heavyball 0.18.0__tar.gz → 0.18.2__tar.gz

{heavyball-0.18.0 → heavyball-0.18.2}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: heavyball
-Version: 0.18.0
+Version: 0.18.2
 Summary: Efficient optimizers
 Home-page: https://github.com/clashluke/heavyball
 Author: Lucas Nestler

{heavyball-0.18.0 → heavyball-0.18.2}/heavyball/cached_delayed_psgd_kron.py

@@ -7,10 +7,9 @@ Source available at https://github.com/evanatyourservice/kron_torch/blob/97a2b5e
 from typing import Optional
 
 import torch
-from heavyball.utils import einsum_base
 
-from .utils import update_param_, warmup, psgd_precond_grad, init_Q_exprs, trust_region_clip_, PSGDBase, \
-    precond_update_prob_schedule, split_p_and_g_in_group, line_to_triu, triu_to_line, set_, einsum_base, promote
+from .utils import update_param_, warmup, init_Q_exprs, trust_region_clip_, PSGDBase, split_p_and_g_in_group, \
+    line_to_triu, triu_to_line, set_, einsum_base, promote
 
 
 class ForeachCachedDelayedPSGDKron(PSGDBase):
@@ -42,7 +41,9 @@ class ForeachCachedDelayedPSGDKron(PSGDBase):
                  max_size_triangular=2048, min_ndim_triangular=2, memory_save_mode=None,
                  momentum_into_precond_update=True, warmup_steps: int = 1, merge_dims: bool = False,
                  split: bool = False, clip_fn: Optional[callable] = None, store_triu_as_line: bool = True,
-                 foreach: bool = True, q_dtype='float32', stochastic_schedule: bool = True):
+                 foreach: bool = True, q_dtype='float32', stochastic_schedule: bool = True,  #
+                 # expert parameters
+                 precond_init_scale=1.0, precond_lr=0.1):
         if not 0.0 <= lr:
             raise ValueError(f"Invalid learning rate: {lr}")
         if not 0.0 <= beta < 1.0:
@@ -55,14 +56,11 @@ class ForeachCachedDelayedPSGDKron(PSGDBase):
 
         defaults = dict(lr=lr, beta=beta, weight_decay=weight_decay, max_size_triangular=max_size_triangular,
                         min_ndim_triangular=min_ndim_triangular, memory_save_mode=memory_save_mode,
-                        momentum_into_precond_update=momentum_into_precond_update, precond_lr=0.1,
-                        # precond lr hardcoded to 0.1
-                        precond_init_scale=1.0,  # precond init scale hardcoded to 1.0
-                        step=0, warmup_steps=warmup_steps, merge_dims=merge_dims, split=split,
-                        store_triu_as_line=store_triu_as_line, q_dtype=q_dtype)
+                        momentum_into_precond_update=momentum_into_precond_update, precond_lr=precond_lr,
+                        precond_init_scale=precond_init_scale, step=0, warmup_steps=warmup_steps, merge_dims=merge_dims,
+                        split=split, store_triu_as_line=store_triu_as_line, q_dtype=q_dtype)
         super().__init__(params, defaults, foreach, stochastic_schedule, clip_fn, preconditioner_update_probability)
 
-
     def _step(self, group):
         momentum_into_precond_update = group.get("momentum_into_precond_update", True)
         precond_init_scale = group['precond_init_scale']
@@ -115,6 +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]
@@ -126,7 +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-0.18.0 → heavyball-0.18.2}/heavyball/cached_psgd_kron.py

