torchzero 0.3.8__py3-none-any.whl → 0.3.10__py3-none-any.whl

torchzero/modules/momentum/matrix_momentum.py

@@ -2,7 +2,7 @@ from typing import Literal
 
 import torch
 
-from ...core import Module, apply
+from ...core import Module, apply_transform, Chainable
 from ...utils import NumberList, TensorList, as_tensorlist
 from ...utils.derivatives import hvp, hvp_fd_central, hvp_fd_forward
 
@@ -13,105 +13,147 @@ class MatrixMomentum(Module):
 
     `mu` is supposed to be smaller than (1/largest eigenvalue), otherwise this will be very unstable.
 
-    Orr, Genevieve, and Todd Leen. "Using curvature information for fast stochastic search." Advances in neural information processing systems 9 (1996).
+    Args:
+        mu (float, optional): this has a similar role to (1 - beta) in normal momentum. Defaults to 0.1.
+        beta (float, optional): decay for the buffer, this is not part of the original update rule. Defaults to 1.
+        hvp_method (str, optional):
+            How to calculate hessian-vector products.
+            Exact - "autograd", or finite difference - "forward", "central". Defaults to 'forward'.
+        h (float, optional): finite difference step size if hvp_method is set to finite difference. Defaults to 1e-3.
+        hvp_tfm (Chainable | None, optional): optional module applied to hessian-vector products. Defaults to None.
+
+    Reference:
+        Orr, Genevieve, and Todd Leen. "Using curvature information for fast stochastic search." Advances in neural information processing systems 9 (1996).
     """
-    def __init__(self, mu=0.1, beta:float=1, hvp_mode: Literal['autograd', 'forward', 'central'] = 'forward', h=1e-3, hvp_tfm=None):
-        defaults = dict(mu=mu, beta=beta, hvp_mode=hvp_mode, h=h)
+
+    def __init__(
+        self,
+        mu=0.1,
+        beta: float = 1,
+        hvp_method: Literal["autograd", "forward", "central"] = "forward",
+        h: float = 1e-3,
+        hvp_tfm: Chainable | None = None,
+    ):
+        defaults = dict(mu=mu, beta=beta, hvp_method=hvp_method, h=h)
         super().__init__(defaults)
 
         if hvp_tfm is not None:
             self.set_child('hvp_tfm', hvp_tfm)
 
     @torch.no_grad
-    def step(self, vars):
-        assert vars.closure is not None
-        prev_update = self.get_state('prev_update', params=vars.params, cls=TensorList)
-        hvp_mode = self.settings[vars.params[0]]['hvp_mode']
-        h = self.settings[vars.params[0]]['h']
+    def step(self, var):
+        assert var.closure is not None
+        prev_update = self.get_state(var.params, 'prev_update', cls=TensorList)
+        hvp_method = self.settings[var.params[0]]['hvp_method']
+        h = self.settings[var.params[0]]['h']
 
-        mu,beta = self.get_settings('mu','beta', params=vars.params, cls=NumberList)
+        mu,beta = self.get_settings(var.params, 'mu','beta', cls=NumberList)
 
-        if hvp_mode == 'autograd':
+        if hvp_method == 'autograd':
             with torch.enable_grad():
-                grad = vars.get_grad(create_graph=True)
-                hvp_ = TensorList(hvp(vars.params, grads=grad, vec=prev_update, allow_unused=True, retain_graph=False)).detach_()
+                grad = var.get_grad(create_graph=True)
+                hvp_ = TensorList(hvp(var.params, grads=grad, vec=prev_update, allow_unused=True, retain_graph=False)).detach_()
 
-        elif hvp_mode == 'forward':
-            vars.get_grad()
-            l, hvp_ = hvp_fd_forward(vars.closure, vars.params, vec=prev_update, g_0=vars.grad, h=h, normalize=True)
-            if vars.loss_approx is None: vars.loss_approx = l
+        elif hvp_method == 'forward':
+            var.get_grad()
+            l, hvp_ = hvp_fd_forward(var.closure, var.params, vec=prev_update, g_0=var.grad, h=h, normalize=True)
+            if var.loss_approx is None: var.loss_approx = l
 
