torchzero 0.3.15__py3-none-any.whl → 0.4.0__py3-none-any.whl

torchzero/modules/second_order/rsn.py

@@ -5,9 +5,10 @@ from typing import Literal
 
 import torch
 
-from ...core import Chainable, Module, apply_transform
-from ...utils import Distributions, TensorList, vec_to_tensors
-from ...utils.linalg.linear_operator import Sketched
+from ...core import Chainable, Transform, HVPMethod
+from ...utils import vec_to_tensors
+from ...linalg.linear_operator import Sketched
+
 from .newton import _newton_step
 
 def _qr_orthonormalize(A:torch.Tensor):
@@ -15,9 +16,9 @@ def _qr_orthonormalize(A:torch.Tensor):
     if m < n:
         q, _ = torch.linalg.qr(A.T) # pylint:disable=not-callable
         return q.T
-    else:
-        q, _ = torch.linalg.qr(A) # pylint:disable=not-callable
-        return q
+
+    q, _ = torch.linalg.qr(A) # pylint:disable=not-callable
+    return q
 
 def _orthonormal_sketch(m, n, dtype, device, generator):
     return _qr_orthonormalize(torch.randn(m, n, dtype=dtype, device=device, generator=generator))
@@ -25,26 +26,31 @@ def _orthonormal_sketch(m, n, dtype, device, generator):
 def _gaussian_sketch(m, n, dtype, device, generator):
     return torch.randn(m, n, dtype=dtype, device=device, generator=generator) / math.sqrt(m)
 
-class RSN(Module):
-    """Randomized Subspace Newton. Performs a Newton step in a random subspace.
+def _rademacher_sketch(m, n, dtype, device, generator):
+    rademacher = torch.bernoulli(torch.full((m,n), 0.5), generator = generator).mul_(2).sub_(1)
+    return rademacher.mul_(1 / math.sqrt(m))
+
+class SubspaceNewton(Transform):
+    """Subspace Newton. Performs a Newton step in a subspace (random or spanned by past gradients).
 
     Args:
         sketch_size (int):
             size of the random sketch. This many hessian-vector products will need to be evaluated each step.
         sketch_type (str, optional):
             - "orthonormal" - random orthonormal basis. Orthonormality is necessary to use linear operator based modules such as trust region, but it can be slower to compute.
+            - "rademacher" - approximately orthonormal scaled random rademacher basis.
             - "gaussian" - random gaussian (not orthonormal) basis.
             - "common_directions" - uses history steepest descent directions as the basis[2]. It is orthonormalized on-line using Gram-Schmidt.
-            - "mixed" - random orthonormal basis but with three directions set to gradient, slow EMA and fast EMA (default).
+            - "mixed" - random orthonormal basis but with four directions set to gradient, slow and fast gradient EMAs, and previous update direction (default).
         damping (float, optional): hessian damping (scale of identity matrix added to hessian). Defaults to 0.
         hvp_method (str, optional):
             How to compute hessian-matrix product:
-            - "batched" - uses batched autograd
+            - "batched_autograd" - uses batched autograd
             - "autograd" - uses unbatched autograd
             - "forward" - uses finite difference with forward formula, performing 1 backward pass per Hvp.
             - "central" - uses finite difference with a more accurate central formula, performing 2 backward passes per Hvp.
 
-            . Defaults to "batched".
+            . Defaults to "batched_autograd".
         h (float, optional): finite difference step size. Defaults to 1e-2.
         use_lstsq (bool, optional): whether to use least squares to solve ``Hx=g``. Defaults to False.
         update_freq (int, optional): frequency of updating the hessian. Defaults to 1.
@@ -93,7 +99,7 @@ class RSN(Module):
         sketch_size: int,
         sketch_type: Literal["orthonormal", "gaussian", "common_directions", "mixed"] = "mixed",
         damping:float=0,
-        hvp_method: Literal["batched", "autograd", "forward", "central"] = "batched",
+        hvp_method: HVPMethod = "batched_autograd",
         h: float = 1e-2,
         use_lstsq: bool = True,
         update_freq: int = 1,
@@ -102,115 +108,119 @@ class RSN(Module):
         seed: int | None = None,
         inner: Chainable | None = None,
     ):
-        defaults = dict(sketch_size=sketch_size, sketch_type=sketch_type,seed=seed,hvp_method=hvp_method, h=h, damping=damping, use_lstsq=use_lstsq, H_tfm=H_tfm, eigval_fn=eigval_fn, update_freq=update_freq)
-        super().__init__(defaults)
-
-        if inner is not None:
-            self.set_child("inner", inner)
+        defaults = locals().copy()
+        del defaults['self'], defaults['inner'], defaults["update_freq"]
+        super().__init__(defaults, update_freq=update_freq, inner=inner)
 
