torchzero 0.3.15__py3-none-any.whl → 0.4.1__py3-none-any.whl

torchzero/modules/step_size/lr.py

@@ -2,7 +2,7 @@
 import torch
 import random
 
-from ...core import Transform
+from ...core import TensorTransform
 from ...utils import NumberList, TensorList, generic_ne, unpack_dicts
 
 def lazy_lr(tensors: TensorList, lr: float | list, inplace:bool):
@@ -12,24 +12,24 @@ def lazy_lr(tensors: TensorList, lr: float | list, inplace:bool):
         return tensors * lr
     return tensors
 
-class LR(Transform):
+class LR(TensorTransform):
     """Learning rate. Adding this module also adds support for LR schedulers."""
     def __init__(self, lr: float):
         defaults=dict(lr=lr)
-        super().__init__(defaults, uses_grad=False)
+        super().__init__(defaults)
 
     @torch.no_grad
-    def apply_tensors(self, tensors, params, grads, loss, states, settings):
+    def multi_tensor_apply(self, tensors, params, grads, loss, states, settings):
         return lazy_lr(TensorList(tensors), lr=[s['lr'] for s in settings], inplace=True)
 
-class StepSize(Transform):
+class StepSize(TensorTransform):
     """this is exactly the same as LR, except the `lr` parameter can be renamed to any other name to avoid clashes"""
     def __init__(self, step_size: float, key = 'step_size'):
         defaults={"key": key, key: step_size}
-        super().__init__(defaults, uses_grad=False)
+        super().__init__(defaults)
 
     @torch.no_grad
-    def apply_tensors(self, tensors, params, grads, loss, states, settings):
+    def multi_tensor_apply(self, tensors, params, grads, loss, states, settings):
         return lazy_lr(TensorList(tensors), lr=[s[s['key']] for s in settings], inplace=True)
 
 
@@ -38,8 +38,8 @@ def _warmup_lr(step: int, start_lr: float | NumberList, end_lr: float | NumberLi
     if step > steps: return end_lr
     return start_lr + (end_lr - start_lr) * (step / steps)
 
-class Warmup(Transform):
-    """Learning rate warmup, linearly increases learning rate multiplier from :code:`start_lr` to :code:`end_lr` over :code:`steps` steps.
+class Warmup(TensorTransform):
+    """Learning rate warmup, linearly increases learning rate multiplier from ``start_lr`` to ``end_lr`` over ``steps`` steps.
 
     Args:
         steps (int, optional): number of steps to perform warmup for. Defaults to 100.
@@ -51,7 +51,7 @@ class Warmup(Transform):
 
         .. code-block:: python
 
-            opt = tz.Modular(
+            opt = tz.Optimizer(
                 model.parameters(),
                 tz.m.Adam(),
                 tz.m.LR(1e-2),
@@ -64,7 +64,7 @@ class Warmup(Transform):
         super().__init__(defaults, uses_grad=False)
 
     @torch.no_grad
-    def apply_tensors(self, tensors, params, grads, loss, states, settings):
+    def multi_tensor_apply(self, tensors, params, grads, loss, states, settings):
         start_lr, end_lr = unpack_dicts(settings, 'start_lr', 'end_lr', cls = NumberList)
         num_steps = settings[0]['steps']
         step = self.global_state.get('step', 0)
@@ -77,7 +77,7 @@ class Warmup(Transform):
         self.global_state['step'] = step + 1
         return tensors
 
-class WarmupNormClip(Transform):
+class WarmupNormClip(TensorTransform):
     """Warmup via clipping of the update norm.
 
     Args:
@@ -90,7 +90,7 @@ class WarmupNormClip(Transform):
 
         .. code-block:: python
 
-            opt = tz.Modular(
+            opt = tz.Optimizer(
                 model.parameters(),
                 tz.m.Adam(),
                 tz.m.WarmupNormClip(steps=1000)
@@ -102,7 +102,7 @@ class WarmupNormClip(Transform):
         super().__init__(defaults, uses_grad=False)
 
     @torch.no_grad
-    def apply_tensors(self, tensors, params, grads, loss, states, settings):
+    def multi_tensor_apply(self, tensors, params, grads, loss, states, settings):
         start_norm, end_norm = unpack_dicts(settings, 'start_norm', 'end_norm', cls = NumberList)
         num_steps = settings[0]['steps']
         step = self.global_state.get('step', 0)
@@ -118,8 +118,8 @@ class WarmupNormClip(Transform):
         return tensors
 
 
-class RandomStepSize(Transform):
-    """Uses random global or layer-wise step size from `low` to `high`.
+class RandomStepSize(TensorTransform):
+    """Uses random global or layer-wise step size from ``low`` to ``high``.
 