@@ -7,10 +7,9 @@ Source available at https://github.com/evanatyourservice/kron_torch/blob/97a2b5e
 from typing import Optional
 
 import torch
-from heavyball.utils import einsum_base
 
-from .utils import update_param_, warmup, psgd_precond_grad, init_Q_exprs, trust_region_clip_, PSGDBase, \
-    precond_update_prob_schedule, split_p_and_g_in_group, line_to_triu, triu_to_line, set_, einsum_base, promote
+from .utils import update_param_, warmup, init_Q_exprs, trust_region_clip_, PSGDBase, \
+    split_p_and_g_in_group, line_to_triu, triu_to_line, set_, einsum_base, promote
 
 
 class ForeachCachedPSGDKron(PSGDBase):
@@ -40,7 +39,9 @@ class ForeachCachedPSGDKron(PSGDBase):
                  max_size_triangular=2048, min_ndim_triangular=2, memory_save_mode=None,
                  momentum_into_precond_update=True, warmup_steps: int = 1, merge_dims: bool = False,
                  split: bool = False, clip_fn: Optional[callable] = None, store_triu_as_line: bool = True,
-                 foreach: bool = True, q_dtype='float32', stochastic_schedule: bool = True):
+                 foreach: bool = True, q_dtype='float32', stochastic_schedule: bool = True,  #
+                 # expert parameters
+                 precond_init_scale=1.0, precond_lr=0.1):
         if not 0.0 <= lr:
             raise ValueError(f"Invalid learning rate: {lr}")
         if not 0.0 <= beta < 1.0:
@@ -53,15 +54,11 @@ class ForeachCachedPSGDKron(PSGDBase):
 
         defaults = dict(lr=lr, beta=beta, weight_decay=weight_decay, max_size_triangular=max_size_triangular,
                         min_ndim_triangular=min_ndim_triangular, memory_save_mode=memory_save_mode,
-                        momentum_into_precond_update=momentum_into_precond_update, precond_lr=0.1,
-                        # precond lr hardcoded to 0.1
-                        precond_init_scale=1.0,  # precond init scale hardcoded to 1.0
-                        step=0, warmup_steps=warmup_steps, merge_dims=merge_dims, split=split,
-                        store_triu_as_line=store_triu_as_line,
-                        q_dtype=q_dtype)
+                        momentum_into_precond_update=momentum_into_precond_update, precond_lr=precond_lr,
+                        precond_init_scale=precond_init_scale, step=0, warmup_steps=warmup_steps, merge_dims=merge_dims,
+                        split=split, store_triu_as_line=store_triu_as_line, q_dtype=q_dtype)
         super().__init__(params, defaults, foreach, stochastic_schedule, clip_fn, preconditioner_update_probability)
 
-
     def _step(self, group):
         momentum_into_precond_update = group.get("momentum_into_precond_update", True)
         precond_init_scale = group['precond_init_scale']
@@ -114,19 +111,18 @@ 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]
-                self.do_update(group,[p], [ea if momentum_into_precond_update else g], [q32], precond_lr, [q_orig], store_triu_as_line)
+                self.do_update(group, [p], [ea if momentum_into_precond_update else g], [q32], precond_lr, [q_orig],
+                               store_triu_as_line)
                 for c_, q_ in zip(cached_q, q):
                     if q_.ndim == 2:
                         torch.matmul(q_.T.conj(), q_, out=c_)
                     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-0.18.0 → heavyball-0.18.2}/heavyball/delayed_psgd.py

@@ -39,7 +39,9 @@ class ForeachDelayedPSGD(PSGDBase):
                  max_size_triangular=2048, min_ndim_triangular=2, memory_save_mode=None,
                  momentum_into_precond_update=True, warmup_steps: int = 1, merge_dims: bool = False,
                  split: bool = False, clip_fn: callable = None, store_triu_as_line: bool = True,
-                 foreach: bool = True, q_dtype='float32', stochastic_schedule: bool = True):
+                 foreach: bool = True, q_dtype='float32', stochastic_schedule: bool = True,  #
+                 # expert parameters
+                 precond_init_scale=1.0, precond_lr=0.1):
         if not 0.0 <= lr:
             raise ValueError(f"Invalid learning rate: {lr}")
         if not 0.0 <= beta < 1.0:
@@ -52,9 +54,8 @@ class ForeachDelayedPSGD(PSGDBase):
 