     @torch.no_grad
-    def update(self, var):
-        step = self.global_state.get('step', 0)
-        self.global_state['step'] = step + 1
-
-        if step % self.defaults['update_freq'] == 0:
+    def update_states(self, objective, states, settings):
+        fs = settings[0]
+        params = objective.params
+        generator = self.get_generator(params[0].device, fs["seed"])
 
-            closure = var.closure
-            if closure is None:
-                raise RuntimeError("RSN requires closure")
-            params = var.params
-            generator = self.get_generator(params[0].device, self.defaults["seed"])
+        ndim = sum(p.numel() for p in params)
 
-            ndim = sum(p.numel() for p in params)
+        device=params[0].device
+        dtype=params[0].dtype
 
-            device=params[0].device
-            dtype=params[0].dtype
+        # sample sketch matrix S: (ndim, sketch_size)
+        sketch_size = min(fs["sketch_size"], ndim)
+        sketch_type = fs["sketch_type"]
+        hvp_method = fs["hvp_method"]
 
-            # sample sketch matrix S: (ndim, sketch_size)
-            sketch_size = min(self.defaults["sketch_size"], ndim)
-            sketch_type = self.defaults["sketch_type"]
-            hvp_method = self.defaults["hvp_method"]
+        if sketch_type in ('normal', 'gaussian'):
+            S = _gaussian_sketch(ndim, sketch_size, device=device, dtype=dtype, generator=generator)
 
-            if sketch_type in ('normal', 'gaussian'):
-                S = _gaussian_sketch(ndim, sketch_size, device=device, dtype=dtype, generator=generator)
+        elif sketch_type == "rademacher":
+            S = _rademacher_sketch(ndim, sketch_size, device=device, dtype=dtype, generator=generator)
 
-            elif sketch_type == 'orthonormal':
-                S = _orthonormal_sketch(ndim, sketch_size, device=device, dtype=dtype, generator=generator)
+        elif sketch_type == 'orthonormal':
+            S = _orthonormal_sketch(ndim, sketch_size, device=device, dtype=dtype, generator=generator)
 
-            elif sketch_type == 'common_directions':
-                # Wang, Po-Wei, Ching-pei Lee, and Chih-Jen Lin. "The common-directions method for regularized empirical risk minimization." Journal of Machine Learning Research 20.58 (2019): 1-49.
-                g_list = var.get_grad(create_graph=hvp_method in ("batched", "autograd"))
-                g = torch.cat([t.ravel() for t in g_list])
+        elif sketch_type == 'common_directions':
+            # Wang, Po-Wei, Ching-pei Lee, and Chih-Jen Lin. "The common-directions method for regularized empirical risk minimization." Journal of Machine Learning Research 20.58 (2019): 1-49.
+            g_list = objective.get_grads(create_graph=hvp_method in ("batched_autograd", "autograd"))
+            g = torch.cat([t.ravel() for t in g_list])
 
-                # initialize directions deque
-                if "directions" not in self.global_state:
+            # initialize directions deque
+            if "directions" not in self.global_state:
 