-        elif hvp_mode == 'central':
-            l, hvp_ = hvp_fd_central(vars.closure, vars.params, vec=prev_update, h=h, normalize=True)
-            if vars.loss_approx is None: vars.loss_approx = l
+        elif hvp_method == 'central':
+            l, hvp_ = hvp_fd_central(var.closure, var.params, vec=prev_update, h=h, normalize=True)
+            if var.loss_approx is None: var.loss_approx = l
 
         else:
-            raise ValueError(hvp_mode)
+            raise ValueError(hvp_method)
 
         if 'hvp_tfm' in self.children:
-            hvp_ = TensorList(apply(self.children['hvp_tfm'], hvp_, params=vars.params, grads=vars.grad, vars=vars))
+            hvp_ = TensorList(apply_transform(self.children['hvp_tfm'], hvp_, params=var.params, grads=var.grad, var=var))
 
-        update = TensorList(vars.get_update())
+        update = TensorList(var.get_update())
 
         hvp_ = as_tensorlist(hvp_)
         update.add_(prev_update - hvp_*mu)
         prev_update.set_(update * beta)
-        vars.update = update
-        return vars
+        var.update = update
+        return var
 
 
 class AdaptiveMatrixMomentum(Module):
     """
-    Mu here is estimated as ||s_k||/||y_k||.
+    May be useful for ill conditioned stochastic quadratic objectives but I need to test this.
+    Evaluates hessian vector product on each step (via finite difference or autograd).
+
+    This version estimates mu via a simple heuristic: ||s||/||y||, where s is parameter difference, y is gradient difference.
+
+    Args:
+        mu_mul (float, optional): multiplier to the estimated mu. Defaults to 1.
+        beta (float, optional): decay for the buffer, this is not part of the original update rule. Defaults to 1.
+        hvp_method (str, optional):
+            How to calculate hessian-vector products.
+            Exact - "autograd", or finite difference - "forward", "central". Defaults to 'forward'.
+        h (float, optional): finite difference step size if hvp_method is set to finite difference. Defaults to 1e-3.
+        hvp_tfm (Chainable | None, optional): optional module applied to hessian-vector products. Defaults to None.
+
+    Reference:
+        Orr, Genevieve, and Todd Leen. "Using curvature information for fast stochastic search." Advances in neural information processing systems 9 (1996).
     """
-    def __init__(self, mu_mul:float=1, beta:float=1, eps=1e-4, hvp_mode: Literal['autograd', 'forward', 'central'] = 'forward', h=1e-3, hvp_tfm=None):
-        defaults = dict(mu_mul=mu_mul, beta=beta, hvp_mode=hvp_mode, h=h, eps=eps)
+
+    def __init__(
+        self,
+        mu_mul: float = 1,
+        beta: float = 1,
+        eps=1e-4,
+        hvp_method: Literal["autograd", "forward", "central"] = "forward",
+        h: float = 1e-3,
+        hvp_tfm: Chainable | None = None,
+    ):
+        defaults = dict(mu_mul=mu_mul, beta=beta, hvp_method=hvp_method, h=h, eps=eps)
         super().__init__(defaults)
 
         if hvp_tfm is not None:
             self.set_child('hvp_tfm', hvp_tfm)
 
     @torch.no_grad
-    def step(self, vars):
-        assert vars.closure is not None
-        prev_update, prev_params, prev_grad = self.get_state('prev_update', 'prev_params', 'prev_grad', params=vars.params, cls=TensorList)
+    def step(self, var):
+        assert var.closure is not None
+        prev_update, prev_params, prev_grad = self.get_state(var.params, 'prev_update', 'prev_params', 'prev_grad', cls=TensorList)
 
-        settings = self.settings[vars.params[0]]
-        hvp_mode = settings['hvp_mode']
+        settings = self.settings[var.params[0]]
+        hvp_method = settings['hvp_method']
         h = settings['h']
         eps = settings['eps']
 
