torchzero 0.4.1__py3-none-any.whl → 0.4.2__py3-none-any.whl

torchzero/modules/second_order/newton.py

@@ -7,6 +7,7 @@ from ...core import Chainable, Transform, Objective, HessianMethod
 from ...utils import vec_to_tensors_
 from ...linalg.linear_operator import Dense, DenseWithInverse, Eigendecomposition
 from ...linalg import torch_linalg
+from ...linalg.eigh import regularize_eigh
 
 def _try_lu_solve(H: torch.Tensor, g: torch.Tensor):
     try:
@@ -30,6 +31,8 @@ def _newton_update_state_(
     H: torch.Tensor,
     damping: float,
     eigval_fn: Callable | None,
+    eigv_tol: float | None,
+    truncate: int | None,
     precompute_inverse: bool,
     use_lstsq: bool,
 ):
@@ -39,10 +42,11 @@ def _newton_update_state_(
         reg = torch.eye(H.size(0), device=H.device, dtype=H.dtype).mul_(damping)
         H += reg
 
-    # if eigval_fn is given, we don't need H or H_inv, we store factors
-    if eigval_fn is not None:
+    # if any args require eigendecomp, we don't need H or H_inv, we store factors
+    if any(i is not None for i in [eigval_fn, eigv_tol, truncate]):
         L, Q = torch_linalg.eigh(H, retry_float64=True)
-        L = eigval_fn(L)
+        if eigval_fn is not None: L = eigval_fn(L)
+        L, Q = regularize_eigh(L, Q, truncate=truncate, tol=eigv_tol)
         state["L"] = L
         state["Q"] = Q
         return
@@ -216,6 +220,8 @@ class Newton(Transform):
         self,
         damping: float = 0,
         eigval_fn: Callable[[torch.Tensor], torch.Tensor] | None = None,
+        eigv_tol: float | None = None,
+        truncate: int | None = None,
         update_freq: int = 1,
         precompute_inverse: bool | None = None,
         use_lstsq: bool = False,
@@ -242,6 +248,8 @@ class Newton(Transform):
             H=H,
             damping = fs["damping"],
             eigval_fn = fs["eigval_fn"],
+            eigv_tol = fs["eigv_tol"],
+            truncate = fs["truncate"],
             precompute_inverse = precompute_inverse,
             use_lstsq = fs["use_lstsq"]
         )

torchzero/modules/second_order/newton_cg.py

@@ -226,7 +226,8 @@ class NewtonCGSteihaug(Transform):
         tol: float = 1e-8,
         reg: float = 1e-8,
         solver: Literal['cg', "minres"] = 'cg',
-        adapt_tol: bool = True,
+        adapt_tol: bool = False,
+        terminate_on_tr: bool = True,
         npc_terminate: bool = False,
 
         # hvp settings
@@ -272,7 +273,6 @@ class NewtonCGSteihaug(Transform):
         npc_terminate=fs["npc_terminate"]
         miniter=fs["miniter"]
         max_history=fs["max_history"]
-        adapt_tol=fs["adapt_tol"]
 
 
         # ------------------------------- trust region ------------------------------- #
@@ -294,9 +294,13 @@ class NewtonCGSteihaug(Transform):
             finfo = torch.finfo(orig_params[0].dtype)
             if trust_radius < finfo.tiny * 2:
                 trust_radius = self.global_state['trust_radius'] = init
-                if adapt_tol:
+
+                if fs["adapt_tol"]:
                     self.global_state["tol_mul"] = self.global_state.get("tol_mul", 1) * 0.1
 
+                if fs["terminate_on_tr"]:
+                    objective.should_terminate = True
+
             elif trust_radius > finfo.max / 2:
                 trust_radius = self.global_state['trust_radius'] = init
 

torchzero/modules/second_order/nystrom.py

@@ -5,7 +5,7 @@ import torch
 
 from ...core import Chainable, Transform, HVPMethod
 from ...utils import TensorList, vec_to_tensors
-from ...linalg import nystrom_pcg, nystrom_sketch_and_solve, nystrom_approximation, cg, regularize_eigh, OrthogonalizeMethod
+from ...linalg import nystrom_pcg, nystrom_sketch_and_solve, nystrom_approximation, cg, regularize_eigh, OrthogonalizeMethod, orthogonalize
 from ...linalg.linear_operator import Eigendecomposition, ScaledIdentity
 