+                g_norm = torch.linalg.vector_norm(g) # pylint:disable=not-callable
+                if g_norm < torch.finfo(g.dtype).tiny * 2:
+                    g = torch.randn_like(g)
                     g_norm = torch.linalg.vector_norm(g) # pylint:disable=not-callable
-                    if g_norm < torch.finfo(g.dtype).tiny * 2:
-                        g = torch.randn_like(g)
-                        g_norm = torch.linalg.vector_norm(g) # pylint:disable=not-callable
-
-                    self.global_state["directions"] = deque([g / g_norm], maxlen=sketch_size)
-                    S = self.global_state["directions"][0].unsqueeze(1)
-
-                # add new steepest descent direction orthonormal to existing columns
-                else:
-                    S = torch.stack(tuple(self.global_state["directions"]), dim=1)
-                    p = g - S @ (S.T @ g)
-                    p_norm = torch.linalg.vector_norm(p) # pylint:disable=not-callable
-                    if p_norm > torch.finfo(p.dtype).tiny * 2:
-                        p = p / p_norm
-                        self.global_state["directions"].append(p)
-                        S = torch.cat([S, p.unsqueeze(1)], dim=1)
-
-            elif sketch_type == "mixed":
-                g_list = var.get_grad(create_graph=hvp_method in ("batched", "autograd"))
-                g = torch.cat([t.ravel() for t in g_list])
-
-                if "slow_ema" not in self.global_state:
-                    self.global_state["slow_ema"] = torch.randn_like(g) * 1e-2
-                    self.global_state["fast_ema"] = torch.randn_like(g) * 1e-2
-
-                slow_ema = self.global_state["slow_ema"]
-                fast_ema = self.global_state["fast_ema"]
-                slow_ema.lerp_(g, 0.001)
-                fast_ema.lerp_(g, 0.1)
-
-                S = torch.stack([g, slow_ema, fast_ema], dim=1)
-                if sketch_size > 3:
-                    S_random = _gaussian_sketch(ndim, sketch_size - 3, device=device, dtype=dtype, generator=generator)
-                    S = torch.cat([S, S_random], dim=1)
-
-                S = _qr_orthonormalize(S)
-
-            else:
-                raise ValueError(f'Unknown sketch_type {sketch_type}')
-
-            # form sketched hessian
-            HS, _ = var.hessian_matrix_product(S, at_x0=True, rgrad=None, hvp_method=self.defaults["hvp_method"], normalize=True, retain_graph=False, h=self.defaults["h"])
-            H_sketched = S.T @ HS
 
-            self.global_state["H_sketched"] = H_sketched
-            self.global_state["S"] = S
+                self.global_state["directions"] = deque([g / g_norm], maxlen=sketch_size)
+                S = self.global_state["directions"][0].unsqueeze(1)
 
-    def apply(self, var):
+            # add new steepest descent direction orthonormal to existing columns
+            else:
+                S = torch.stack(tuple(self.global_state["directions"]), dim=1)
+                p = g - S @ (S.T @ g)
+                p_norm = torch.linalg.vector_norm(p) # pylint:disable=not-callable
+                if p_norm > torch.finfo(p.dtype).tiny * 2:
+                    p = p / p_norm
+                    self.global_state["directions"].append(p)
+                    S = torch.cat([S, p.unsqueeze(1)], dim=1)
+
+        elif sketch_type == "mixed":
+            g_list = objective.get_grads(create_graph=hvp_method in ("batched_autograd", "autograd"))
+            g = torch.cat([t.ravel() for t in g_list])
+
+            # initialize state
+            if "slow_ema" not in self.global_state:
+                self.global_state["slow_ema"] = torch.randn_like(g) * 1e-2
+                self.global_state["fast_ema"] = torch.randn_like(g) * 1e-2
+                self.global_state["p_prev"] = torch.randn_like(g)
+
+            # previous update direction
+            p_cur = torch.cat([t.ravel() for t in params])
+            prev_dir = p_cur - self.global_state["p_prev"]
+            self.global_state["p_prev"] = p_cur
+
+            # EMAs
+            slow_ema = self.global_state["slow_ema"]
+            fast_ema = self.global_state["fast_ema"]
+            slow_ema.lerp_(g, 0.001)
+            fast_ema.lerp_(g, 0.1)
+
+            # form and orthogonalize sketching matrix
+            S = torch.stack([g, slow_ema, fast_ema, prev_dir], dim=1)
+            if sketch_size > 4:
+                S_random = _gaussian_sketch(ndim, sketch_size - 3, device=device, dtype=dtype, generator=generator)
+                S = torch.cat([S, S_random], dim=1)
+
+            S = _qr_orthonormalize(S)
+
+        else:
+            raise ValueError(f'Unknown sketch_type {sketch_type}')
+
+        # form sketched hessian
+        HS, _ = objective.hessian_matrix_product(S, rgrad=None, at_x0=True,
+                                                 hvp_method=fs["hvp_method"], h=fs["h"])
+        H_sketched = S.T @ HS
+
+        self.global_state["H_sketched"] = H_sketched
+        self.global_state["S"] = S
+
+    def apply_states(self, objective, states, settings):
         S: torch.Tensor = self.global_state["S"]
+
         d_proj = _newton_step(
-            var=var,
+            objective=objective,
             H=self.global_state["H_sketched"],
             damping=self.defaults["damping"],
-            inner=self.children.get("inner", None),
             H_tfm=self.defaults["H_tfm"],
             eigval_fn=self.defaults["eigval_fn"],
             use_lstsq=self.defaults["use_lstsq"],
             g_proj = lambda g: S.T @ g
         )
-        d = S @ d_proj
-        var.update = vec_to_tensors(d, var.params)
 
