titans-pytorch 0.0.1__tar.gz → 0.0.8__tar.gz

{titans_pytorch-0.0.1 → titans_pytorch-0.0.8}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: titans-pytorch
-Version: 0.0.1
+Version: 0.0.8
 Summary: Titans
 Project-URL: Homepage, https://pypi.org/project/titans-pytorch/
 Project-URL: Repository, https://github.com/lucidrains/titans-pytorch
@@ -39,6 +39,8 @@ Requires-Dist: einx>=0.3.0
 Requires-Dist: tensordict>=0.6.2
 Requires-Dist: torch>=2.3
 Provides-Extra: examples
+Requires-Dist: local-attention>=1.10.0; extra == 'examples'
+Requires-Dist: taylor-series-linear-attention; extra == 'examples'
 Provides-Extra: test
 Requires-Dist: pytest; extra == 'test'
 Description-Content-Type: text/markdown
@@ -49,7 +51,7 @@ 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
 
@@ -63,13 +65,16 @@ $ pip install titans-pytorch
 import torch
 from titans_pytorch import NeuralMemory
 
-x = torch.randn(2, 64, 32)
+mem = NeuralMemory(
+    dim = 384,
+    chunk_size = 64,
+    pre_rmsnorm = True
+).cuda()
 
-mem = NeuralMemory(32)
+seq = torch.randn(2, 1024, 384).cuda()
+retrieved = mem(seq)
 
-out = mem(x)
-
-assert x.shape == out.shape
+assert seq.shape == retrieved.shape
 ```
 
 ## Citations

{titans_pytorch-0.0.1 → titans_pytorch-0.0.8}/README.md

@@ -4,7 +4,7 @@
 
 ## 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
 
@@ -18,13 +18,16 @@ $ pip install titans-pytorch
 import torch
 from titans_pytorch import NeuralMemory
 
-x = torch.randn(2, 64, 32)
+mem = NeuralMemory(
+    dim = 384,
+    chunk_size = 64,
+    pre_rmsnorm = True
+).cuda()
 
-mem = NeuralMemory(32)
+seq = torch.randn(2, 1024, 384).cuda()
+retrieved = mem(seq)
 
-out = mem(x)
-
-assert x.shape == out.shape
+assert seq.shape == retrieved.shape
 ```
 
 ## Citations

titans_pytorch-0.0.8/data/README.md

@@ -0,0 +1,3 @@
+# Data source
+
+The enwik8 data was downloaded from the Hutter prize page: http://prize.hutter1.net/

{titans_pytorch-0.0.1 → titans_pytorch-0.0.8}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "titans-pytorch"
-version = "0.0.1"
+version = "0.0.8"
 description = "Titans"
 authors = [
     { name = "Phil Wang", email = "lucidrains@gmail.com" }
@@ -36,7 +36,10 @@ Homepage = "https://pypi.org/project/titans-pytorch/"
 Repository = "https://github.com/lucidrains/titans-pytorch"
 
 [project.optional-dependencies]
-examples = []
+examples = [
+    "local-attention>=1.10.0",
+    "taylor-series-linear-attention"
+]
 test = [
     "pytest"
 ]

{titans_pytorch-0.0.1 → titans_pytorch-0.0.8}/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)
 
@@ -50,6 +55,10 @@ def pack_one_with_inverse(t, pattern):
 
     return packed, inverse
 
+def softclamp_max(t, max_value):
+    range_value = max_value / 2
+    return ((t / range_value).tanh() * range_value) + range_value
+
 # classes
 
 class MLP(Module):
@@ -84,11 +93,17 @@ class NeuralMemory(Module):
     def __init__(
         self,
         dim,
+        chunk_size = 1,
         model: Module | None = None,
-        store_memory_loss_fn: Callable = default_loss_fn
+        store_memory_loss_fn: Callable = default_loss_fn,
+        pre_rmsnorm = False,
+        max_adaptive_step_size = 1e-5
     ):
         super().__init__()
 
+        self.retrieve_norm = nn.RMSNorm(dim) if pre_rmsnorm else nn.Identity()
+        self.store_norm = nn.RMSNorm(dim) if pre_rmsnorm else nn.Identity()
+
         if not exists(model):
             model = MLP(dim, depth = 4)
 
@@ -98,11 +113,15 @@ 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 +138,25 @@ 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 = nn.Sequential(LinearNoBias(dim, 1), nn.Sigmoid()) # 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 -> ...')
+        )
+
+        self.max_adaptive_step_size = max_adaptive_step_size
+
+        # 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 +172,18 @@ class NeuralMemory(Module):
         past_state: tuple[dict[str, Tensor], dict[str, Tensor]]
     ):
 
+        seq = self.store_norm(seq)
+
+        # curtail sequence by multiple of the chunk size
+        # only a complete chunk of the sequence provides the memory for the next chunk
+
+        seq_len, chunk_size = seq.shape[-2], self.chunk_size
+        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 +195,19 @@ class NeuralMemory(Module):
 
         batch = seq.shape[0]
 
-        adaptive_lr = self.to_adaptive_step(seq)
-        adaptive_momentum = self.to_momentum(seq)
+        adaptive_lr = softclamp_max(self.to_adaptive_step(seq), self.max_adaptive_step_size)
 