 class NystromSketchAndSolve(Transform):
@@ -75,7 +75,7 @@ class NystromSketchAndSolve(Transform):
     """
     def __init__(
         self,
-        rank: int,
+        rank: int = 100,
         reg: float | None = 1e-2,
         eigv_tol: float = 0,
         truncate: int | None = None,
@@ -109,17 +109,15 @@ class NystromSketchAndSolve(Transform):
 
         generator = self.get_generator(params[0].device, seed=fs['seed'])
         try:
+            Omega = torch.randn([ndim, min(fs["rank"], ndim)], device=device, dtype=dtype, generator=generator)
+            Omega = orthogonalize(Omega, fs["orthogonalize_method"])
+            HOmega = H_mm(Omega)
+
             # compute the approximation
             L, Q = nystrom_approximation(
-                A_mv=H_mv,
-                A_mm=H_mm,
-                ndim=ndim,
-                rank=min(fs["rank"], ndim),
+                Omega=Omega,
+                AOmega=HOmega,
                 eigv_tol=fs["eigv_tol"],
-                orthogonalize_method=fs["orthogonalize_method"],
-                dtype=dtype,
-                device=device,
-                generator=generator,
             )
 
             # regularize
@@ -225,7 +223,7 @@ class NystromPCG(Transform):
     """
     def __init__(
         self,
-        rank: int,
+        rank: int = 100,
         maxiter=None,
         tol=1e-8,
         reg: float = 1e-6,
@@ -260,16 +258,13 @@ class NystromPCG(Transform):
             generator = self.get_generator(device, seed=fs['seed'])
 
             try:
+                Omega = torch.randn(ndim, min(fs["rank"], ndim), device=device, dtype=dtype, generator=generator)
+                HOmega = H_mm(orthogonalize(Omega, fs["orthogonalize_method"]))
+                # compute the approximation
                 L, Q = nystrom_approximation(
-                    A_mv=None,
-                    A_mm=H_mm,
-                    ndim=ndim,
-                    rank=min(fs["rank"], ndim),
+                    Omega=Omega,
+                    AOmega=HOmega,
                     eigv_tol=fs["eigv_tol"],
-                    orthogonalize_method=fs["orthogonalize_method"],
-                    dtype=dtype,
-                    device=device,
-                    generator=generator,
                 )
 
                 self.global_state["L"] = L

torchzero/modules/second_order/rsn.py

@@ -25,9 +25,23 @@ def _orthonormal_sketch(m, n, dtype, device, generator):
     return _qr_orthonormalize(torch.randn(m, n, dtype=dtype, device=device, generator=generator))
 
 def _rademacher_sketch(m, n, dtype, device, generator):
-    rademacher = torch.bernoulli(torch.full((m,n), 0.5), generator = generator).mul_(2).sub_(1)
+    rademacher = torch.bernoulli(torch.full((m,n), 0.5, device=device, dtype=dtype), generator = generator).mul_(2).sub_(1)
     return rademacher.mul_(1 / math.sqrt(m))
 
+def _row_sketch(m, n, dtype, device, generator):
+    weights = torch.ones(m, dtype=dtype, device=device)
+    indices = torch.multinomial(weights, n, replacement=False, generator=generator)
+
+    P = torch.zeros(m, n, dtype=dtype, device=device)
+    P[indices, range(n)] = 1
+    return P
+
+def _topk_rows(grad, m, n, dtype, device, generator):
+    _, indices = torch.topk(grad.abs(), n)
+    P = torch.zeros(m, n, dtype=dtype, device=device)
+    P[indices, range(n)] = 1
+    return P
+
 class SubspaceNewton(Transform):
     """Subspace Newton. Performs a Newton step in a subspace (random or spanned by past gradients).
 