     Args:
         low (float, optional): minimum learning rate. Defaults to 0.
@@ -133,7 +133,7 @@ class RandomStepSize(Transform):
         super().__init__(defaults, uses_grad=False)
 
     @torch.no_grad
-    def apply_tensors(self, tensors, params, grads, loss, states, settings):
+    def multi_tensor_apply(self, tensors, params, grads, loss, states, settings):
         s = settings[0]
         parameterwise = s['parameterwise']
 

torchzero/modules/termination/termination.py

@@ -1,11 +1,11 @@
 import time
 from abc import ABC, abstractmethod
 from collections.abc import Sequence
-from typing import cast
+from typing import cast, final
 
 import torch
 
-from ...core import Module, Var
+from ...core import Module, Objective
 from ...utils import Metrics, TensorList, safe_dict_update_, tofloat
 
 
@@ -16,14 +16,15 @@ class TerminationCriteriaBase(Module):
         super().__init__(defaults)
 
     @abstractmethod
-    def termination_criteria(self, var: Var) -> bool:
+    def termination_criteria(self, objective: Objective) -> bool:
         ...
 
-    def should_terminate(self, var: Var) -> bool:
+    @final
+    def should_terminate(self, objective: Objective) -> bool:
         n_bad = self.global_state.get('_n_bad', 0)
         n = self.defaults['_n']
 
-        if self.termination_criteria(var):
+        if self.termination_criteria(objective):
             n_bad += 1
             if n_bad >= n:
                 self.global_state['_n_bad'] = 0
@@ -36,12 +37,12 @@ class TerminationCriteriaBase(Module):
         return False
 
 
-    def update(self, var):
-        var.should_terminate = self.should_terminate(var)
-        if var.should_terminate: self.global_state['_n_bad'] = 0
+    def update(self, objective):
+        objective.should_terminate = self.should_terminate(objective)
+        if objective.should_terminate: self.global_state['_n_bad'] = 0
 
-    def apply(self, var):
-        return var
+    def apply(self, objective):
+        return objective
 
 
 class TerminateAfterNSteps(TerminationCriteriaBase):
@@ -49,7 +50,7 @@ class TerminateAfterNSteps(TerminationCriteriaBase):
         defaults = dict(steps=steps)
         super().__init__(defaults)
 
-    def termination_criteria(self, var):
+    def termination_criteria(self, objective):
         step = self.global_state.get('step', 0)
         self.global_state['step'] = step + 1
 
@@ -61,16 +62,17 @@ class TerminateAfterNEvaluations(TerminationCriteriaBase):
         defaults = dict(maxevals=maxevals)
         super().__init__(defaults)
 
-    def termination_criteria(self, var):
+    def termination_criteria(self, objective):
         maxevals = self.defaults['maxevals']
-        return var.modular.num_evaluations >= maxevals
+        assert objective.modular is not None
+        return objective.modular.num_evaluations >= maxevals
 
 class TerminateAfterNSeconds(TerminationCriteriaBase):
     def __init__(self, seconds:float, sec_fn = time.time):
         defaults = dict(seconds=seconds, sec_fn=sec_fn)
         super().__init__(defaults)
 
-    def termination_criteria(self, var):
+    def termination_criteria(self, objective):
         max_seconds = self.defaults['seconds']
         sec_fn = self.defaults['sec_fn']
 
@@ -88,10 +90,10 @@ class TerminateByGradientNorm(TerminationCriteriaBase):
         defaults = dict(tol=tol, ord=ord)
         super().__init__(defaults, n=n)
 
-    def termination_criteria(self, var):
+    def termination_criteria(self, objective):
         tol = self.defaults['tol']
         ord = self.defaults['ord']
-        return TensorList(var.get_grad()).global_metric(ord) <= tol
+        return TensorList(objective.get_grads()).global_metric(ord) <= tol
 
 
 class TerminateByUpdateNorm(TerminationCriteriaBase):
@@ -100,20 +102,20 @@ class TerminateByUpdateNorm(TerminationCriteriaBase):
         defaults = dict(tol=tol, ord=ord)
         super().__init__(defaults, n=n)
 