         defaults = dict(lr=lr, beta=beta, weight_decay=weight_decay, max_size_triangular=max_size_triangular,
                         min_ndim_triangular=min_ndim_triangular, memory_save_mode=memory_save_mode,
-                        momentum_into_precond_update=momentum_into_precond_update, precond_lr=0.1,
-                        # precond lr hardcoded to 0.1
-                        precond_init_scale=1.0,  # precond init scale hardcoded to 1.0
+                        momentum_into_precond_update=momentum_into_precond_update, precond_lr=precond_lr,
+                        precond_init_scale=precond_init_scale,
                         step=0, warmup_steps=warmup_steps, merge_dims=merge_dims, split=split,
                         store_triu_as_line=store_triu_as_line, q_dtype=q_dtype)
         super().__init__(params, defaults, foreach, stochastic_schedule, clip_fn, preconditioner_update_probability)

{heavyball-0.18.0 → heavyball-0.18.2}/heavyball/p_adam.py

@@ -5,10 +5,10 @@ Source available at https://github.com/evanatyourservice/kron_torch/blob/97a2b5e
 """
 
 import torch
-from heavyball.utils import triu_to_line, line_to_triu, identity
 
-from .utils import update_param_, warmup, psgd_precond_grad, init_Q_exprs, PSGDBase, precond_update_prob_schedule, \
-    exp_avg_sq_, beta_debias, split_p_and_g_in_group, promote
+from heavyball.utils import triu_to_line, line_to_triu, identity
+from .utils import update_param_, warmup, psgd_precond_grad, init_Q_exprs, PSGDBase, exp_avg_sq_, beta_debias, \
+    split_p_and_g_in_group, promote
 
 
 class ForeachPaLMPAdam(PSGDBase):
@@ -39,7 +39,9 @@ class ForeachPaLMPAdam(PSGDBase):
                  momentum_into_precond_update=True, warmup_steps: int = 1, betas=(None, None), beta: float = 0.9,
                  beta2_scale: float = 0.8, merge_dims: bool = False, split: bool = False, clip_fn: callable = None,
                  store_triu_as_line: bool = True, foreach: bool = True, q_dtype='float32',
-                 stochastic_schedule: bool = True):
+                 stochastic_schedule: bool = True,  #
+                 # expert parameters
+                 precond_init_scale=1.0, precond_lr=0.1):
         if not 0.0 <= lr:
             raise ValueError(f"Invalid learning rate: {lr}")
         if not 0.0 <= weight_decay:
@@ -52,11 +54,10 @@ class ForeachPaLMPAdam(PSGDBase):
 
         defaults = dict(lr=lr, weight_decay=weight_decay, max_size_triangular=max_size_triangular,
                         min_ndim_triangular=min_ndim_triangular, memory_save_mode=memory_save_mode,
-                        momentum_into_precond_update=momentum_into_precond_update, precond_lr=0.1,
-                        # precond lr hardcoded to 0.1
-                        precond_init_scale=1.0,  # precond init scale hardcoded to 1.0
-                        step=0, warmup_steps=warmup_steps, beta=beta, beta2_scale=beta2_scale, merge_dims=merge_dims,
-                        split=split, store_triu_as_line=store_triu_as_line, q_dtype=q_dtype)
+                        momentum_into_precond_update=momentum_into_precond_update, precond_lr=precond_lr,
+                        precond_init_scale=precond_init_scale, step=0, warmup_steps=warmup_steps, beta=beta,
+                        beta2_scale=beta2_scale, merge_dims=merge_dims, split=split,
+                        store_triu_as_line=store_triu_as_line, q_dtype=q_dtype)
         super().__init__(params, defaults, foreach, stochastic_schedule, clip_fn, preconditioner_update_probability)
 
     def _step(self, group):