-        return var
+        d = S @ d_proj
+        objective.updates = vec_to_tensors(d, objective.params)
+        return objective
 
-    def get_H(self, var=...):
+    def get_H(self, objective=...):
         eigval_fn = self.defaults["eigval_fn"]
         H_sketched: torch.Tensor = self.global_state["H_sketched"]
         S: torch.Tensor = self.global_state["S"]

torchzero/modules/smoothing/laplacian.py

@@ -4,7 +4,7 @@ from collections.abc import Iterable
 import torch
 
 from ...utils.tensorlist import TensorList
-from ...core import Transform, Target
+from ...core import TensorTransform
 
 
 def vector_laplacian_smoothing(input: torch.Tensor, sigma: float = 1) -> torch.Tensor:
@@ -55,7 +55,7 @@ def _precompute_denominator(tensor: torch.Tensor, sigma) -> torch.Tensor:
     v[-1] = 1
     return 1 - sigma * torch.fft.fft(v) # pylint: disable = not-callable
 
-class LaplacianSmoothing(Transform):
+class LaplacianSmoothing(TensorTransform):
     """Applies laplacian smoothing via a fast Fourier transform solver which can improve generalization.
 
     Args:
@@ -70,29 +70,30 @@ class LaplacianSmoothing(Transform):
             what to set on var.
 
     Examples:
-        Laplacian Smoothing Gradient Descent optimizer as in the paper
+    Laplacian Smoothing Gradient Descent optimizer as in the paper
 
-        .. code-block:: python
+    ```python
 
-            opt = tz.Modular(
-                model.parameters(),
-                tz.m.LaplacianSmoothing(),
-                tz.m.LR(1e-2),
-            )
+    opt = tz.Modular(
+        model.parameters(),
+        tz.m.LaplacianSmoothing(),
+        tz.m.LR(1e-2),
+    )
+    ```
 
     Reference:
         Osher, S., Wang, B., Yin, P., Luo, X., Barekat, F., Pham, M., & Lin, A. (2022). Laplacian smoothing gradient descent. Research in the Mathematical Sciences, 9(3), 55.
 
     """
-    def __init__(self, sigma:float = 1, layerwise=True, min_numel = 4, target: Target = 'update'):
+    def __init__(self, sigma:float = 1, layerwise=True, min_numel = 4):
         defaults = dict(sigma = sigma, layerwise=layerwise, min_numel=min_numel)
-        super().__init__(defaults, uses_grad=False, target=target)
+        super().__init__(defaults)
         # precomputed denominator for when layerwise=False
         self.global_state['full_denominator'] = None
 
 
     @torch.no_grad
-    def apply_tensors(self, tensors, params, grads, loss, states, settings):
+    def multi_tensor_apply(self, tensors, params, grads, loss, states, settings):
         layerwise = settings[0]['layerwise']
 
         # layerwise laplacian smoothing

torchzero/modules/smoothing/sampling.py

@@ -7,14 +7,15 @@ from typing import Literal, cast
 
 import torch
 
-from ...core import Chainable, Modular, Module, Var
+from ...core import Chainable, Modular, Module, Objective
 from ...core.reformulation import Reformulation
 from ...utils import Distributions, NumberList, TensorList
 from ..termination import TerminationCriteriaBase, make_termination_criteria
 
 
-def _reset_except_self(optimizer: Modular, var: Var, self: Module):
-    for m in optimizer.unrolled_modules:
+def _reset_except_self(objective: Objective, modules, self: Module):
+    assert objective.modular is not None
+    for m in objective.modular.flat_modules:
         if m is not self:
             m.reset()
 
@@ -98,15 +99,15 @@ class GradientSampling(Reformulation):
             self.set_child('termination', make_termination_criteria(extra=termination))
 