-    def termination_criteria(self, var):
+    def termination_criteria(self, objective):
         step = self.global_state.get('step', 0)
         self.global_state['step'] = step + 1
 
         tol = self.defaults['tol']
         ord = self.defaults['ord']
 
-        p_prev = self.get_state(var.params, 'p_prev', cls=TensorList)
+        p_prev = self.get_state(objective.params, 'p_prev', cls=TensorList)
         if step == 0:
-            p_prev.copy_(var.params)
+            p_prev.copy_(objective.params)
             return False
 
-        should_terminate = (p_prev - var.params).global_metric(ord) <= tol
-        p_prev.copy_(var.params)
+        should_terminate = (p_prev - objective.params).global_metric(ord) <= tol
+        p_prev.copy_(objective.params)
         return should_terminate
 
 
@@ -122,10 +124,10 @@ class TerminateOnNoImprovement(TerminationCriteriaBase):
         defaults = dict(tol=tol)
         super().__init__(defaults, n=n)
 
-    def termination_criteria(self, var):
+    def termination_criteria(self, objective):
         tol = self.defaults['tol']
 
-        f = tofloat(var.get_loss(False))
+        f = tofloat(objective.get_loss(False))
         if 'f_min' not in self.global_state:
             self.global_state['f_min'] = f
             return False
@@ -141,9 +143,9 @@ class TerminateOnLossReached(TerminationCriteriaBase):
         defaults = dict(value=value)
         super().__init__(defaults)
 
-    def termination_criteria(self, var):
+    def termination_criteria(self, objective):
         value = self.defaults['value']
-        return var.get_loss(False) <= value
+        return objective.get_loss(False) <= value
 
 class TerminateAny(TerminationCriteriaBase):
     def __init__(self, *criteria: TerminationCriteriaBase):
@@ -151,9 +153,9 @@ class TerminateAny(TerminationCriteriaBase):
 
         self.set_children_sequence(criteria)
 
-    def termination_criteria(self, var: Var) -> bool:
+    def termination_criteria(self, objective: Objective) -> bool:
         for c in self.get_children_sequence():
-            if cast(TerminationCriteriaBase, c).termination_criteria(var): return True
+            if cast(TerminationCriteriaBase, c).termination_criteria(objective): return True
 
         return False
 
@@ -163,9 +165,9 @@ class TerminateAll(TerminationCriteriaBase):
 
         self.set_children_sequence(criteria)
 
-    def termination_criteria(self, var: Var) -> bool:
+    def termination_criteria(self, objective: Objective) -> bool:
         for c in self.get_children_sequence():
-            if not cast(TerminationCriteriaBase, c).termination_criteria(var): return False
+            if not cast(TerminationCriteriaBase, c).termination_criteria(objective): return False
 
         return True
 
@@ -173,7 +175,7 @@ class TerminateNever(TerminationCriteriaBase):
     def __init__(self):
         super().__init__()
 
-    def termination_criteria(self, var): return False
+    def termination_criteria(self, objective): return False
 
 def make_termination_criteria(
     ftol: float | None = None,

torchzero/modules/trust_region/cubic_regularization.py

@@ -5,7 +5,7 @@ import torch
 
 from ...core import Chainable, Module
 from ...utils import TensorList, vec_to_tensors
-from ...utils.linalg.linear_operator import LinearOperator
+from ...linalg.linear_operator import LinearOperator
 from .trust_region import _RADIUS_KEYS, TrustRegionBase, _RadiusStrategy
 
 
@@ -58,7 +58,7 @@ def ls_cubic_solver(f, g:torch.Tensor, H:LinearOperator, M: float, loss_at_param
     for _ in range(it_max):
         r_try = (r_min + r_max) / 2
         lam = r_try * M
-        s_lam = H.add_diagonal(lam).solve(g).neg()
+        s_lam = H.solve_plus_diag(g, lam).neg()
         # s_lam = -torch.linalg.solve(B + lam*id_matrix, g)
         solver_it += 1
         crit = conv_criterion(s_lam, r_try)
@@ -109,7 +109,7 @@ class CubicRegularization(TrustRegionBase):
 
         .. code-block:: python
 
-            opt = tz.Modular(
+            opt = tz.Optimizer(
                 model.parameters(),
                 tz.m.CubicRegularization(tz.m.Newton()),
             )

torchzero/modules/trust_region/levenberg_marquardt.py

@@ -2,7 +2,7 @@
 import torch
 
 from ...core import Chainable, Module
-from ...utils.linalg import linear_operator
+from ...linalg import linear_operator
 from .trust_region import _RADIUS_KEYS, TrustRegionBase, _RadiusStrategy
 
