titans-pytorch 0.2.4__tar.gz → 0.2.5__tar.gz

{titans_pytorch-0.2.4 → titans_pytorch-0.2.5}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: titans-pytorch
-Version: 0.2.4
+Version: 0.2.5
 Summary: Titans
 Project-URL: Homepage, https://pypi.org/project/titans-pytorch/
 Project-URL: Repository, https://github.com/lucidrains/titans-pytorch

{titans_pytorch-0.2.4 → titans_pytorch-0.2.5}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "titans-pytorch"
-version = "0.2.4"
+version = "0.2.5"
 description = "Titans"
 authors = [
     { name = "Phil Wang", email = "lucidrains@gmail.com" }

{titans_pytorch-0.2.4 → titans_pytorch-0.2.5}/titans_pytorch/mac_transformer.py

@@ -488,7 +488,7 @@ class MemoryAsContextTransformer(Module):
         neural_memory_model: Module | None = None,
         neural_memory_kwargs: dict = dict(),
         neural_memory_layers: tuple[int, ...] | None = None,
-        aux_kv_recon_loss_weight = 0.,
+        aux_kv_recon_loss_weight = 1.,
         use_flex_attn = False,
         sliding_window_attn = False,
         weight_tie_memory_model = False,

{titans_pytorch-0.2.4 → titans_pytorch-0.2.5}/titans_pytorch/memory_models.py

@@ -1,5 +1,5 @@
 import torch
-from torch import nn
+from torch import nn, cat
 import torch.nn.functional as F
 from torch.nn import Module, ModuleList, Parameter, ParameterList
 

{titans_pytorch-0.2.4 → titans_pytorch-0.2.5}/titans_pytorch/neural_memory.py

@@ -304,13 +304,26 @@ class NeuralMemory(Module):
             nn.Sigmoid()
         ) if heads > 1 else None
 
-        # memory mlp
+        # memory model
 
         if not exists(model):
             model = MemoryMLP(dim_head, **default_model_kwargs)
 
+        # validate memory model
+
         assert not exists(next(model.buffers(), None)), 'model cannot have buffers for now'
 
+        test_shape = (3, 2, dim_head)
+
+        with torch.no_grad():
+            try:
+                test_input = torch.randn(test_shape)
+                mem_model_output = model(test_input)
+            except:
+                raise RuntimeError(f'memory model unable to accept a tensor of shape {test_shape}')
+
+            assert mem_model_output.shape == test_shape, 'output of memory model needs to be same shape as input'
+
         # the memory is the weights of the model
 
         self.memory_model = model

{titans_pytorch-0.2.4 → titans_pytorch-0.2.5}/train_mac.py

@@ -30,17 +30,18 @@ SEQ_LEN = 512
 NEURAL_MEMORY_DEPTH = 2
 NUM_PERSIST_MEM = 4
 NUM_LONGTERM_MEM = 4
-NEURAL_MEM_LAYERS = (2, 4, 6)               # layers 2, 4, 6 have neural memory, can add more
+NEURAL_MEM_LAYERS = (2, 4, 6)                   # layers 2, 4, 6 have neural memory, can add more
 NEURAL_MEM_GATE_ATTN_OUTPUT = False
 NEURAL_MEM_MOMENTUM = True
-NEURAL_MEM_QK_NORM = False
+NEURAL_MEM_QK_NORM = True
 WINDOW_SIZE = 32
-NEURAL_MEM_SEGMENT_LEN = WINDOW_SIZE // 2 # set smaller for more granularity for learning rate / momentum etc
+NEURAL_MEM_SEGMENT_LEN = WINDOW_SIZE // 2       # set smaller for more granularity for learning rate / momentum etc
 SLIDING_WINDOWS = True
-WEIGHT_TIE_MEMORY_MODEL = True  # set to have memory MLP shared across layers
-STORE_ATTN_POOL_CHUNKS = True # whether to use attention pooling for chunk derived momentum, per-layer lr mod, decay
+WEIGHT_TIE_MEMORY_MODEL = False                 # set to have memory MLP shared across layers
+PREV_MEM_UPDATE_FOR_WEIGHTS = 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
-KV_RECON_LOSS_WEIGHT = 0.
+KV_RECON_LOSS_WEIGHT = 1.
 
 # experiment related
 
@@ -90,6 +91,7 @@ model = MemoryAsContextTransformer(
     use_flex_attn = USE_FLEX_ATTN,
     sliding_window_attn = SLIDING_WINDOWS,
     weight_tie_memory_model = WEIGHT_TIE_MEMORY_MODEL,
+    prev_neural_mem_update_for_weights = PREV_MEM_UPDATE_FOR_WEIGHTS,
     neural_memory_model = MemoryMLP(
         dim = 64,
         depth = NEURAL_MEMORY_DEPTH