@@ -37,7 +51,9 @@ class SubspaceNewton(Transform):
         sketch_type (str, optional):
             - "common_directions" - uses history steepest descent directions as the basis[2]. It is orthonormalized on-line using Gram-Schmidt (default).
             - "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 (if dimension is large) scaled random rademacher basis. It is recommended to use at least "orthonormal" - it requires QR but it is still very cheap.
+            - "rows" - samples random rows.
+            - "topk" - samples top-rank rows with largest gradient magnitude.
+            - "rademacher" - approximately orthonormal (if dimension is large) scaled random rademacher basis.
             - "mixed" - random orthonormal basis but with four directions set to gradient, slow and fast gradient EMAs, and previous update direction.
         damping (float, optional): hessian damping (scale of identity matrix added to hessian). Defaults to 0.
         hvp_method (str, optional):
@@ -93,13 +109,15 @@ class SubspaceNewton(Transform):
 
     def __init__(
         self,
-        sketch_size: int,
-        sketch_type: Literal["orthonormal", "common_directions", "mixed", "rademacher"] = "common_directions",
+        sketch_size: int = 100,
+        sketch_type: Literal["orthonormal", "common_directions", "mixed", "rademacher", "rows", "topk"] = "common_directions",
         damping:float=0,
         eigval_fn: Callable[[torch.Tensor], torch.Tensor] | None = None,
+        eigv_tol: float | None = None,
+        truncate: int | None = None,
         update_freq: int = 1,
         precompute_inverse: bool = False,
-        use_lstsq: bool = True,
+        use_lstsq: bool = False,
         hvp_method: HVPMethod = "batched_autograd",
         h: float = 1e-2,
         seed: int | None = None,
@@ -131,6 +149,14 @@ class SubspaceNewton(Transform):
         elif sketch_type == 'orthonormal':
             S = _orthonormal_sketch(ndim, sketch_size, device=device, dtype=dtype, generator=generator)
 
+        elif sketch_type == "rows":
+            S = _row_sketch(ndim, sketch_size, device=device, dtype=dtype, generator=generator)
+
+        elif sketch_type == "topk":
+            g_list = objective.get_grads(create_graph=hvp_method in ("batched_autograd", "autograd"))
+            g = torch.cat([t.ravel() for t in g_list])
+            S = _topk_rows(g, 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 = objective.get_grads(create_graph=hvp_method in ("batched_autograd", "autograd"))
@@ -189,6 +215,10 @@ class SubspaceNewton(Transform):
         else:
             raise ValueError(f'Unknown sketch_type {sketch_type}')
 
+        # print(f'{S.shape = }')
+        # I = torch.eye(S.size(1), device=S.device, dtype=S.dtype)
+        # print(f'{torch.nn.functional.mse_loss(S.T @ S, I) = }')
+
         # form sketched hessian
         HS, _ = objective.hessian_matrix_product(S, rgrad=None, at_x0=True,
                                                  hvp_method=fs["hvp_method"], h=fs["h"])
@@ -200,9 +230,10 @@ class SubspaceNewton(Transform):
             H = H_sketched,
             damping = fs["damping"],
             eigval_fn = fs["eigval_fn"],
+            eigv_tol = fs["eigv_tol"],
+            truncate = fs["truncate"],
             precompute_inverse = fs["precompute_inverse"],
             use_lstsq = fs["use_lstsq"]
-
         )
 
         self.global_state["S"] = S

torchzero/modules/trust_region/trust_region.py

@@ -1,7 +1,7 @@
 import math
 import warnings
 from abc import ABC, abstractmethod
-from collections.abc import Callable, Mapping, Sequence
+from collections.abc import Callable, Mapping, Sequence, MutableMapping
 from functools import partial
 from typing import Any, Literal, Protocol, cast, final, overload
 
@@ -203,6 +203,7 @@ def fixed_radius(
 ) -> tuple[float, bool]:
     return init, True
 
+
 _RADIUS_KEYS = Literal['default', 'fixed']
 _RADIUS_STRATEGIES: dict[_RADIUS_KEYS, _RadiusStrategy] = {
     "default": default_radius,

torchzero/utils/benchmarks/logistic.py

@@ -5,39 +5,54 @@ import numpy as np
 import torch
 import tqdm
 
-
-def generate_correlated_logistic_data(n_samples=2000, n_features=32, n_correlated_pairs=512, correlation=0.99, seed=0):
-    """Hard logistic regression dataset with correlated features"""
+def generate_correlated_logistic_data(
+    n_samples=100_000,
+    n_features=32,
+    n_classes=10,
+    n_correlated=768,
+    correlation=0.99,
+    seed=0
+) -> tuple[np.ndarray, np.ndarray]:
+    assert n_classes >= 2
     generator = np.random.default_rng(seed)
 
