heavyball 1.2.2__py3-none-any.whl → 1.3.0__py3-none-any.whl

heavyball/chainable.py

@@ -183,8 +183,8 @@ def update_by_laprop(group, update, grad, param, exp_avg, exp_avg_sq):
 @no_state
 def update_by_schedule_free(group, update, grad, param, z):
     group['weight_sum'] = utils.schedule_free_(group['lr'], group['weight_lr_power'], group.get('weight_sum', 0),
-                                               utils.get_beta1(group), param, z, update, group['r'], group['step'],
-                                               group['weight_decay'])
+                                               utils.get_beta1(group), param, z, update, grad, group['caution'],
+                                               group['r'], group['step'], group['weight_decay'])
     raise SkipUpdate
 
 
@@ -438,14 +438,15 @@ class ChainOpt(utils.StatefulOptimizer):
 
         for param in p:
             state = self.state_(param)
-            if 'step' not in state:
-                if self.compile_step:
-                    step = utils.scalar_guard(0, param)
-                state['step'] = step
-            step = state['step'].add_(1)
+            if 'step' in state:
+                step = state['step']
+            elif self.compile_step:
+                step = utils.scalar_guard(0, param)
+            else:
+                step = 0
             break
 
-        group['step'] = step
+        group['step'] = state['step'] = step = step + 1
 
         if group['warmup_steps'] and step < group['warmup_steps']:
             group['lr'] = group['base_lr'] * step / group['warmup_steps']
@@ -494,7 +495,7 @@ class BaseOpt(ChainOpt):
     auto_fuse: bool = True
 
     def __init__(self, params, defaults, foreach: bool, gradient_clipping: str_or_fn, update_clipping: str_or_fn,
-                 palm: bool = use_default, compile_step: bool = use_default, *fns):
+                 palm: bool = use_default, *fns, compile_step: bool = use_default):
         if default(update_clipping, self.update_clipping) is None:
             if fns and self.auto_fuse:
                 args, kwargs = None, None

heavyball/utils.py

@@ -61,22 +61,25 @@ def warmup(lr: float, step: int, warmup_steps: int):
 
 
 @decorator_knowngood
-def _compilable_schedule_free_(p: List[Tensor], z: List[Tensor], ckp1: Tensor, grad: List[Tensor], lr: Tensor,
-                               beta1: Tensor, decay: float):
-    for op, oz, g_ in zip(p, z, grad):
-        g_ = g_.view_as(op)
-        p_, z_, g_ = map(promote, (op, oz, g_))
+def _compilable_schedule_free_(p: List[Tensor], z: List[Tensor], ckp1: Tensor, update: List[Tensor], lr: Tensor,
+                               beta1: Tensor, decay: float, grad: List[Tensor], caution):
+    for op, oz, u_, g_ in zip(p, z, update, grad):
+        u_ = u_.view_as(op)
+        p_, z_, u_ = map(promote, (op, oz, u_))
         if decay != 0:
-            g_ = g_ + p_ * decay
+            u_ = u_ + p_ * decay
+        if caution:
+            u_ = _compilable_cautioning(u_, g_)
         p_ = p_.lerp(z_, ckp1)
-        p_ = p_ + g_ * (lr * (beta1 * (1 - ckp1)) - lr)
-        z_ = z_ + g_ * -lr
+        p_ = p_ + u_ * (lr * (beta1 * (1 - ckp1)) - lr)
+        z_ = z_ + u_ * -lr
         copy_stochastic_(op, p_)
         copy_stochastic_(oz, z_)
 
 
 def schedule_free_(lr: float, weight_lr_power: float, weight_sum: float, beta1: float, parameters: List[Tensor],
-                   z: List[Tensor], grad: List[Tensor], r: float = 0.0, step: int = 0, decay: float = 0.0):
+                   z: List[Tensor], update: List[Tensor], grad: List[Tensor], caution: bool = False, r: float = 0.0,
+                   step: int = 0, decay: float = 0.0):
     weight = abs(lr) ** weight_lr_power * max(step, 1) ** r
     weight_sum = weight_sum + weight
 
@@ -85,9 +88,9 @@ def schedule_free_(lr: float, weight_lr_power: float, weight_sum: float, beta1:
     except ZeroDivisionError:
         ckp1 = 0
 
-    grad, parameters, z = list_guard(grad, parameters, z)
+    update, parameters, z, grad = list_guard(update, parameters, z, grad)
     lr, ckp1, beta1 = scalar_guard(lr, ckp1, beta1, grad[0])
-    _compilable_schedule_free_(parameters, z, ckp1, grad, lr, beta1, decay)
+    _compilable_schedule_free_(parameters, z, ckp1, update, lr, beta1, decay, grad, caution)
     return weight_sum
 
