heavyball 0.18.1__tar.gz → 0.18.2__tar.gz

{heavyball-0.18.1 → heavyball-0.18.2}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: heavyball
-Version: 0.18.1
+Version: 0.18.2
 Summary: Efficient optimizers
 Home-page: https://github.com/clashluke/heavyball
 Author: Lucas Nestler

{heavyball-0.18.1 → heavyball-0.18.2}/heavyball/cached_delayed_psgd_kron.py

@@ -113,6 +113,8 @@ class ForeachCachedDelayedPSGDKron(PSGDBase):
             q_orig = Q_list.pop(0)
             ea = exp_avg_list.pop(0)
 
+            new = torch.einsum(self.state_(p)['cache_expr'], *cached_q, ea)
+
             if self.should_update(group):
                 q = line_to_triu(q_orig) if store_triu_as_line else q_orig
                 q32 = [promote(q_) for q_ in q]
@@ -124,7 +126,7 @@ class ForeachCachedDelayedPSGDKron(PSGDBase):
                     else:
                         torch.mul(q_.conj(), q_, out=c_)
 
-            set_(g, torch.einsum(self.state_(p)['cache_expr'], *cached_q, ea))
+            set_(g, new)
         grad_list = self.clip_fn(grad_list)
 
         lr = -warmup(lr, group['step'], group['warmup_steps'])

{heavyball-0.18.1 → heavyball-0.18.2}/heavyball/cached_psgd_kron.py

@@ -111,8 +111,6 @@ class ForeachCachedPSGDKron(PSGDBase):
             q_orig = Q_list.pop(0)
             ea = exp_avg_list.pop(0)
 
-            new = torch.einsum(self.state_(p)['cache_expr'], *cached_q, ea)
-
             if self.should_update(group):
                 q = line_to_triu(q_orig) if store_triu_as_line else q_orig
                 q32 = [promote(q_) for q_ in q]
@@ -124,7 +122,7 @@ class ForeachCachedPSGDKron(PSGDBase):
                     else:
                         torch.mul(q_.conj(), q_, out=c_)
 
-            set_(g, new)
+            set_(g, torch.einsum(self.state_(p)['cache_expr'], *cached_q, ea))
 
         grad_list = self.clip_fn(grad_list)
 

{heavyball-0.18.1 → heavyball-0.18.2}/heavyball/utils.py

@@ -817,14 +817,15 @@ class PSGDBase(StatefulOptimizer):
 
     def do_update(self, group, p_list, grad_list, q_list, precond_lr, original_q: Optional[List] = None,
                   store_triu_as_line=False):
-        if self.should_update(group, self.balance_probability, 'balance_prob'):
-            for g, q in zip(grad_list, q_list):
-                if g.dim() > 1:
-                    psgd_balance_Q(q)
-
         for i, (p, grad, Q) in enumerate(zip(p_list, grad_list, q_list)):
             psgd_update_precond(Q, self.state_(p)["exprs"], torch.randn_like(grad), grad, precond_lr, self._tiny)
-            if original_q:
+
+        for g, q in zip(grad_list, q_list):
+            if g.dim() > 1:
+                psgd_balance_Q(q)
+
+        if original_q:
+            for q in q_list:
                 if store_triu_as_line:
                     update_triu_(original_q[i], Q)
                 else:

{heavyball-0.18.1 → heavyball-0.18.2}/heavyball.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: heavyball
-Version: 0.18.1
+Version: 0.18.2
 Summary: Efficient optimizers
 Home-page: https://github.com/clashluke/heavyball
 Author: Lucas Nestler

{heavyball-0.18.1 → heavyball-0.18.2}/setup.py

@@ -10,7 +10,7 @@ setuptools.setup(
     name='heavyball',
     license='BSD',
     description='Efficient optimizers',
-    version='0.18.1',
+    version='0.18.2',
     long_description=README,
     url='https://github.com/clashluke/heavyball',
     packages=setuptools.find_packages(),

{heavyball-0.18.1 → heavyball-0.18.2}/test/test_stochastic_updates.py

@@ -16,7 +16,7 @@ def get_memory():
 
 
 @pytest.mark.parametrize("opt", heavyball.__all__)
-@pytest.mark.parametrize("size,depth", [(128, 2)])
+@pytest.mark.parametrize("size,depth", [(128, 1)])
 def test_foreach(opt, size, depth: int, iterations: int = 8192, outer_iterations: int = 3):
     set_torch()
 
@@ -28,12 +28,13 @@ def test_foreach(opt, size, depth: int, iterations: int = 8192, outer_iterations
     losses = []
 
     for stochastic in [False, True]:
+        print('stochastic', stochastic)
         torch.manual_seed(0x2131290)
         peaks.append([])
         losses.append([])
 
         for i in range(outer_iterations):
-            model = nn.Sequential(*[nn.Linear(size, size) for _ in range(depth)]).cuda()
+            model = nn.Sequential(*[nn.Linear(size, size, bias=False) for _ in range(depth)]).cuda()
             o = get_optim(opt, model.parameters(), lr=1e-3, stochastic_schedule=stochastic)
 
             for _ in range(iterations):