titans-pytorch 0.1.8__tar.gz → 0.1.9__tar.gz

{titans_pytorch-0.1.8 → titans_pytorch-0.1.9}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: titans-pytorch
-Version: 0.1.8
+Version: 0.1.9
 Summary: Titans
 Project-URL: Homepage, https://pypi.org/project/titans-pytorch/
 Project-URL: Repository, https://github.com/lucidrains/titans-pytorch

{titans_pytorch-0.1.8 → titans_pytorch-0.1.9}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "titans-pytorch"
-version = "0.1.8"
+version = "0.1.9"
 description = "Titans"
 authors = [
     { name = "Phil Wang", email = "lucidrains@gmail.com" }

{titans_pytorch-0.1.8 → titans_pytorch-0.1.9}/titans_pytorch/mac_transformer.py

@@ -83,6 +83,14 @@ def identity(t):
 def round_up_multiple(seq, mult):
     return ceil(seq / mult) * mult
 
+def pack_with_inverse(t, pattern):
+    packed, packed_shape = pack(t, pattern)
+
+    def inverse(out, inv_pattern = None):
+        return unpack(out, packed_shape, default(inv_pattern, pattern))
+
+    return packed, inverse
+
 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)
@@ -576,7 +584,7 @@ class MemoryAsContextTransformer(Module):
         x, inverse_segment = pad_and_segment_with_inverse(x, segment_len)
 
         mems = repeat(self.longterm_mems, 'n d -> b n d', b = x.shape[0])
-        x, mem_ps = pack((x, mems), 'b * d')
+        x, inverse_pack_mems = pack_with_inverse((x, mems), 'b * d')
 
         x = inverse_segment(x)
 
@@ -634,7 +642,7 @@ class MemoryAsContextTransformer(Module):
 
         x, inverse_segment = pad_and_segment_with_inverse(x, segment_len + num_longterm_mem_tokens)
 
-        x, _ = unpack(x, mem_ps, 'b * d')
+        x, _ = inverse_pack_mems(x)
 
         x = inverse_segment(x)
 

{titans_pytorch-0.1.8 → titans_pytorch-0.1.9}/titans_pytorch/titans.py

@@ -391,6 +391,8 @@ class NeuralMemory(Module):
 
         # whether to use averaging of chunks, or attention pooling
 
+        assert not (attn_pool_chunks and chunk_size == 1), '`attn_pool_chunks` cannot be set to True if `chunk_size` is set to 1'
+
         if not attn_pool_chunks:
             chunk_reduce_module = Reduce('b (n c) ... -> b n ...', 'mean', c = chunk_size)
         else: