titans-pytorch 0.2.16__tar.gz → 0.2.17__tar.gz

{titans_pytorch-0.2.16 → titans_pytorch-0.2.17}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: titans-pytorch
-Version: 0.2.16
+Version: 0.2.17
 Summary: Titans
 Project-URL: Homepage, https://pypi.org/project/titans-pytorch/
 Project-URL: Repository, https://github.com/lucidrains/titans-pytorch

{titans_pytorch-0.2.16 → titans_pytorch-0.2.17}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "titans-pytorch"
-version = "0.2.16"
+version = "0.2.17"
 description = "Titans"
 authors = [
     { name = "Phil Wang", email = "lucidrains@gmail.com" }

{titans_pytorch-0.2.16 → titans_pytorch-0.2.17}/tests/test_titans.py

@@ -93,6 +93,34 @@ def test_neural_mem_chaining_chunks():
 
     assert torch.allclose(parallel_retrieved, torch.cat((first_retrieved, second_retrieved, third_retrieved), dim = 1), atol = 1e-5)
 
+def test_neural_mem_chaining_with_weight_residual():
+    mem  = NeuralMemory(
+        dim = 384,
+        dim_head = 64,
+        heads = 2,
+        chunk_size = 64
+    )
+
+    mem2 = NeuralMemory(
+        dim = 384,
+        dim_head = 64,
+        heads = 2,
+        chunk_size = 64
+    )
+
+    seq = torch.randn(2, 256, 384)
+
+    seq, state = mem(seq)
+
+    parallel_retrieved, _ = mem2(seq, prev_weights = state.updates)
+
+    seq_first, seq_second = seq[:, :128], seq[:, 128:]
+
+    first_retrieved, state1 = mem2(seq_first, prev_weights = state.updates)
+    second_retrieved, state2 = mem2(seq_second, state = state1, prev_weights = state.updates)
+
+    assert torch.allclose(parallel_retrieved, torch.cat((first_retrieved, second_retrieved), dim = 1), atol = 1e-6)
+
 def test_neural_mem_chaining_with_batch_size():
     mem  = NeuralMemory(
         dim = 384,

{titans_pytorch-0.2.16 → titans_pytorch-0.2.17}/titans_pytorch/neural_memory.py

@@ -564,7 +564,12 @@ class NeuralMemory(Module):
         # maybe add previous layer weight
 
         if exists(prev_weights):
-            prev_weights = prev_weights.apply(lambda t: t[:, -1:])
+
+            start_index = math.ceil(seq_index / chunk_size)
+            end_index = start_index + num_chunks
+
+            prev_weights = prev_weights.apply(lambda t: t[:, start_index:end_index])
+
             weights_for_surprise = weights_for_surprise + prev_weights
 
         # flatten batch and time if surprise depends on previous layer memory model

{titans_pytorch-0.2.16 → titans_pytorch-0.2.17}/train_mac.py

@@ -38,7 +38,7 @@ NEURAL_MEM_MAX_LR = 1e-1
 WINDOW_SIZE = 32
 NEURAL_MEM_SEGMENT_LEN = 4                      # set smaller for more granularity for learning rate / momentum etc
 NEURAL_MEM_BATCH_SIZE = 128                     # set smaller to update the neural memory weights more often as it traverses the sequence
-NEURAL_MEM_WEIGHT_RESIDUAL = True
+NEURAL_MEM_WEIGHT_RESIDUAL = False
 SLIDING_WINDOWS = True
 STORE_ATTN_POOL_CHUNKS = True                   # whether to use attention pooling for chunk derived momentum, per-layer lr mod, decay
 MEMORY_MODEL_PER_LAYER_LEARNED_LR = True