 
@@ -909,13 +912,12 @@ def copy_stochastic_(target: Tensor, source: Tensor):
 @decorator_knowngood
 def _compilable_update_(p: List[Tensor], u: List[Tensor], decay: Tensor, lr: Tensor, caution: bool,
                         g: List[Optional[Tensor]]):
-    u = [u_.view_as(p_) for u_, p_ in zip(u, p)]
-    p32, u32 = [list(map(promote, x)) for x in [p, u]]
-
-    for p32_, u32_, g_, p_ in zip(p32, u32, g, p):  # lr is data-dependent -> can't compile a foreach
+    for u_, g_, p_ in zip(u, g, p):  # lr is data-dependent -> can't compile a foreach
+        u_ = promote(u_.view_as(p_))
+        p32_ = promote(p_)
         if caution:
-            u32_ = _compilable_cautioning(promote(g_), u32_)
-        p32_ = p32_ * (1 - decay * lr) + u32_ * -lr
+            u_ = _compilable_cautioning(promote(g_), u_)
+        p32_ = p32_ * (1 - decay * lr) + u_ * -lr
         copy_stochastic_(p_, p32_)
 
 
@@ -1220,13 +1222,16 @@ def update_triu_(q_state, materialised):
         assert shape0 == shape1
         copy_stochastic_(q, m)
 
+
 _warned = set()
 
+
 def warn_once(msg):
     if msg not in _warned:
         warnings.warn(msg)
         _warned.add(msg)
 
+
 def psgd_should_update(group, prob: Union[float, callable], rng: Optional[random.Random] = None,
                        name: str = 'cumulative_prob'):
     group[f'{name}_prob_step'] = group.get(f'{name}_prob_step', 0) + 1
@@ -1369,4 +1374,16 @@ def fused_hook(parameters, optimizer, *args, **kwargs):
             seen_params.clear()
 
     for p in parameters:
-        p.register_post_accumulate_grad_hook(_step)
+        p.register_post_accumulate_grad_hook(_step)
+
+
+@decorator_knowngood
+def _compilable_caution_no_scale(g: Tensor, update: Tensor):
+    mask = g.signbit() ^ update.signbit()  # "Mask if they point in different directions"
+    update = update.masked_fill(mask, 0)
+    return update
+
+
+def disable_caution_scaling():
+    global _compilable_cautioning
+    _compilable_cautioning = _compilable_caution_no_scale

{heavyball-1.2.2.dist-info → heavyball-1.3.0.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: heavyball
-Version: 1.2.2
+Version: 1.3.0
 Summary: Efficient optimizers
 Home-page: https://github.com/clashluke/heavyball
 Author: Lucas Nestler

heavyball-1.3.0.dist-info/RECORD

@@ -0,0 +1,8 @@
+heavyball/__init__.py,sha256=miRgcXlzLWTNzojeRF5hEcg-x_GqfMHjRzOaiR_zO3U,10981
+heavyball/chainable.py,sha256=U-yxSKYIcnMJFvqQlVBKL9BSpFvUAVPg4FxgI5sN21g,21119
+heavyball/utils.py,sha256=0bqa2J3oIp3qvxpUsvQgSh5RBkO9fR9WX8nFFWvcLG0,47901
+heavyball-1.3.0.dist-info/LICENSE,sha256=CGdGJim64YifGmUVPaeyRsxkvyExtClswhRNIp8FY_U,1322
+heavyball-1.3.0.dist-info/METADATA,sha256=1zOefT9mm8_SnOReLEAHHJkY10CC9aWquTCanWHS4ww,12022
+heavyball-1.3.0.dist-info/WHEEL,sha256=GJ7t_kWBFywbagK5eo9IoUwLW6oyOeTKmQ-9iHFVNxQ,92
+heavyball-1.3.0.dist-info/top_level.txt,sha256=SzCxSVg_qCUPA4kZObW3Zyo4v-d_mMOD-p7a-WXTl2E,10
+heavyball-1.3.0.dist-info/RECORD,,

heavyball-1.2.2.dist-info/RECORD

@@ -1,8 +0,0 @@
-heavyball/__init__.py,sha256=miRgcXlzLWTNzojeRF5hEcg-x_GqfMHjRzOaiR_zO3U,10981
-heavyball/chainable.py,sha256=O3vBuwDKSkcl6yGrcEgSylqZ6htjRTg1NIA1sNY4KcA,21076
-heavyball/utils.py,sha256=n80bsTZ2NUD4L3YERaC7ydKaQzW4kSDNoBTLF0DfE1g,47393
-heavyball-1.2.2.dist-info/LICENSE,sha256=CGdGJim64YifGmUVPaeyRsxkvyExtClswhRNIp8FY_U,1322
-heavyball-1.2.2.dist-info/METADATA,sha256=k_qyB9aR8PpreGu_XxNv9z5pyTnH1TMcA8Ah3k1hUgQ,12022
-heavyball-1.2.2.dist-info/WHEEL,sha256=GJ7t_kWBFywbagK5eo9IoUwLW6oyOeTKmQ-9iHFVNxQ,92
-heavyball-1.2.2.dist-info/top_level.txt,sha256=SzCxSVg_qCUPA4kZObW3Zyo4v-d_mMOD-p7a-WXTl2E,10
-heavyball-1.2.2.dist-info/RECORD,,