 
@@ -32,38 +32,31 @@ class LevenbergMarquardt(TrustRegionBase):
         max_attempts (max_attempts, optional):
             maximum number of trust region size size reductions per step. A zero update vector is returned when
             this limit is exceeded. Defaults to 10.
+        adaptive (bool, optional):
+            if True, trust radius is multiplied by square root of gradient norm.
         fallback (bool, optional):
             if ``True``, when ``hess_module`` maintains hessian inverse which can't be inverted efficiently, it will
             be inverted anyway. When ``False`` (default), a ``RuntimeError`` will be raised instead.
         inner (Chainable | None, optional): preconditioning is applied to output of thise module. Defaults to None.
 
-    Examples:
-        Gauss-Newton with Levenberg-Marquardt trust-region
+    ### Examples:
 
-        .. code-block:: python
+    Gauss-Newton with Levenberg-Marquardt trust-region
 
-            opt = tz.Modular(
-                model.parameters(),
-                tz.m.LevenbergMarquardt(tz.m.GaussNewton()),
-            )
+    ```python
+    opt = tz.Optimizer(
+        model.parameters(),
+        tz.m.LevenbergMarquardt(tz.m.GaussNewton()),
+    )
+    ```
 
-        LM-SR1
-
-        .. code-block:: python
-
-            opt = tz.Modular(
-                model.parameters(),
-                tz.m.LevenbergMarquardt(tz.m.SR1(inverse=False)),
-            )
-
-        First order trust region (hessian is assumed to be identity)
-
-        .. code-block:: python
-
-            opt = tz.Modular(
-                model.parameters(),
-                tz.m.LevenbergMarquardt(tz.m.Identity()),
-            )
+    LM-SR1
+    ```python
+    opt = tz.Optimizer(
+        model.parameters(),
+        tz.m.LevenbergMarquardt(tz.m.SR1(inverse=False)),
+    )
+    ```
 
     """
     def __init__(
@@ -78,11 +71,12 @@ class LevenbergMarquardt(TrustRegionBase):
         max_attempts: int = 10,
         radius_strategy: _RadiusStrategy | _RADIUS_KEYS = 'default',
         y: float = 0,
+        adaptive: bool = False,
         fallback: bool = False,
         update_freq: int = 1,
         inner: Chainable | None = None,
     ):
-        defaults = dict(y=y, fallback=fallback)
+        defaults = dict(y=y, fallback=fallback, adaptive=adaptive)
         super().__init__(
             defaults=defaults,
             hess_module=hess_module,
@@ -103,6 +97,7 @@ class LevenbergMarquardt(TrustRegionBase):
 
     def trust_solve(self, f, g, H, radius, params, closure, settings):
         y = settings['y']
+        adaptive = settings["adaptive"]
 
         if isinstance(H, linear_operator.DenseInverse):
             if settings['fallback']:
@@ -117,12 +112,14 @@ class LevenbergMarquardt(TrustRegionBase):
                 )
 
         reg = 1/radius
+        if adaptive: reg = reg * torch.linalg.vector_norm(g).sqrt()
+
         if y == 0:
-            return H.add_diagonal(reg).solve(g)
+            return H.solve_plus_diag(g, reg) # pyright:ignore[reportAttributeAccessIssue]
 
         diag = H.diagonal()
         diag = torch.where(diag < torch.finfo(diag.dtype).tiny * 2, 1, diag)
         if y != 1: diag = (diag*y) + (1-y)
-        return H.add_diagonal(diag*reg).solve(g)
+        return H.solve_plus_diag(g, diag*reg)
 
 

torchzero/modules/trust_region/trust_cg.py

@@ -1,7 +1,7 @@
 import torch
 
 from ...core import Chainable, Module
-from ...utils.linalg import cg, linear_operator
+from ...linalg import cg, linear_operator
 from .trust_region import _RADIUS_KEYS, TrustRegionBase, _RadiusStrategy
 
 
@@ -47,7 +47,7 @@ class TrustCG(TrustRegionBase):
 