     @torch.no_grad
-    def pre_step(self, var):
-        params = TensorList(var.params)
+    def pre_step(self, objective):
+        params = TensorList(objective.params)
 
         fixed = self.defaults['fixed']
 
         # check termination criteria
         if 'termination' in self.children:
             termination = cast(TerminationCriteriaBase, self.children['termination'])
-            if termination.should_terminate(var):
+            if termination.should_terminate(objective):
 
                 # decay sigmas
                 states = [self.state[p] for p in params]
@@ -118,7 +119,7 @@ class GradientSampling(Reformulation):
 
                 # reset on sigmas decay
                 if self.defaults['reset_on_termination']:
-                    var.post_step_hooks.append(partial(_reset_except_self, self=self))
+                    objective.post_step_hooks.append(partial(_reset_except_self, self=self))
 
                 # clear perturbations
                 self.global_state.pop('perts', None)
@@ -136,7 +137,7 @@ class GradientSampling(Reformulation):
             self.global_state['perts'] = perts
 
     @torch.no_grad
-    def closure(self, backward, closure, params, var):
+    def closure(self, backward, closure, params, objective):
         params = TensorList(params)
         loss_agg = None
         grad_agg = None
@@ -160,7 +161,7 @@ class GradientSampling(Reformulation):
 
         # evaluate at x_0
         if include_x0:
-            f_0 = cast(torch.Tensor, var.get_loss(backward=backward))
+            f_0 = objective.get_loss(backward=backward)
 
             isfinite = math.isfinite(f_0)
             if isfinite:
@@ -168,7 +169,7 @@ class GradientSampling(Reformulation):
                 loss_agg = f_0
 
             if backward:
-                g_0 = var.get_grad()
+                g_0 = objective.get_grads()
                 if isfinite: grad_agg = g_0
 
         # evaluate at x_0 + p for each perturbation

torchzero/modules/step_size/adaptive.py

@@ -5,9 +5,9 @@ from typing import Any, Literal
 
 import torch
 
-from ...core import Chainable, Transform
+from ...core import Chainable, TensorTransform
 from ...utils import NumberList, TensorList, tofloat, unpack_dicts, unpack_states
-from ...utils.linalg.linear_operator import ScaledIdentity
+from ...linalg.linear_operator import ScaledIdentity
 from ..functional import epsilon_step_size
 
 def _acceptable_alpha(alpha, param:torch.Tensor):
@@ -16,7 +16,7 @@ def _acceptable_alpha(alpha, param:torch.Tensor):
         return False
     return True
 
-def _get_H(self: Transform, var):
+def _get_H(self: TensorTransform, var):
     n = sum(p.numel() for p in var.params)
     p = var.params[0]
     alpha = self.global_state.get('alpha', 1)
@@ -25,7 +25,7 @@ def _get_H(self: Transform, var):
     return ScaledIdentity(1 / alpha, shape=(n,n), device=p.device, dtype=p.dtype)
 
 
-class PolyakStepSize(Transform):
+class PolyakStepSize(TensorTransform):
     """Polyak's subgradient method with known or unknown f*.
 
     Args:
@@ -47,7 +47,7 @@ class PolyakStepSize(Transform):
         super().__init__(defaults, uses_grad=use_grad, uses_loss=True, inner=inner)
 
     @torch.no_grad
-    def update_tensors(self, tensors, params, grads, loss, states, settings):
+    def multi_tensor_update(self, tensors, params, grads, loss, states, settings):
         assert grads is not None and loss is not None
         tensors = TensorList(tensors)
         grads = TensorList(grads)
@@ -79,15 +79,15 @@ class PolyakStepSize(Transform):
         self.global_state['alpha'] = alpha
 
     @torch.no_grad
-    def apply_tensors(self, tensors, params, grads, loss, states, settings):
+    def multi_tensor_apply(self, tensors, params, grads, loss, states, settings):
         alpha = self.global_state.get('alpha', 1)
         if not _acceptable_alpha(alpha, tensors[0]): alpha = epsilon_step_size(TensorList(tensors))
 
         torch._foreach_mul_(tensors, alpha * unpack_dicts(settings, 'alpha', cls=NumberList))
         return tensors
 
-    def get_H(self, var):
-        return _get_H(self, var)
+    def get_H(self, objective):
+        return _get_H(self, objective)
 
 
 def _bb_short(s: TensorList, y: TensorList, sy, eps):
@@ -116,7 +116,7 @@ def _bb_geom(s: TensorList, y: TensorList, sy, eps, fallback:bool):
         return None
     return (short * long) ** 0.5
 
-class BarzilaiBorwein(Transform):
+class BarzilaiBorwein(TensorTransform):
     """Barzilai-Borwein step size method.
 
