hyper-connections 0.3.16__tar.gz → 0.4.1__tar.gz

{hyper_connections-0.3.16 → hyper_connections-0.4.1}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: hyper-connections
-Version: 0.3.16
+Version: 0.4.1
 Summary: Hyper-Connections
 Project-URL: Homepage, https://pypi.org/project/hyper-connections/
 Project-URL: Repository, https://github.com/lucidrains/hyper-connections

{hyper_connections-0.3.16 → hyper_connections-0.4.1}/hyper_connections/manifold_constrained_hyper_connections.py

@@ -21,6 +21,7 @@ s - residual streams
 t - residual streams + num branch inputs
 f - number of fractions (division of feature dimension space)
 v - number of views for branch input
+p - proposals
 """
 
 # helper functions
@@ -212,7 +213,10 @@ class ManifoldConstrainedHyperConnections(Module):
         num_fracs = 1,                      # https://arxiv.org/abs/2503.14125
         sinkhorn_iters = 20,
         log_domain_sinkhorn = False,
+        residual_constraint_fn: Callable | None = None,
         forward_method_names: tuple[str, ...] = (),
+        num_dynamic_alpha_proposals = 1,
+
     ):
         """
         Appendix J, Algorithm2 in - https://arxiv.org/abs/2409.19606
@@ -256,6 +260,11 @@ class ManifoldConstrainedHyperConnections(Module):
         assert num_input_views >= 1
         self.num_input_views = num_input_views
 
+        # number of dynamic alpha proposals, for averaging Hres across proposals
+
+        self.has_dynamic_alpha_proposals = num_dynamic_alpha_proposals > 1
+        self.num_dynamic_alpha_proposals = num_dynamic_alpha_proposals
+
         # width connection
 
         init_alpha0 = torch.zeros((num_residual_streams_fracs, num_input_views_fracs))
@@ -263,7 +272,7 @@ class ManifoldConstrainedHyperConnections(Module):
 
         self.static_alpha = nn.Parameter(cat((init_alpha0, torch.eye(num_residual_streams_fracs)), dim = 1))
 
-        self.dynamic_alpha_fn = nn.Parameter(torch.zeros(dim, num_residual_streams_fracs + num_input_views_fracs))
+        self.dynamic_alpha_fn = nn.Parameter(torch.zeros(num_dynamic_alpha_proposals, dim, num_residual_streams_fracs + num_input_views_fracs))
 
         self.pre_branch_scale = nn.Parameter(torch.ones(1) * 1e-2)
         self.residual_scale = nn.Parameter(torch.ones(1) * 1e-2)
@@ -280,10 +289,13 @@ class ManifoldConstrainedHyperConnections(Module):
 
             self.h_post_scale = nn.Parameter(torch.ones(()) * 1e-2)
 
-        # sinkhorn related
+        # Hres constraint related
+        # by default is sinkhorn
 
-        self.sinkhorn_iters = sinkhorn_iters
-        self.log_domain_sinkhorn = log_domain_sinkhorn
+        self.residual_constraint_fn = default(
+            residual_constraint_fn,
+            partial(sinkhorn_knopps if not log_domain_sinkhorn else log_domain_sinkhorn_knopps, iters = sinkhorn_iters)
+        )
 
         # dropouts
 
@@ -346,7 +358,7 @@ class ManifoldConstrainedHyperConnections(Module):
 
         normed = normed.float()
 
-        wc_weight = normed @ self.dynamic_alpha_fn.float()
+        wc_weight = einsum(normed, self.dynamic_alpha_fn.float(), '... d, p d e -> p ... e')
 
         pre_branch_scale = repeat(self.pre_branch_scale.float(), '1 -> s', s = self.num_fracs)
         residual_scale = repeat(self.residual_scale.float(), '1 -> s', s = self.num_fracs * streams)
@@ -366,12 +378,15 @@ class ManifoldConstrainedHyperConnections(Module):
 
         alpha_pre = alpha_pre.sigmoid()
 
-        sinkhorn_fn = sinkhorn_knopps if not self.log_domain_sinkhorn else log_domain_sinkhorn_knopps
-
-        alpha_residual = sinkhorn_fn(alpha_residual, self.sinkhorn_iters)
+        alpha_residual = self.residual_constraint_fn(alpha_residual)
 
         alpha = cat((alpha_pre, alpha_residual), dim = -1)
 