-        mu_mul, beta = self.get_settings('mu_mul','beta', params=vars.params, cls=NumberList)
+        mu_mul, beta = self.get_settings(var.params, 'mu_mul','beta', cls=NumberList)
 
-        if hvp_mode == 'autograd':
+        if hvp_method == 'autograd':
             with torch.enable_grad():
-                grad = vars.get_grad(create_graph=True)
-                hvp_ = TensorList(hvp(vars.params, grads=grad, vec=prev_update, allow_unused=True, retain_graph=False)).detach_()
+                grad = var.get_grad(create_graph=True)
+                hvp_ = TensorList(hvp(var.params, grads=grad, vec=prev_update, allow_unused=True, retain_graph=False)).detach_()
 
-        elif hvp_mode == 'forward':
-            vars.get_grad()
-            l, hvp_ = hvp_fd_forward(vars.closure, vars.params, vec=prev_update, g_0=vars.grad, h=h, normalize=True)
-            if vars.loss_approx is None: vars.loss_approx = l
+        elif hvp_method == 'forward':
+            var.get_grad()
+            l, hvp_ = hvp_fd_forward(var.closure, var.params, vec=prev_update, g_0=var.grad, h=h, normalize=True)
+            if var.loss_approx is None: var.loss_approx = l
 
-        elif hvp_mode == 'central':
-            l, hvp_ = hvp_fd_central(vars.closure, vars.params, vec=prev_update, h=h, normalize=True)
-            if vars.loss_approx is None: vars.loss_approx = l
+        elif hvp_method == 'central':
+            l, hvp_ = hvp_fd_central(var.closure, var.params, vec=prev_update, h=h, normalize=True)
+            if var.loss_approx is None: var.loss_approx = l
 
         else:
-            raise ValueError(hvp_mode)
+            raise ValueError(hvp_method)
 
         if 'hvp_tfm' in self.children:
-            hvp_ = TensorList(apply(self.children['hvp_tfm'], hvp_, params=vars.params, grads=vars.grad, vars=vars))
+            hvp_ = TensorList(apply_transform(self.children['hvp_tfm'], hvp_, params=var.params, grads=var.grad, var=var))
 
         # adaptive part
-        update = TensorList(vars.get_update())
+        update = TensorList(var.get_update())
 
-        s_k = vars.params - prev_params
-        prev_params.copy_(vars.params)
+        s_k = var.params - prev_params
+        prev_params.copy_(var.params)
 
-        assert vars.grad is not None
-        y_k = vars.grad - prev_grad
-        prev_grad.copy_(vars.grad)
+        assert var.grad is not None
+        y_k = var.grad - prev_grad
+        prev_grad.copy_(var.grad)
 
         ada_mu = (s_k.global_vector_norm() / (y_k.global_vector_norm() + eps)) * mu_mul
 
@@ -119,6 +161,6 @@ class AdaptiveMatrixMomentum(Module):
         hvp_ = as_tensorlist(hvp_)
         update.add_(prev_update - hvp_*ada_mu)
         prev_update.set_(update * beta)
-        vars.update = update
-        return vars
+        var.update = update
+        return var
 

torchzero/modules/momentum/momentum.py

@@ -3,11 +3,22 @@ from typing import Literal
 import torch
 
 from ...core import Target, Transform
-from ...utils import NumberList, TensorList
+from ...utils import NumberList, TensorList, unpack_dicts, unpack_states
 from .ema import EMA
 
 
 class HeavyBall(EMA):
+    """Polyak's momentum (heavy-ball method).
+
+    Args:
+        momentum (float, optional): momentum (beta). Defaults to 0.9.
+        dampening (float, optional): momentum dampening. Defaults to 0.
+        debiased (bool, optional): whether to debias the EMA like in Adam. Defaults to False.
+        lerp (bool, optional):
+            whether to use linear interpolation, if True, this becomes exponential moving average. Defaults to False.
+        ema_init (str, optional): initial values for the EMA, "zeros" or "update".
+        target (Target, optional): target to apply EMA to. Defaults to 'update'.
+    """
     def __init__(self, momentum:float=0.9, dampening:float=0, debiased: bool = False, lerp=False, ema_init: Literal['zeros', 'update'] = 'update', target: Target = 'update'):
         super().__init__(momentum=momentum, dampening=dampening, debiased=debiased, lerp=lerp, ema_init=ema_init, target=target)
 