-    # ------------------------------------- X ------------------------------------ #
     X = generator.standard_normal(size=(n_samples, n_features))
-    weights = generator.uniform(-2, 2, n_features)
+    weights = generator.uniform(-2, 2, size=(n_features, n_classes))
+
+    used_pairs = set()
+    n_correlated = min(n_correlated, n_features * (n_features - 1) // 2)
 
-    used_pairs = []
-    for i in range(n_correlated_pairs):
+    for _ in range(n_correlated):
         idxs = None
         while idxs is None or idxs in used_pairs:
-            idxs = tuple(generator.choice(n_features, size=2, replace=False).tolist())
+            pair = generator.choice(n_features, size=2, replace=False)
+            pair.sort()
+            idxs = tuple(pair)
 
-        used_pairs.append(idxs)
+        used_pairs.add(idxs)
         idx1, idx2 = idxs
 
         noise = generator.standard_normal(n_samples) * np.sqrt(1 - correlation**2)
         X[:, idx2] = correlation * X[:, idx1] + noise
 
         w = generator.integers(1, 51)
-        weights[idx1] = w
-        weights[idx2] = -w
+        cls = generator.integers(0, n_classes)
+        weights[idx1, cls] = w
+        weights[idx2, cls] = -w
 
-    # ---------------------------------- logits ---------------------------------- #
     logits = X @ weights
-    probabilities = 1 / (1 + np.exp(-logits))
-    y = generator.binomial(1, probabilities).astype(np.float32)
 
-    X = X - X.mean(0, keepdims=True)
-    X = X / X.std(0, keepdims=True)
-    return X, y
+    logits -= logits.max(axis=1, keepdims=True)
+    exp_logits = np.exp(logits)
+    probabilities = exp_logits / exp_logits.sum(axis=1, keepdims=True)
+
+    y_one_hot = generator.multinomial(1, pvals=probabilities)
+    y = np.argmax(y_one_hot, axis=1)
+
+    X -= X.mean(0, keepdims=True)
+    X /= X.std(0, keepdims=True)
+
+    return X, y.astype(np.int64)
 
 
 # if __name__ == '__main__':
@@ -101,7 +116,7 @@ def run_logistic_regression(X: torch.Tensor, y: torch.Tensor, opt_fn, max_steps:
         # this is for tests
         if _assert_on_evaluated_same_params:
             for p in evaluated_params:
-                assert not _tensorlist_equal(p, model.parameters()), f"evaluated same parameters on epoch {epoch}"
+                assert not _tensorlist_equal(p, model.parameters()), f"{optimizer} evaluated same parameters on epoch {epoch}"
 
             evaluated_params.append([p.clone() for p in model.parameters()])
 

torchzero/utils/params.py

@@ -3,7 +3,7 @@ from collections.abc import Sequence, Iterable, Mapping
 import warnings
 import torch, numpy as np
 
-
+from .torch_tools import set_storage_
 
 Params = Iterable[torch.Tensor | tuple[str, torch.Tensor] | Mapping[str, Any]]
 
@@ -147,3 +147,15 @@ def _set_update_and_grad_(
 
     return param_groups
 
+
+def _set_fake_params_(fake_params: Iterable[torch.Tensor], storage: Iterable[torch.Tensor]):
+    """sets ``fake_params`` storage to ``storage`` while they remain the same python object"""
+    for fake_p, s in zip(fake_params, storage):
+        fake_p.set_(s.view_as(s).requires_grad_()) # pyright: ignore[reportArgumentType]
+
+def _empty_fake_param_storage_(fake_params: Iterable[torch.Tensor]):
+    """sets ``fake_params`` storage to empty while they remain the same python object"""
+    for p in fake_params:
+        set_storage_(p, torch.empty(0, device=p.device, dtype=p.dtype))
+
+

torchzero/utils/tensorlist.py

@@ -330,10 +330,10 @@ class TensorList(list[torch.Tensor | Any]):
 
     def global_vector_norm(self, ord:float = 2) -> torch.Tensor:
         # return torch.linalg.vector_norm(self.to_vec(), ord = ord) # pylint:disable = not-callable
-        if ord == 1: return self.global_sum()
-        if ord % 2 == 0: return self.pow(ord).global_sum().pow(1/ord)
         if ord == torch.inf: return self.abs().global_max()
         if ord == -torch.inf: return self.abs().global_min()
+        if ord == 1: return self.abs().global_sum()
+        if ord % 2 == 0: return self.pow(ord).global_sum().pow(1/ord)
         if ord == 0: return (self != 0).global_sum().to(self[0].dtype)
 
         return self.abs().pow_(ord).global_sum().pow(1/ord)