{heavyball-0.18.0 → heavyball-0.18.2}/heavyball/psgd_kron.py

@@ -9,7 +9,7 @@ from typing import Optional
 import torch
 
 from .utils import update_param_, warmup, psgd_precond_grad, init_Q_exprs, trust_region_clip_, PSGDBase, \
-    precond_update_prob_schedule, split_p_and_g_in_group, line_to_triu, triu_to_line, set_, promote
+    split_p_and_g_in_group, line_to_triu, triu_to_line, set_, promote
 
 
 class ForeachPSGDKron(PSGDBase):
@@ -39,7 +39,9 @@ class ForeachPSGDKron(PSGDBase):
                  max_size_triangular=2048, min_ndim_triangular=2, memory_save_mode=None,
                  momentum_into_precond_update=True, warmup_steps: int = 1, merge_dims: bool = False,
                  split: bool = False, clip_fn: Optional[callable] = None, store_triu_as_line: bool = True,
-                 foreach: bool = True, q_dtype='float32', stochastic_schedule: bool = True):
+                 foreach: bool = True, q_dtype='float32', stochastic_schedule: bool = True,  #
+                 # expert parameters
+                 precond_init_scale=1.0, precond_lr=0.1):
         if not 0.0 <= lr:
             raise ValueError(f"Invalid learning rate: {lr}")
         if not 0.0 <= beta < 1.0:
@@ -52,14 +54,11 @@ class ForeachPSGDKron(PSGDBase):
 
         defaults = dict(lr=lr, beta=beta, weight_decay=weight_decay, max_size_triangular=max_size_triangular,
                         min_ndim_triangular=min_ndim_triangular, memory_save_mode=memory_save_mode,
-                        momentum_into_precond_update=momentum_into_precond_update, precond_lr=0.1,
-                        # precond lr hardcoded to 0.1
-                        precond_init_scale=1.0,  # precond init scale hardcoded to 1.0
-                        step=0, warmup_steps=warmup_steps, merge_dims=merge_dims, split=split,
-                        store_triu_as_line=store_triu_as_line, q_dtype=q_dtype)
+                        momentum_into_precond_update=momentum_into_precond_update, precond_lr=precond_lr,
+                        precond_init_scale=precond_init_scale, step=0, warmup_steps=warmup_steps, merge_dims=merge_dims,
+                        split=split, store_triu_as_line=store_triu_as_line, q_dtype=q_dtype)
         super().__init__(params, defaults, foreach, stochastic_schedule, clip_fn, preconditioner_update_probability)
 
-
     def _step(self, group):
         momentum_into_precond_update = group.get("momentum_into_precond_update", True)
         precond_init_scale = group['precond_init_scale']
@@ -104,7 +103,7 @@ class ForeachPSGDKron(PSGDBase):
 
             if self.should_update(group):
                 q32 = [promote(q_) for q_ in q]
-                self.do_update(group,[p], [g], [q32], precond_lr, [q_orig], store_triu_as_line)
+                self.do_update(group, [p], [g], [q32], precond_lr, [q_orig], store_triu_as_line)
             set_(g, psgd_precond_grad(q, self.state_(p)["exprs"], ea))
 
         grad_list = self.clip_fn(grad_list)

{heavyball-0.18.0 → heavyball-0.18.2}/heavyball/pure_psgd.py

@@ -5,10 +5,10 @@ Source available at https://github.com/evanatyourservice/kron_torch/blob/97a2b5e
 """
 
 import torch
-from heavyball.utils import copy_stochastic_list_, identity
 
-from .utils import update_param_, warmup, psgd_precond_grad, init_Q_exprs, PSGDBase, precond_update_prob_schedule, \
-    split_p_and_g_in_group, line_to_triu, triu_to_line, promote
+from heavyball.utils import identity
+from .utils import update_param_, warmup, psgd_precond_grad, init_Q_exprs, PSGDBase, split_p_and_g_in_group, \
+    line_to_triu, triu_to_line, promote
 