@@ -30,14 +41,23 @@ def nag_(
 
 
 class NAG(Transform):
+    """Nesterov accelerated gradient method (nesterov momentum).
+
+    Args:
+        momentum (float, optional): momentum (beta). Defaults to 0.9.
+        dampening (float, optional): momentum dampening. Defaults to 0.
+        lerp (bool, optional):
+            whether to use linear interpolation, if True, this becomes similar to exponential moving average. Defaults to False.
+        target (Target, optional): target to apply EMA to. Defaults to 'update'.
+    """
     def __init__(self, momentum:float=0.9, dampening:float=0, lerp=False, target: Target = 'update'):
         defaults = dict(momentum=momentum,dampening=dampening, lerp=lerp)
         super().__init__(defaults, uses_grad=False, target=target)
 
     @torch.no_grad
-    def transform(self, tensors, params, grads, vars):
-        velocity = self.get_state('velocity', params=params, cls=TensorList)
+    def apply(self, tensors, params, grads, loss, states, settings):
+        velocity = unpack_states(states, tensors, 'velocity', cls=TensorList)
         lerp = self.settings[params[0]]['lerp']
 
-        momentum,dampening = self.get_settings('momentum','dampening', params=params, cls=NumberList)
+        momentum,dampening = unpack_dicts(settings, 'momentum','dampening', cls=NumberList)
         return nag_(TensorList(tensors), velocity_=velocity,momentum=momentum,dampening=dampening,lerp=lerp)

torchzero/modules/ops/accumulate.py

@@ -5,61 +5,91 @@ from typing import Literal
 import torch
 
 from ...core import Target, Transform
-from ...utils import TensorList, NumberList
+from ...utils import TensorList, NumberList, unpack_states, unpack_dicts
 
 class AccumulateSum(Transform):
+    """Accumulates sum of all past updates.
+
+    Args:
+        decay (float, optional): decays the accumulator. Defaults to 0.
+        target (Target, optional): target. Defaults to 'update'.
+    """
     def __init__(self, decay: float = 0, target: Target = 'update',):
         defaults = dict(decay=decay)
         super().__init__(defaults, uses_grad=False, target=target)
 
     @torch.no_grad
-    def transform(self, tensors, params, grads, vars):
-        sum = self.get_state('sum', params=params, cls=TensorList)
-        decay = self.get_settings('decay', params=params, cls=NumberList)
-        return sum.add_(tensors).lazy_mul(1-decay, clone=True)
+    def apply(self, tensors, params, grads, loss, states, settings):
+        sum = unpack_states(states, tensors, 'sum', cls=TensorList)
+        decay = [1-s['decay'] for s in settings]
+        return sum.add_(tensors).lazy_mul(decay, clone=True)
 
 class AccumulateMean(Transform):
+    """Accumulates mean of all past updates.
+
+    Args:
+        decay (float, optional): decays the accumulator. Defaults to 0.
+        target (Target, optional): target. Defaults to 'update'.
+    """
     def __init__(self, decay: float = 0, target: Target = 'update',):
         defaults = dict(decay=decay)
         super().__init__(defaults, uses_grad=False, target=target)
 
     @torch.no_grad
-    def transform(self, tensors, params, grads, vars):
+    def apply(self, tensors, params, grads, loss, states, settings):
         step = self.global_state['step'] = self.global_state.get('step', 0) + 1
-        mean = self.get_state('mean', params=params, cls=TensorList)
-        decay = self.get_settings('decay', params=params, cls=NumberList)
-        return mean.add_(tensors).lazy_mul(1-decay, clone=True).div_(step)
+        mean = unpack_states(states, tensors, 'mean', cls=TensorList)
+        decay = [1-s['decay'] for s in settings]
+        return mean.add_(tensors).lazy_mul(decay, clone=True).div_(step)
 
 class AccumulateProduct(Transform):
+    """Accumulates product of all past updates.
+
+    Args:
+        decay (float, optional): decays the accumulator. Defaults to 0.
+        target (Target, optional): target. Defaults to 'update'.
+    """
     def __init__(self, decay: float = 0, target: Target = 'update',):
         defaults = dict(decay=decay)
         super().__init__(defaults, uses_grad=False, target=target)
 