+        if self.has_dynamic_alpha_proposals:
+            alpha = reduce(alpha, 'p ... -> ...', 'mean')
+        else:
+            alpha = rearrange(alpha, '1 ... -> ...')
+
         alpha = self.split_fracs(alpha) # (batch, seq, fracs1, streams, fracs2, input + residual streams)
 
         # beta for weights from branch output back to residual streams

{hyper_connections-0.3.16 → hyper_connections-0.4.1}/hyper_connections/vit.py

@@ -66,12 +66,12 @@ class Attention(Module):
         return self.to_out(out)
 
 class Transformer(Module):
-    def __init__(self, dim, depth, heads, dim_head, mlp_dim, dropout = 0., num_residual_streams = 4):
+    def __init__(self, dim, depth, heads, dim_head, mlp_dim, dropout = 0., num_residual_streams = 4, num_dynamic_alpha_proposals = 1):
         super().__init__()
         self.norm = nn.LayerNorm(dim)
         self.layers = ModuleList([])
 
-        init_hyper_conn, self.expand_streams, self.reduce_streams = mHC.get_init_and_expand_reduce_stream_functions(num_residual_streams)
+        init_hyper_conn, self.expand_streams, self.reduce_streams = mHC.get_init_and_expand_reduce_stream_functions(num_residual_streams, num_dynamic_alpha_proposals = num_dynamic_alpha_proposals)
 
         for _ in range(depth):
             self.layers.append(ModuleList([
@@ -92,7 +92,7 @@ class Transformer(Module):
         return self.norm(x)
 
 class ViT(Module):
-    def __init__(self, *, image_size, patch_size, num_classes, dim, depth, heads, mlp_dim, pool = 'cls', channels = 3, dim_head = 64, dropout = 0., emb_dropout = 0., num_residual_streams = 4):
+    def __init__(self, *, image_size, patch_size, num_classes, dim, depth, heads, mlp_dim, pool = 'cls', channels = 3, dim_head = 64, dropout = 0., emb_dropout = 0., num_residual_streams = 4, num_dynamic_alpha_proposals = 1):
         super().__init__()
         image_height, image_width = pair(image_size)
         patch_height, patch_width = pair(patch_size)
@@ -117,7 +117,7 @@ class ViT(Module):
 
         self.dropout = nn.Dropout(emb_dropout)
 
-        self.transformer = Transformer(dim, depth, heads, dim_head, mlp_dim, dropout)
+        self.transformer = Transformer(dim, depth, heads, dim_head, mlp_dim, dropout, num_residual_streams, num_dynamic_alpha_proposals)
 
         self.pool = pool
         self.to_latent = nn.Identity()

{hyper_connections-0.3.16 → hyper_connections-0.4.1}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "hyper-connections"
-version = "0.3.16"
+version = "0.4.1"
 description = "Hyper-Connections"
 authors = [
     { name = "Phil Wang", email = "lucidrains@gmail.com" }

{hyper_connections-0.3.16 → hyper_connections-0.4.1}/tests/test_hyper_connections.py

@@ -1,11 +1,12 @@
 import pytest
+param = pytest.mark.parametrize
 
 import torch
 from torch import nn
 
-@pytest.mark.parametrize('num_fracs', (1, 4))
-@pytest.mark.parametrize('disable', (False, True))
-@pytest.mark.parametrize('manifold_constrained', (False, True))
+@param('num_fracs', (1, 4))
+@param('disable', (False, True))
+@param('manifold_constrained', (False, True))
 def test_readme(
     num_fracs,
     disable,
@@ -208,7 +209,11 @@ def test_mhc_dtype_restoration():
 
     assert residual.dtype == torch.half
 
-def test_mhc_vit():
+@param('num_dynamic_alpha_proposals', (1, 2))
+def test_mhc_vit(
+    num_dynamic_alpha_proposals
+):
+
     from hyper_connections.vit import ViT
 
     v = ViT(
@@ -221,7 +226,8 @@ def test_mhc_vit():
         mlp_dim = 2048,
         dropout = 0.1,
         emb_dropout = 0.1,
-        num_residual_streams = 4
+        num_residual_streams = 4,
+        num_dynamic_alpha_proposals = num_dynamic_alpha_proposals
     )
 
     img = torch.randn(1, 3, 256, 256)