         .. code-block:: python
 
-            opt = tz.Modular(
+            opt = tz.Optimizer(
                 model.parameters(),
                 tz.m.TrustCG(hess_module=tz.m.SR1(inverse=False)),
             )

torchzero/modules/trust_region/trust_region.py

@@ -7,9 +7,16 @@ from typing import Any, Literal, Protocol, cast, final, overload
 
 import torch
 
-from ...core import Chainable, Module, Var, apply_transform
-from ...utils import TensorList, safe_dict_update_, tofloat, vec_to_tensors, generic_finfo, generic_vector_norm
-from ...utils.linalg.linear_operator import LinearOperator
+from ...core import Chainable, Module, Objective
+from ...linalg.linear_operator import LinearOperator
+from ...utils import (
+    TensorList,
+    generic_finfo,
+    generic_vector_norm,
+    safe_dict_update_,
+    tofloat,
+    vec_to_tensors,
+)
 
 
 def _flatten_tensors(tensors: list[torch.Tensor]):
@@ -256,24 +263,24 @@ class TrustRegionBase(Module, ABC):
         """Solve Hx=g with a trust region penalty/bound defined by `radius`"""
         ... # pylint:disable=unnecessary-ellipsis
 
-    def trust_region_update(self, var: Var, H: LinearOperator | None) -> None:
+    def trust_region_update(self, objective: Objective, H: LinearOperator | None) -> None:
         """updates the state of this module after H or B have been updated, if necessary"""
 
-    def trust_region_apply(self, var: Var, tensors:list[torch.Tensor], H: LinearOperator | None) -> Var:
-        """Solves the trust region subproblem and outputs ``Var`` with the solution direction."""
+    def trust_region_apply(self, objective: Objective, tensors:list[torch.Tensor], H: LinearOperator | None) -> Objective:
+        """Solves the trust region subproblem and outputs ``Objective`` with the solution direction."""
         assert H is not None
 
-        params = TensorList(var.params)
+        params = TensorList(objective.params)
         settings = self.settings[params[0]]
         g = _flatten_tensors(tensors)
 
         max_attempts = settings['max_attempts']
 
         # loss at x_0
-        loss = var.loss
-        closure = var.closure
+        loss = objective.loss
+        closure = objective.closure
         if closure is None: raise RuntimeError("Trust region requires closure")
-        if loss is None: loss = var.get_loss(False)
+        if loss is None: loss = objective.get_loss(False)
         loss = tofloat(loss)
 
         # trust region step and update
@@ -313,38 +320,36 @@ class TrustRegionBase(Module, ABC):
             )
 
         assert d is not None
-        if success: var.update = vec_to_tensors(d, params)
-        else: var.update = params.zeros_like()
+        if success: objective.updates = vec_to_tensors(d, params)
+        else: objective.updates = params.zeros_like()
 
-        return var
+        return objective
 
 
     @final
     @torch.no_grad
-    def update(self, var):
+    def update(self, objective):
         step = self.global_state.get('step', 0)
         self.global_state['step'] = step + 1
 
         if step % self.defaults["update_freq"] == 0:
 
             hessian_module = self.children['hess_module']
-            hessian_module.update(var)
-            H = hessian_module.get_H(var)
+            hessian_module.update(objective)
+            H = hessian_module.get_H(objective)
             self.global_state["H"] = H
 
-            self.trust_region_update(var, H=H)
+            self.trust_region_update(objective, H=H)
 
 
     @final
     @torch.no_grad
-    def apply(self, var):
+    def apply(self, objective):
         H = self.global_state.get('H', None)
 
         # -------------------------------- inner step -------------------------------- #
-        update = var.get_update()
-        if 'inner' in self.children:
-            update = apply_transform(self.children['inner'], update, params=var.params, grads=var.grad, var=var)
+        objective = self.inner_step("inner", objective, must_exist=False)
 
         # ----------------------------------- apply ---------------------------------- #
-        return self.trust_region_apply(var=var, tensors=update, H=H)
+        return self.trust_region_apply(objective=objective, tensors=objective.get_updates(), H=H)
 

torchzero/modules/variance_reduction/svrg.py

@@ -3,15 +3,16 @@ from functools import partial
 
 import torch
 
-from ...core.module import Module
+from ...core import Module, Objective
 from ...utils import tofloat
 
 
-def _reset_except_self(optimizer, var, self: Module):
-    for m in optimizer.unrolled_modules:
+def _reset_except_self(objective: Objective, modules, self: Module):
+    for m in modules:
         if m is not self:
             m.reset()
 
+
 class SVRG(Module):
     """Stochastic variance reduced gradient method (SVRG).
 