     @torch.no_grad
-    def transform(self, tensors, params, grads, vars):
-        prod = self.get_state('prod', params=params, cls=TensorList)
-        decay = self.get_settings('decay', params=params, cls=NumberList)
-        return prod.mul_(tensors).lazy_mul(1-decay, clone=True)
+    def apply(self, tensors, params, grads, loss, states, settings):
+        prod = unpack_states(states, tensors, 'prod', cls=TensorList)
+        decay = [1-s['decay'] for s in settings]
+        return prod.mul_(tensors).lazy_mul(decay, clone=True)
 
 class AccumulateMaximum(Transform):
+    """Accumulates maximum of all past updates.
+
+    Args:
+        decay (float, optional): decays the accumulator. Defaults to 0.
+        target (Target, optional): target. Defaults to 'update'.
+    """
     def __init__(self, decay: float = 0, target: Target = 'update',):
         defaults = dict(decay=decay)
         super().__init__(defaults, uses_grad=False, target=target)
 
     @torch.no_grad
-    def transform(self, tensors, params, grads, vars):
-        maximum = self.get_state('maximum', params=params, cls=TensorList)
-        decay = self.get_settings('decay', params=params, cls=NumberList)
-        return maximum.maximum_(tensors).lazy_mul(1-decay, clone=True)
+    def apply(self, tensors, params, grads, loss, states, settings):
+        maximum = unpack_states(states, tensors, 'maximum', cls=TensorList)
+        decay = [1-s['decay'] for s in settings]
+        return maximum.maximum_(tensors).lazy_mul(decay, clone=True)
 
 class AccumulateMinimum(Transform):
+    """Accumulates minimum of all past updates.
+
+    Args:
+        decay (float, optional): decays the accumulator. Defaults to 0.
+        target (Target, optional): target. Defaults to 'update'.
+    """
     def __init__(self, decay: float = 0, target: Target = 'update',):
         defaults = dict(decay=decay)
         super().__init__(defaults, uses_grad=False, target=target)
 
     @torch.no_grad
-    def transform(self, tensors, params, grads, vars):
-        minimum = self.get_state('minimum', params=params, cls=TensorList)
-        decay = self.get_settings('decay', params=params, cls=NumberList)
-        return minimum.minimum_(tensors).lazy_mul(1-decay, clone=True)
+    def apply(self, tensors, params, grads, loss, states, settings):
+        minimum = unpack_states(states, tensors, 'minimum', cls=TensorList)
+        decay = [1-s['decay'] for s in settings]
+        return minimum.minimum_(tensors).lazy_mul(decay, clone=True)
 

torchzero/modules/ops/binary.py

@@ -7,7 +7,7 @@ from typing import Any
 
 import torch
 
-from ...core import Chainable, Module, Target, Vars, maybe_chain
+from ...core import Chainable, Module, Target, Var, maybe_chain
 from ...utils import TensorList, tensorlist
 
 
@@ -26,25 +26,25 @@ class BinaryOperation(Module, ABC):
                 self.operands[k] = v
 
     @abstractmethod
-    def transform(self, vars: Vars, update: list[torch.Tensor], **operands: Any | list[torch.Tensor]) -> Iterable[torch.Tensor]:
+    def transform(self, var: Var, update: list[torch.Tensor], **operands: Any | list[torch.Tensor]) -> Iterable[torch.Tensor]:
         """applies the operation to operands"""
         raise NotImplementedError
 