     Args:
@@ -144,7 +144,7 @@ class BarzilaiBorwein(Transform):
         self.global_state['reset'] = True
 
     @torch.no_grad
-    def update_tensors(self, tensors, params, grads, loss, states, settings):
+    def multi_tensor_update(self, tensors, params, grads, loss, states, settings):
         step = self.global_state.get('step', 0)
         self.global_state['step'] = step + 1
 
@@ -175,11 +175,11 @@ class BarzilaiBorwein(Transform):
         prev_p.copy_(params)
         prev_g.copy_(g)
 
-    def get_H(self, var):
-        return _get_H(self, var)
+    def get_H(self, objective):
+        return _get_H(self, objective)
 
     @torch.no_grad
-    def apply_tensors(self, tensors, params, grads, loss, states, settings):
+    def multi_tensor_apply(self, tensors, params, grads, loss, states, settings):
         alpha = self.global_state.get('alpha', None)
 
         if not _acceptable_alpha(alpha, tensors[0]):
@@ -189,7 +189,7 @@ class BarzilaiBorwein(Transform):
         return tensors
 
 
-class BBStab(Transform):
+class BBStab(TensorTransform):
     """Stabilized Barzilai-Borwein method (https://arxiv.org/abs/1907.06409).
 
     This clips the norm of the Barzilai-Borwein update by ``delta``, where ``delta`` can be adaptive if ``c`` is specified.
@@ -228,7 +228,7 @@ class BBStab(Transform):
         self.global_state['reset'] = True
 
     @torch.no_grad
-    def update_tensors(self, tensors, params, grads, loss, states, settings):
+    def multi_tensor_update(self, tensors, params, grads, loss, states, settings):
         step = self.global_state.get('step', 0)
         self.global_state['step'] = step + 1
 
@@ -287,11 +287,11 @@ class BBStab(Transform):
         prev_p.copy_(params)
         prev_g.copy_(g)
 
-    def get_H(self, var):
-        return _get_H(self, var)
+    def get_H(self, objective):
+        return _get_H(self, objective)
 
     @torch.no_grad
-    def apply_tensors(self, tensors, params, grads, loss, states, settings):
+    def multi_tensor_apply(self, tensors, params, grads, loss, states, settings):
         alpha = self.global_state.get('alpha', None)
 
         if not _acceptable_alpha(alpha, tensors[0]):
@@ -301,7 +301,7 @@ class BBStab(Transform):
         return tensors
 
 
-class AdGD(Transform):
+class AdGD(TensorTransform):
     """AdGD and AdGD-2 (https://arxiv.org/abs/2308.02261)"""
     def __init__(self, variant:Literal[1,2]=2, alpha_0:float = 1e-7, sqrt:bool=True, use_grad=True, inner: Chainable | None = None,):
         defaults = dict(variant=variant, alpha_0=alpha_0, sqrt=sqrt)
@@ -313,7 +313,7 @@ class AdGD(Transform):
         self.global_state['reset'] = True
 
     @torch.no_grad
-    def update_tensors(self, tensors, params, grads, loss, states, settings):
+    def multi_tensor_update(self, tensors, params, grads, loss, states, settings):
         variant = settings[0]['variant']
         theta_0 = 0 if variant == 1 else 1/3
         theta = self.global_state.get('theta', theta_0)
@@ -371,7 +371,7 @@ class AdGD(Transform):
         prev_g.copy_(g)
 
     @torch.no_grad
-    def apply_tensors(self, tensors, params, grads, loss, states, settings):
+    def multi_tensor_apply(self, tensors, params, grads, loss, states, settings):
         alpha = self.global_state.get('alpha', None)
 
         if not _acceptable_alpha(alpha, tensors[0]):
@@ -383,5 +383,5 @@ class AdGD(Transform):
         torch._foreach_mul_(tensors, alpha)
         return tensors
 
-    def get_H(self, var):
-        return _get_H(self, var)
+    def get_H(self, objective):
+        return _get_H(self, objective)

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.
@@ -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:
@@ -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']