hippoformer 0.0.8__py3-none-any.whl → 0.0.10__py3-none-any.whl

hippoformer/hippoformer.py

@@ -38,6 +38,40 @@ def pack_with_inverse(t, pattern):
 def l2norm(t):
     return F.normalize(t, dim = -1)
 
+# Muon - Jordan et al from oss community - applied to the latest version of titans
+
+def newtonschulz5(
+    t,
+    steps = 5,
+    eps = 1e-7,
+    coefs = (3.4445, -4.7750, 2.0315)
+):
+    not_weights = t.ndim <= 3
+
+    if not_weights:
+        return t
+
+    shape = t.shape
+    should_transpose = shape[-2] > shape[-1]
+
+    if should_transpose:
+        t = t.transpose(-1, -2)
+
+    t, inv_pack = pack_with_inverse(t, '* i j')
+    t = t / t.norm(dim = (-1, -2), keepdim = True).clamp(min = eps)
+
+    a, b, c = coefs
+
+    for _ in range(steps):
+        A = t @ t.transpose(-1, -2)
+        B = b * A + c * A @ A
+        t = a * t + B @ t
+
+    if should_transpose:
+        t = t.transpose(-1, -2)
+
+    return inv_pack(t)
+
 # sensory encoder decoder for 2d
 
 grid_sensory_enc_dec = (
@@ -209,6 +243,7 @@ class mmTEM(Module):
         loss_weight_consistency = 1.,
         loss_weight_relational = 1.,
         integration_ratio_learned = True,
+        muon_update = False,
         assoc_scan_kwargs: dict = dict()
     ):
         super().__init__()
@@ -287,6 +322,10 @@ class mmTEM(Module):
         self.loss_weight_consistency = loss_weight_consistency
         self.register_buffer('zero', tensor(0.), persistent = False)
 
+        # update with muon
+
+        self.muon_update = muon_update
+
         # there is an integration ratio for error correction, but unclear what value this is fixed to or whether it is learned
 
         self.integration_ratio = nn.Parameter(tensor(0.), requires_grad = integration_ratio_learned)
@@ -416,16 +455,26 @@ class mmTEM(Module):
 
             update = self.assoc_scan(grad, expanded_beta.sigmoid(), momentum)
 
+            # store next momentum
+
+            next_momentum[key] = update[:, -1]
+
+            # maybe muon
+
+            if self.muon_update:
+                update = newtonschulz5(update)
+
+            # with forget gating
+
             expanded_forget = repeat(forget, 'b t -> b t w', w = grad.shape[-1])
 
-            acc_update = self.assoc_scan(update, expanded_forget.sigmoid())
+            acc_update = self.assoc_scan(-update, expanded_forget.sigmoid(), param)
 
             acc_update = inverse_pack(acc_update)
 
             # set the next params and momentum, which can be passed back in
 
-            next_params[key] = param - acc_update[:, -1]
-            next_momentum[key] = update[:, -1]
+            next_params[key] =  acc_update[:, -1]
 
         # losses
 

{hippoformer-0.0.8.dist-info → hippoformer-0.0.10.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: hippoformer
-Version: 0.0.8
+Version: 0.0.10
 Summary: hippoformer
 Project-URL: Homepage, https://pypi.org/project/hippoformer/
 Project-URL: Repository, https://github.com/lucidrains/hippoformer

hippoformer-0.0.10.dist-info/RECORD

@@ -0,0 +1,6 @@
+hippoformer/__init__.py,sha256=A7N8GsRAZH4yP-L5hb7IVDnNjnhfjNyolg5MZ6vnGyE,71
+hippoformer/hippoformer.py,sha256=GWFJy2idp0FWBoVFw8T_6inTXYtY4i47hfhKj88_I0A,14463
+hippoformer-0.0.10.dist-info/METADATA,sha256=IB7iybYMwOkee3Q5ji-B_dnOB62LyK_6t1FPM_UT-FM,2773
+hippoformer-0.0.10.dist-info/WHEEL,sha256=qtCwoSJWgHk21S1Kb4ihdzI2rlJ1ZKaIurTj_ngOhyQ,87
+hippoformer-0.0.10.dist-info/licenses/LICENSE,sha256=1yCiA9b5nhslTavxPjsQAO-wpOnwJR9-l8LTVi7GJuk,1066
+hippoformer-0.0.10.dist-info/RECORD,,

hippoformer-0.0.8.dist-info/RECORD

@@ -1,6 +0,0 @@
-hippoformer/__init__.py,sha256=A7N8GsRAZH4yP-L5hb7IVDnNjnhfjNyolg5MZ6vnGyE,71
-hippoformer/hippoformer.py,sha256=jTR7XRB4bgP6i2VkzOg_Z20Xx7IDg5ifBXYjs6tmrFs,13473
-hippoformer-0.0.8.dist-info/METADATA,sha256=DCgss4I14pexPOjxk319tiqQeBpz5aVRGRQDNvskL9g,2772
-hippoformer-0.0.8.dist-info/WHEEL,sha256=qtCwoSJWgHk21S1Kb4ihdzI2rlJ1ZKaIurTj_ngOhyQ,87
-hippoformer-0.0.8.dist-info/licenses/LICENSE,sha256=1yCiA9b5nhslTavxPjsQAO-wpOnwJR9-l8LTVi7GJuk,1066
-hippoformer-0.0.8.dist-info/RECORD,,