     @torch.no_grad
-    def step(self, vars: Vars) -> Vars:
+    def step(self, var: Var) -> Var:
         # pass cloned update to all module operands
         processed_operands: dict[str, Any | list[torch.Tensor]] = self.operands.copy()
 
         for k,v in self.operands.items():
             if k in self.children:
                 v: Module
-                updated_vars = v.step(vars.clone(clone_update=True))
-                processed_operands[k] = updated_vars.get_update()
-                vars.update_attrs_from_clone_(updated_vars) # update loss, grad, etc if this module calculated them
+                updated_var = v.step(var.clone(clone_update=True))
+                processed_operands[k] = updated_var.get_update()
+                var.update_attrs_from_clone_(updated_var) # update loss, grad, etc if this module calculated them
 
-        transformed = self.transform(vars, update=vars.get_update(), **processed_operands)
-        vars.update = list(transformed)
-        return vars
+        transformed = self.transform(var, update=var.get_update(), **processed_operands)
+        var.update = list(transformed)
+        return var
 
 
 class Add(BinaryOperation):
@@ -53,9 +53,9 @@ class Add(BinaryOperation):
         super().__init__(defaults, other=other)
 
     @torch.no_grad
-    def transform(self, vars, update: list[torch.Tensor], other: float | list[torch.Tensor]):
-        if isinstance(other, (int,float)): torch._foreach_add_(update, other * self.settings[vars.params[0]]['alpha'])
-        else: torch._foreach_add_(update, other, alpha=self.settings[vars.params[0]]['alpha'])
+    def transform(self, var, update: list[torch.Tensor], other: float | list[torch.Tensor]):
+        if isinstance(other, (int,float)): torch._foreach_add_(update, other * self.settings[var.params[0]]['alpha'])
+        else: torch._foreach_add_(update, other, alpha=self.settings[var.params[0]]['alpha'])
         return update
 
 class Sub(BinaryOperation):
@@ -64,9 +64,9 @@ class Sub(BinaryOperation):
         super().__init__(defaults, other=other)
 
     @torch.no_grad
-    def transform(self, vars, update: list[torch.Tensor], other: float | list[torch.Tensor]):
-        if isinstance(other, (int,float)): torch._foreach_sub_(update, other * self.settings[vars.params[0]]['alpha'])
-        else: torch._foreach_sub_(update, other, alpha=self.settings[vars.params[0]]['alpha'])
+    def transform(self, var, update: list[torch.Tensor], other: float | list[torch.Tensor]):
+        if isinstance(other, (int,float)): torch._foreach_sub_(update, other * self.settings[var.params[0]]['alpha'])
+        else: torch._foreach_sub_(update, other, alpha=self.settings[var.params[0]]['alpha'])
         return update
 
 class RSub(BinaryOperation):
@@ -74,7 +74,7 @@ class RSub(BinaryOperation):
         super().__init__({}, other=other)
 
     @torch.no_grad
-    def transform(self, vars, update: list[torch.Tensor], other: float | list[torch.Tensor]):
+    def transform(self, var, update: list[torch.Tensor], other: float | list[torch.Tensor]):
         return other - TensorList(update)
 
 class Mul(BinaryOperation):
@@ -82,7 +82,7 @@ class Mul(BinaryOperation):
         super().__init__({}, other=other)
 
     @torch.no_grad
-    def transform(self, vars, update: list[torch.Tensor], other: float | list[torch.Tensor]):
+    def transform(self, var, update: list[torch.Tensor], other: float | list[torch.Tensor]):
         torch._foreach_mul_(update, other)
         return update
 
@@ -91,7 +91,7 @@ class Div(BinaryOperation):
         super().__init__({}, other=other)
 
     @torch.no_grad
-    def transform(self, vars, update: list[torch.Tensor], other: float | list[torch.Tensor]):
+    def transform(self, var, update: list[torch.Tensor], other: float | list[torch.Tensor]):
         torch._foreach_div_(update, other)
         return update
 
@@ -100,7 +100,7 @@ class RDiv(BinaryOperation):
         super().__init__({}, other=other)
 
