = params.clone().zero_()
+
+        return init_weights, init_momentum
+
+    def store_memories(
+        self,
+        seq,
+        past_state: tuple[dict[str, Tensor], dict[str, Tensor]]
+    ):
+
+        curr_weights = TensorDict(dict(self.memory_model.named_parameters()))
+
+        past_state = tuple(TensorDict(d) for d in past_state)
+        past_weights, past_momentum = past_state
+
+        curr_weights = curr_weights + past_weights
+
+        # pack batch and sequence dimension
+
+        batch = seq.shape[0]
+
+        adaptive_lr = self.to_adaptive_step(seq)
+        adaptive_momentum = self.to_momentum(seq)
+
+        decay_factor = self.to_decay_factor(seq)
+
+        # keys and values
+
+        seq = rearrange(seq, 'b n d -> (b n) d')
+        keys, values = self.to_keys_values(seq).chunk(2, dim = -1)
+
+        # 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)
+
+        grads = TensorDict(grads)
+
+        # restore batch and sequence dimension
+
+        grads = grads.apply(lambda t: rearrange(t, '(b n) ... -> b n ...', b = batch))
+
+        # multiply gradients with learned adaptive step size
+
+        surprises = grads.apply(lambda t: einx.multiply('b n ..., b n -> b n ...', t, -adaptive_lr))
+
+        # derive momentum with associative scan - eq (10)
+
+        next_momentum = 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
+
+        # use associative scan again for learned forgetting (weight decay) - eq (13)
+
+        updates = TensorDict()
+
+        for param_name, momentum in next_momentum.items():
+            momentum, inverse_pack = pack_one_with_inverse(momentum, 'b n *')
+
+            _, update = associative_scan(binary_operator, (1. - decay_factor, momentum)) # momentum is S / surprise in the paper
+
+            update = inverse_pack(update)
+
+            updates[param_name] = update
+
+        # compute the next weight per batch
+
+        last_update = updates.apply(lambda t: t[:, -1])
+
+        next_state = (curr_weights + last_update, next_momentum)
+
+        return updates, next_state, aux_store_loss.mean()
+
+    def retrieve_memories(
+        self,
+        seq,
+        past_weights: dict[str, Tensor] | None = None,
+    ):
+        batch = seq.shape[0]
+
+        # 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
+
+        curr_weights = TensorDict(dict(self.memory_model.named_parameters()))
+
+        if exists(past_weights):
+            past_weights = TensorDict(past_weights)
+            assert past_weights.keys() == curr_weights.keys()
+
+            curr_weights = curr_weights + past_weights
+
+        # sequence Float['b n d'] to queries
+
+        queries = self.to_queries(seq)
+
+        # 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')
+
+        # forward functional call
+
+        values = functional_call(self.memory_model, dict(curr_weights), queries)
+
+        # reconstitute batch dimension
+
+        values = rearrange(values, '(b n) 1 d -> b n d', b = batch)
+
+        return values
+
+    def forward(
+        self,
+        seq,
+        past_state: tuple[dict[str, Tensor], dict[str, Tensor]] | None = None,
+        return_next_memories = False
+    ):
+        batch = seq.shape[0]
+
+        if exists(past_state):
+            past_state = tuple(TensorDict(d) for d in past_state)
+
+        if not exists(past_state):
+            past_state = self.init_weights_and_momentum()
+
+        updates, next_memories, aux_kv_mse_loss = self.store_memories(seq, past_state)
+
+        past_weights, _ = past_state
+
+        retrieved = self.retrieve_memories(seq, past_weights + updates)
+
+        if not return_next_memories:
+            return retrieved
+
+        return retrieved, next_memories, aux_kv_mse_loss

titans_pytorch-0.0.1.dist-info/METADATA

@@ -0,0 +1,84 @@
+Metadata-Version: 2.4
+Name: titans-pytorch
+Version: 0.0.1
+Summary: Titans
+Project-URL: Homepage, https://pypi.org/project/titans-pytorch/
+Project-URL: Repository, https://github.com/lucidrains/titans-pytorch
+Author-email: Phil Wang <lucidrains@gmail.com>
+License: MIT License
+        
+        Copyright (c) 2025 Phil Wang
+        
+        Permission is hereby granted, free of charge, to any person obtaining a copy
+        of this software and associated documentation files (the "Software"), to deal
+        in the Software without restriction, including without limitation the rights
+        to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+        copies of the Software, and to permit persons to whom the Software is
+        furnished to do so, subject to the following conditions:
+        
+        The above copyright notice and this permission notice shall be included in all
+        copies or substantial portions of the Software.
+        
+        THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+        IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+        FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+        AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+        LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+        OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+        SOFTWARE.
+License-File: LICENSE
+Keywords: artificial intelligence,deep learning,linear attention,neural memory module,test time training
+Classifier: Development Status :: 4 - Beta
+Classifier: Intended Audience :: Developers
+Classifier: License :: OSI Approved :: MIT License
+Classifier: Programming Language :: Python :: 3.9
+Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
+Requires-Python: >=3.9
+Requires-Dist: einops>=0.8.0
+Requires-Dist: einx>=0.3.0
+Requires-Dist: tensordict>=0.6.2
+Requires-Dist: torch>=2.3
+Provides-Extra: examples
+Provides-Extra: test
+Requires-Dist: pytest; extra == 'test'
+Description-Content-Type: text/markdown
+
+<img src="./fig2.png" width="400px"></img>
+
+<img src="./fig1.png" width="400px"></img>
+
+## 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.
+
+## Install
+
+```bash
+$ pip install titans-pytorch
+```
+
+## Usage
+
+```python
+import torch
+from titans_pytorch import NeuralMemory
+
+x = torch.randn(2, 64, 32)
+
+mem = NeuralMemory(32)
+
+out = mem(x)
+
+assert x.shape == out.shape
+```
+
+## Citations
+
+```bibtex
+@inproceedings{Behrouz2024TitansLT,
+    title   = {Titans: Learning to Memorize at Test Time},
+    author  = {Ali Behrouz and Peilin Zhong and Vahab S. Mirrokni},
+    year    = {2024},
+    url     = {https://api.semanticscholar.org/CorpusID:275212078}
+}
+```

titans_pytorch-0.0.1.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=xty74Q3xQ174uycscfOnh-zgxGMH882lrIA_KGvxTUU,7802
+titans_pytorch-0.0.1.dist-info/METADATA,sha256=HqR3VxpV5e-dPLLEbuOekC161-2r2WwKBCvK7E2MhAs,2968
+titans_pytorch-0.0.1.dist-info/WHEEL,sha256=qtCwoSJWgHk21S1Kb4ihdzI2rlJ1ZKaIurTj_ngOhyQ,87
+titans_pytorch-0.0.1.dist-info/licenses/LICENSE,sha256=1yCiA9b5nhslTavxPjsQAO-wpOnwJR9-l8LTVi7GJuk,1066
+titans_pytorch-0.0.1.dist-info/RECORD,,

titans_pytorch-0.0.1.dist-info/WHEEL

@@ -0,0 +1,4 @@
+Wheel-Version: 1.0
+Generator: hatchling 1.27.0
+Root-Is-Purelib: true
+Tag: py3-none-any

titans_pytorch-0.0.1.dist-info/licenses/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2025 Phil Wang
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.