titans-pytorch 0.0.20__tar.gz → 0.0.22__tar.gz

{titans_pytorch-0.0.20 → titans_pytorch-0.0.22}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: titans-pytorch
-Version: 0.0.20
+Version: 0.0.22
 Summary: Titans
 Project-URL: Homepage, https://pypi.org/project/titans-pytorch/
 Project-URL: Repository, https://github.com/lucidrains/titans-pytorch

{titans_pytorch-0.0.20 → titans_pytorch-0.0.22}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "titans-pytorch"
-version = "0.0.20"
+version = "0.0.22"
 description = "Titans"
 authors = [
     { name = "Phil Wang", email = "lucidrains@gmail.com" }

{titans_pytorch-0.0.20 → titans_pytorch-0.0.22}/tests/test_titans.py

@@ -2,8 +2,10 @@ import torch
 import pytest
 
 @pytest.mark.parametrize('seq_len', (32, 1024, 77))
+@pytest.mark.parametrize('max_grad_norm', (None, 2.))
 def test_titans(
-    seq_len
+    seq_len,
+    max_grad_norm
 ):
 
     from titans_pytorch import NeuralMemory
@@ -11,6 +13,7 @@ def test_titans(
     mem = NeuralMemory(
         dim = 384,
         chunk_size = 64,
+        max_grad_norm = max_grad_norm
     )
 
     seq = torch.randn(2, seq_len, 384)

{titans_pytorch-0.0.20 → titans_pytorch-0.0.22}/titans_pytorch/titans.py

@@ -40,6 +40,9 @@ def exists(v):
 def default(v, d):
     return v if exists(v) else d
 
+def identity(t):
+    return t
+
 def round_down_multiple(seq, mult):
     return seq // mult * mult
 
@@ -97,6 +100,9 @@ class MemoryMLP(Module):
 
 # main neural memory
 
+def default_adaptive_step_transform(adaptive_step, max_lr = 1e-1):
+    return adaptive_step.sigmoid() * max_lr
+
 def default_loss_fn(pred, target):
     return (pred - target).pow(2).mean(dim = -1).sum()
 
@@ -109,6 +115,7 @@ class NeuralMemory(Module):
         heads = 1,
         model: Module | None = None,
         store_memory_loss_fn: Callable = default_loss_fn,
+        adaptive_step_transform: Callable = default_adaptive_step_transform,
         pre_rmsnorm = True,
         post_rmsnorm = True,
         max_grad_norm: float | None = None,
@@ -188,6 +195,8 @@ class NeuralMemory(Module):
             Rearrange('b n h -> (b h) n')
         )
 
+        self.adaptive_step_transform = adaptive_step_transform
+
         # allow for softclamp the gradient norms for storing memories
 
         self.max_grad_norm = max_grad_norm
@@ -242,7 +251,8 @@ class NeuralMemory(Module):
 
         # pack batch and sequence dimension
 
-        adaptive_lr = (self.to_adaptive_step(seq).sigmoid() * -15).exp() # from 1. - 1e-7
+        adaptive_lr = self.to_adaptive_step(seq)
+        adaptive_lr = self.adaptive_step_transform(adaptive_lr)
 
         adaptive_momentum = self.to_momentum(seq).sigmoid()
         decay_factor = self.to_decay_factor(seq).sigmoid()

{titans_pytorch-0.0.20 → titans_pytorch-0.0.22}/train.py

@@ -62,11 +62,12 @@ def decode_tokens(tokens):
 
 titans_neural_memory = NeuralMemory(
     dim = 384,
-    chunk_size = WINDOW_SIZE,
+    chunk_size = 4,
     pre_rmsnorm = True,
     post_rmsnorm = True,
-    dim_head = 32,
-    heads = 8,
+    dim_head = 64,
+    heads = 4,
+    max_grad_norm = 1.,
     use_accelerated_scan = True,
     default_mlp_kwargs = dict(
         depth = NEURAL_MEMORY_DEPTH