     @torch.no_grad
-    def transform(self, vars, update: list[torch.Tensor], other: float | list[torch.Tensor]):
+    def transform(self, var, update: list[torch.Tensor], other: float | list[torch.Tensor]):
         return other / TensorList(update)
 
 class Pow(BinaryOperation):
@@ -108,7 +108,7 @@ class Pow(BinaryOperation):
         super().__init__({}, exponent=exponent)
 
     @torch.no_grad
-    def transform(self, vars, update: list[torch.Tensor], exponent: float | list[torch.Tensor]):
+    def transform(self, var, update: list[torch.Tensor], exponent: float | list[torch.Tensor]):
         torch._foreach_pow_(update, exponent)
         return update
 
@@ -117,7 +117,7 @@ class RPow(BinaryOperation):
         super().__init__({}, other=other)
 
     @torch.no_grad
-    def transform(self, vars, update: list[torch.Tensor], other: float | list[torch.Tensor]):
+    def transform(self, var, update: list[torch.Tensor], other: float | list[torch.Tensor]):
         if isinstance(other, (int, float)): return torch._foreach_pow(other, update) # no in-place
         torch._foreach_pow_(other, update)
         return other
@@ -128,8 +128,8 @@ class Lerp(BinaryOperation):
         super().__init__(defaults, end=end)
 
     @torch.no_grad
-    def transform(self, vars, update: list[torch.Tensor], end: list[torch.Tensor]):
-        torch._foreach_lerp_(update, end, weight=self.get_settings('weight',params=vars))
+    def transform(self, var, update: list[torch.Tensor], end: list[torch.Tensor]):
+        torch._foreach_lerp_(update, end, weight=self.get_settings(var.params, 'weight'))
         return update
 
 class CopySign(BinaryOperation):
@@ -137,7 +137,7 @@ class CopySign(BinaryOperation):
         super().__init__({}, other=other)
 
     @torch.no_grad
-    def transform(self, vars, update: list[torch.Tensor], other: list[torch.Tensor]):
+    def transform(self, var, update: list[torch.Tensor], other: list[torch.Tensor]):
         return [u.copysign_(o) for u, o in zip(update, other)]
 
 class RCopySign(BinaryOperation):
@@ -145,7 +145,7 @@ class RCopySign(BinaryOperation):
         super().__init__({}, other=other)
 
     @torch.no_grad
-    def transform(self, vars, update: list[torch.Tensor], other: list[torch.Tensor]):
+    def transform(self, var, update: list[torch.Tensor], other: list[torch.Tensor]):
         return [o.copysign_(u) for u, o in zip(update, other)]
 CopyMagnitude = RCopySign
 
@@ -154,7 +154,7 @@ class Clip(BinaryOperation):
         super().__init__({}, min=min, max=max)
 
     @torch.no_grad
-    def transform(self, vars, update: list[torch.Tensor], min: float | list[torch.Tensor] | None, max: float | list[torch.Tensor] | None):
+    def transform(self, var, update: list[torch.Tensor], min: float | list[torch.Tensor] | None, max: float | list[torch.Tensor] | None):
         return TensorList(update).clamp_(min=min,  max=max)
 
 class MirroredClip(BinaryOperation):
@@ -163,7 +163,7 @@ class MirroredClip(BinaryOperation):
         super().__init__({}, value=value)
 
     @torch.no_grad
-    def transform(self, vars, update: list[torch.Tensor], value: float | list[torch.Tensor]):
+    def transform(self, var, update: list[torch.Tensor], value: float | list[torch.Tensor]):
         min = -value if isinstance(value, (int,float)) else [-v for v in value]
         return TensorList(update).clamp_(min=min,  max=value)
 
@@ -174,8 +174,8 @@ class Graft(BinaryOperation):
         super().__init__(defaults, magnitude=magnitude)
 
     @torch.no_grad
-    def transform(self, vars, update: list[torch.Tensor], magnitude: list[torch.Tensor]):
-        tensorwise, ord, eps = itemgetter('tensorwise','ord','eps')(self.settings[vars.params[0]])
+    def transform(self, var, update: list[torch.Tensor], magnitude: list[torch.Tensor]):
+        tensorwise, ord, eps = itemgetter('tensorwise','ord','eps')(self.settings[var.params[0]])
         return TensorList(update).graft_(magnitude, tensorwise=tensorwise, ord=ord, eps=eps)
 
 class RGraft(BinaryOperation):
@@ -186,8 +186,8 @@ class RGraft(BinaryOperation):
         super().__init__(defaults, direction=direction)
 