-        decay_factor = self.to_decay_factor(seq)
+        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 +222,24 @@ 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
-
-        # use associative scan again for learned forgetting (weight decay) - eq (13)
+            # derive momentum with associative scan - eq (10)
 
-        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
 
@@ -204,14 +247,28 @@ class NeuralMemory(Module):
 
         next_state = (curr_weights + last_update, next_momentum)
 
-        return updates, next_state, aux_store_loss.mean()
+        return updates, next_state, aux_store_loss.mean() / chunk_size
 
     def retrieve_memories(
         self,
         seq,
         past_weights: dict[str, Tensor] | None = None,
     ):
-        batch = seq.shape[0]
+        chunk_size = self.chunk_size
+        batch, seq_len = seq.shape[:2]
+
+        seq = self.retrieve_norm(seq)
+
+        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 +288,7 @@ 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 +296,12 @@ 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-0.0.8/train.py

@@ -0,0 +1,132 @@
+import random
+import tqdm
+import gzip
+import numpy as np
+
+import torch
+from torch import nn
+from torch.optim import Adam
+from torch.nn import functional as F
+from torch.utils.data import DataLoader, Dataset
+
+from local_attention import LocalTransformer
+
+from taylor_series_linear_attention import TaylorSeriesLinearAttn
+
+from titans_pytorch.titans import NeuralMemory
+
+# constants
+
+NUM_BATCHES = int(1e5)
+BATCH_SIZE = 4
+GRADIENT_ACCUMULATE_EVERY = 4
+LEARNING_RATE = 2e-4
+VALIDATE_EVERY  = 100
+GENERATE_EVERY  = 500
+GENERATE_LENGTH = 512
+SHOULD_GENERATE = False
+SEQ_LEN = 512
+
+# helpers
+
+def cycle(loader):
+    while True:
+        for data in loader:
+            yield data
+
+def decode_token(token):
+    return str(chr(max(32, token)))
+
+def decode_tokens(tokens):
+    return ''.join(list(map(decode_token, tokens)))
+
+# instantiate GPT-like decoder model
+
+titans_neural_memory = NeuralMemory(
+    dim = 384,
+    chunk_size = 64,
+    pre_rmsnorm = True
+)
+
+titans_neural_memory = nn.Sequential(
+    titans_neural_memory,
+    nn.RMSNorm(384)
+)
+
+linear_attn = TaylorSeriesLinearAttn(
+    dim = 384,
+    dim_head = 16,
+    heads = 16,
+    causal = True
+)
+
+model = LocalTransformer(
+    num_tokens = 256,
+    dim = 384,
+    depth = 8,
+    causal = True,
+    local_attn_window_size = 64,
+    max_seq_len = SEQ_LEN,
+    global_attn_layer = titans_neural_memory,
+    layers_insert_global_attn = (4,)
+).cuda()
+
+# prepare enwik8 data
+
+with gzip.open('./data/enwik8.gz') as file:
+    data = np.frombuffer(file.read(int(95e6)), dtype = np.uint8).copy()
+    data_train, data_val = np.split(data, [int(90e6)])
+    data_train, data_val = map(torch.from_numpy, (data_train, data_val))
+
+class TextSamplerDataset(Dataset):
+    def __init__(self, data, seq_len):
+        super().__init__()
+        self.data = data
+        self.seq_len = seq_len
+
+    def __getitem__(self, index):
+        rand_start = torch.randint(0, self.data.size(0) - self.seq_len, (1,))
+        full_seq = self.data[rand_start: rand_start + self.seq_len + 1].long()
+        return full_seq.cuda()
+
+    def __len__(self):
+        return self.data.size(0) // self.seq_len
+
+train_dataset = TextSamplerDataset(data_train, SEQ_LEN)
+val_dataset   = TextSamplerDataset(data_val, SEQ_LEN)
+train_loader  = cycle(DataLoader(train_dataset, batch_size = BATCH_SIZE))
+val_loader    = cycle(DataLoader(val_dataset, batch_size = BATCH_SIZE))
+
+# optimizer
+
+optim = Adam(model.parameters(), lr=LEARNING_RATE)
+
+# training
+
+for i in tqdm.tqdm(range(NUM_BATCHES), mininterval=10., desc='training'):
+    model.train()
+
+    for __ in range(GRADIENT_ACCUMULATE_EVERY):
+        loss = model(next(train_loader), return_loss = True)
+        loss.backward()
+
+    print(f'training loss: {loss.item()}')
+    torch.nn.utils.clip_grad_norm_(model.parameters(), 0.5)
+    optim.step()
+    optim.zero_grad()
+
+    if i % VALIDATE_EVERY == 0:
+        model.eval()
+        with torch.no_grad():
+            loss = model(next(val_loader), return_loss = True)
+            print(f'validation loss: {loss.item()}')
+
+    if SHOULD_GENERATE and i % GENERATE_EVERY == 0:
+        model.eval()
+        inp = random.choice(val_dataset)[:-1]
+        prime = decode_tokens(inp)
+        print(f'%s \n\n %s', (prime, '*' * 100))
+
+        sample = model.generate(inp[None, ...], GENERATE_LENGTH, use_kv_cache = False)
+        output_str = decode_tokens(sample[0])
+        print(output_str)