 
 class ForeachPurePSGD(PSGDBase):
@@ -37,8 +37,10 @@ class ForeachPurePSGD(PSGDBase):
     def __init__(self, params, lr=0.001, weight_decay=0.0, preconditioner_update_probability=None,
                  max_size_triangular=2048, min_ndim_triangular=2, memory_save_mode=None,
                  momentum_into_precond_update=True, warmup_steps: int = 1, merge_dims: bool = False,
-                 split: bool = False, clip_fn: callable = None, store_triu_as_line: bool = True,
-                 foreach: bool = True, q_dtype='float32', stochastic_schedule: bool = True):
+                 split: bool = False, clip_fn: callable = None, store_triu_as_line: bool = True, foreach: bool = True,
+                 q_dtype='float32', stochastic_schedule: bool = True,  #
+                 # expert parameters
+                 precond_init_scale=1.0, precond_lr=0.1):
         if not 0.0 <= lr:
             raise ValueError(f"Invalid learning rate: {lr}")
         if not 0.0 <= weight_decay:
@@ -49,11 +51,9 @@ class ForeachPurePSGD(PSGDBase):
 
         defaults = dict(lr=lr, weight_decay=weight_decay, max_size_triangular=max_size_triangular,
                         min_ndim_triangular=min_ndim_triangular, memory_save_mode=memory_save_mode,
-                        momentum_into_precond_update=momentum_into_precond_update, precond_lr=0.1,
-                        # precond lr hardcoded to 0.1
-                        precond_init_scale=1.0,  # precond init scale hardcoded to 1.0
-                        step=0, warmup_steps=warmup_steps, merge_dims=merge_dims, split=split,
-                        store_triu_as_line=store_triu_as_line, q_dtype=q_dtype)
+                        momentum_into_precond_update=momentum_into_precond_update, precond_lr=precond_lr,
+                        precond_init_scale=precond_init_scale, step=0, warmup_steps=warmup_steps, merge_dims=merge_dims,
+                        split=split, store_triu_as_line=store_triu_as_line, q_dtype=q_dtype)
         super().__init__(params, defaults, foreach, stochastic_schedule, clip_fn, preconditioner_update_probability)
 
     def _step(self, group):
@@ -95,7 +95,7 @@ class ForeachPurePSGD(PSGDBase):
 
             if self.should_update(group):
                 q32 = [promote(q_) for q_ in q]
-                self.do_update(group,[p], [g], [q32], precond_lr, [q_orig], store_triu_as_line)
+                self.do_update(group, [p], [g], [q32], precond_lr, [q_orig], store_triu_as_line)
             psgd_precond_grad(q, self.state_(p)["exprs"], g, inplace=True)
 
         grad_list = self.clip_fn(grad_list)

{heavyball-0.18.0 → 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-0.18.0 → heavyball-0.18.2}/heavyball.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: heavyball
-Version: 0.18.0
+Version: 0.18.2
 Summary: Efficient optimizers
 Home-page: https://github.com/clashluke/heavyball
 Author: Lucas Nestler

{heavyball-0.18.0 → heavyball-0.18.2}/setup.py

@@ -10,7 +10,7 @@ setuptools.setup(
     name='heavyball',
     license='BSD',
     description='Efficient optimizers',
-    version='0.18.0',
+    version='0.18.2',
     long_description=README,
     url='https://github.com/clashluke/heavyball',
     packages=setuptools.find_packages(),

{heavyball-0.18.0 → heavyball-0.18.2}/test/test_stochastic_updates.py

@@ -16,8 +16,8 @@ def get_memory():
 
 
 @pytest.mark.parametrize("opt", heavyball.__all__)
-@pytest.mark.parametrize("size,depth", [(128, 2)])
-def test_foreach(opt, size, depth: int, iterations: int = 1024, outer_iterations: int = 3):
+@pytest.mark.parametrize("size,depth", [(128, 1)])
+def test_foreach(opt, size, depth: int, iterations: int = 8192, outer_iterations: int = 3):
     set_torch()
 
     opt = getattr(heavyball, opt)
@@ -28,12 +28,13 @@ def test_foreach(opt, size, depth: int, iterations: int = 1024, 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):