     @torch.no_grad
-    def transform(self, vars, update: list[torch.Tensor], direction: list[torch.Tensor]):
-        tensorwise, ord, eps = itemgetter('tensorwise','ord','eps')(self.settings[vars.params[0]])
+    def transform(self, var, update: list[torch.Tensor], direction: list[torch.Tensor]):
+        tensorwise, ord, eps = itemgetter('tensorwise','ord','eps')(self.settings[var.params[0]])
         return TensorList(direction).graft_(update, tensorwise=tensorwise, ord=ord, eps=eps)
 
 GraftToUpdate = RGraft
@@ -197,7 +197,7 @@ class Maximum(BinaryOperation):
         super().__init__({}, other=other)
 
     @torch.no_grad
-    def transform(self, vars, update: list[torch.Tensor], other: list[torch.Tensor]):
+    def transform(self, var, update: list[torch.Tensor], other: list[torch.Tensor]):
         torch._foreach_maximum_(update, other)
         return update
 
@@ -206,7 +206,7 @@ class Minimum(BinaryOperation):
         super().__init__({}, other=other)
 
     @torch.no_grad
-    def transform(self, vars, update: list[torch.Tensor], other: list[torch.Tensor]):
+    def transform(self, var, update: list[torch.Tensor], other: list[torch.Tensor]):
         torch._foreach_minimum_(update, other)
         return update
 
@@ -217,7 +217,7 @@ class GramSchimdt(BinaryOperation):
         super().__init__({}, other=other)
 
     @torch.no_grad
-    def transform(self, vars, update: list[torch.Tensor], other: list[torch.Tensor]):
+    def transform(self, var, update: list[torch.Tensor], other: list[torch.Tensor]):
         update = TensorList(update); other = TensorList(other)
         return update - (other*update) / ((other*other) + 1e-8)
 
@@ -229,8 +229,8 @@ class Threshold(BinaryOperation):
         super().__init__(defaults, threshold=threshold, value=value)
 
     @torch.no_grad
-    def transform(self, vars, update: list[torch.Tensor], threshold: list[torch.Tensor] | float, value: list[torch.Tensor] | float):
-        update_above = self.settings[vars.params[0]]['update_above']
+    def transform(self, var, update: list[torch.Tensor], threshold: list[torch.Tensor] | float, value: list[torch.Tensor] | float):
+        update_above = self.settings[var.params[0]]['update_above']
         update = TensorList(update)
         if update_above:
             if isinstance(value, list): return update.where_(update>threshold, value)

torchzero/modules/ops/debug.py

@@ -10,16 +10,16 @@ class PrintUpdate(Module):
         defaults = dict(text=text, print_fn=print_fn)
         super().__init__(defaults)
 
-    def step(self, vars):
-        self.settings[vars.params[0]]["print_fn"](f'{self.settings[vars.params[0]]["text"]}{vars.update}')
-        return vars
+    def step(self, var):
+        self.settings[var.params[0]]["print_fn"](f'{self.settings[var.params[0]]["text"]}{var.update}')
+        return var
 
 class PrintShape(Module):
     def __init__(self, text = 'shapes = ', print_fn = print):
         defaults = dict(text=text, print_fn=print_fn)
         super().__init__(defaults)
 
-    def step(self, vars):
-        shapes = [u.shape for u in vars.update] if vars.update is not None else None
-        self.settings[vars.params[0]]["print_fn"](f'{self.settings[vars.params[0]]["text"]}{shapes}')
-        return vars
+    def step(self, var):
+        shapes = [u.shape for u in var.update] if var.update is not None else None
+        self.settings[var.params[0]]["print_fn"](f'{self.settings[var.params[0]]["text"]}{shapes}')
+        return var