@@ -43,7 +44,7 @@ class SVRG(Module):
     ## Examples:
     SVRG-LBFGS
     ```python
-    opt = tz.Modular(
+    opt = tz.Optimizer(
         model.parameters(),
         tz.m.SVRG(len(dataloader)),
         tz.m.LBFGS(),
@@ -53,7 +54,7 @@ class SVRG(Module):
 
     For extra variance reduction one can use Online versions of algorithms, although it won't always help.
     ```python
-    opt = tz.Modular(
+    opt = tz.Optimizer(
         model.parameters(),
         tz.m.SVRG(len(dataloader)),
         tz.m.Online(tz.m.LBFGS()),
@@ -62,7 +63,7 @@ class SVRG(Module):
 
     Variance reduction can also be applied to gradient estimators.
     ```python
-    opt = tz.Modular(
+    opt = tz.Optimizer(
         model.parameters(),
         tz.m.SPSA(),
         tz.m.SVRG(100),
@@ -71,7 +72,7 @@ class SVRG(Module):
     ```
     ## Notes
 
-    The SVRG gradient is computed as ``g_b(x) - alpha * g_b(x_0) - g_f(x0.)``, where:
+    The SVRG gradient is computed as ``g_b(x) - alpha * (g_b(x_0) - g_f(x_0))``, where:
     - ``x`` is current parameters
     - ``x_0`` is initial parameters, where full gradient was computed
     - ``g_b`` refers to mini-batch gradient at ``x`` or ``x_0``
@@ -83,17 +84,18 @@ class SVRG(Module):
         defaults = dict(svrg_steps = svrg_steps, accum_steps=accum_steps, reset_before_accum=reset_before_accum, svrg_loss=svrg_loss, alpha=alpha)
         super().__init__(defaults)
 
+
     @torch.no_grad
-    def step(self, var):
-        params = var.params
-        closure = var.closure
+    def update(self, objective):
+        params = objective.params
+        closure = objective.closure
         assert closure is not None
 
         if "full_grad" not in self.global_state:
 
             # -------------------------- calculate full gradient ------------------------- #
-            if "full_closure" in var.storage:
-                full_closure = var.storage['full_closure']
+            if "full_closure" in objective.storage:
+                full_closure = objective.storage['full_closure']
                 with torch.enable_grad():
                     full_loss = full_closure()
                     if all(p.grad is None for p in params):
@@ -116,12 +118,12 @@ class SVRG(Module):
 
                 # accumulate grads
                 accumulator = self.get_state(params, 'accumulator')
-                grad = var.get_grad()
+                grad = objective.get_grads()
                 torch._foreach_add_(accumulator, grad)
 
                 # accumulate loss
                 loss_accumulator = self.global_state.get('loss_accumulator', 0)
-                loss_accumulator += tofloat(var.loss)
+                loss_accumulator += tofloat(objective.loss)
                 self.global_state['loss_accumulator'] = loss_accumulator
 
                 # on nth step, use the accumulated gradient
@@ -136,10 +138,10 @@ class SVRG(Module):
 
                 # otherwise skip update until enough grads are accumulated
                 else:
-                    var.update = None
-                    var.stop = True
-                    var.skip_update = True
-                    return var
+                    objective.updates = None
+                    objective.stop = True
+                    objective.skip_update = True
+                    return
 
 
         svrg_steps = self.defaults['svrg_steps']
@@ -194,7 +196,7 @@ class SVRG(Module):
 
             return closure(False)
 
-        var.closure = svrg_closure
+        objective.closure = svrg_closure
 
         # --- after svrg_steps steps reset so that new full gradient is calculated on next step --- #
         if current_svrg_step >= svrg_steps:
@@ -203,6 +205,6 @@ class SVRG(Module):
             del self.global_state['full_loss']
             del self.global_state['x_0']
             if self.defaults['reset_before_accum']:
-                var.post_step_hooks.append(partial(_reset_except_self, self=self))
+                objective.post_step_hooks.append(partial(_reset_except_self, self=self))
 
-        return var
+    def apply(self, objective): return objective

torchzero/modules/weight_decay/__init__.py

@@ -1 +1,2 @@
-from .weight_decay import WeightDecay, DirectWeightDecay, decay_weights_, RelativeWeightDecay
+from .weight_decay import WeightDecay, DirectWeightDecay, decay_weights_, RelativeWeightDecay
+from .reinit import RandomReinitialize