{torchzero-0.4.1.dist-info → torchzero-0.4.2.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: torchzero
-Version: 0.4.1
+Version: 0.4.2
 Summary: Modular optimization library for PyTorch.
 Author-email: Ivan Nikishev <nkshv2@gmail.com>
 Project-URL: Homepage, https://github.com/inikishev/torchzero

{torchzero-0.4.1.dist-info → torchzero-0.4.2.dist-info}/RECORD

@@ -5,50 +5,54 @@ tests/test_objective.py,sha256=HY0rK0z6PpiXvEsCu4mLgTlSVKusnT69S2GbuVcwMRo,7119
 tests/test_opts.py,sha256=hw7CCw7FD_RJSdiSacyXUSM7DI-_RfP8wJlsz079SNw,44263
 tests/test_tensorlist.py,sha256=B0Tq4_r-1DOYpS360X7IsLQiWn5fukhIMDKZM6zVO2Y,72164
 tests/test_utils_optimizer.py,sha256=_JoMqvXXZ6TxugS_CmfmP55Vvp0XrSPCjSz2nJJmaoI,8399
-torchzero/__init__.py,sha256=nit4KxrRoW6hJDGOy0jkphuawY5gAvPqrYY11Yct6fA,133
-torchzero/core/__init__.py,sha256=h9Ck7XX2XuJUTojU2IMa_2TprXZHbgo748txa3z7-2o,341
-torchzero/core/chain.py,sha256=dtFpxnw8vcbi3EeAANXyPtUmyPyv_VuZrTiPlLRmh7c,1899
-torchzero/core/functional.py,sha256=TSygtyQHDhqf998--hF48yIFr-y3Ycz8arjjR8x1ILU,3156
+torchzero/__init__.py,sha256=SZLJgf_sjHyqtTzz0f70AtHP_V_WloX1KQF8mm34zdg,175
+torchzero/_minimize/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
+torchzero/_minimize/methods.py,sha256=1oktoSdWtiA0JEF34yTkY3_nPB5Q5ODHl18C0mcglNw,2445
+torchzero/_minimize/minimize.py,sha256=JJBmREQvhDxyqGM62xharsuebyefxRADkd6Bg_TE-DQ,17236
+torchzero/core/__init__.py,sha256=lufcll5r98gTjVfQSvz6-wfI0qMAgZtLLSByHuHTats,358
+torchzero/core/chain.py,sha256=-6vW-L5pzg2Rwpq3LKIAoqJGPvCkHKjt_B1boGikQmM,1900
+torchzero/core/functional.py,sha256=D125Hso8fHMSKlyhkir3GGJzXxuIitXmVhKn2Y9x-Ck,3272
 torchzero/core/modular.py,sha256=Xpp6jfiKArC3Q42G63I9qj3eWcYt-l7d-EIm-59ADcI,9584
-torchzero/core/module.py,sha256=HfbPfxXxgyBf9wQl7Fpw6B6Ux6UYfvPEmITC64ozb_Q,18012
+torchzero/core/module.py,sha256=DKGLwLWm9LkOBYZHW9QBoXo9eBgnYz7nmoCXJ0gl0e0,21210
 torchzero/core/objective.py,sha256=kEIlry7Bxf_zDUoqAIKUTRvvJmCEpn0Ad2crNt18GCc,40005
 torchzero/core/reformulation.py,sha256=UyAS_xq5sy_mMpmkvtwpHrZHd6Y2RgyPwN0zZlyxFTI,3857
-torchzero/core/transform.py,sha256=aJRBtvYjKqD-Ic_AkzeSINYDsTaBAErA-kocEl3PHZw,12244
+torchzero/core/transform.py,sha256=WlHoc5cCY1vXQrwMsIG0g3Kle93kBSbrBfxGz5X9_0Q,12251
 torchzero/linalg/__init__.py,sha256=wlry3dbncdsySKk6sSdiRefTcc8dIh4DcA0wFyU1MC8,407
 torchzero/linalg/benchmark.py,sha256=wiIMn-GY2xxWbHVf8CPbJddUPeUPq9OUDkvbp1iILYI,479
-torchzero/linalg/eigh.py,sha256=YC8x5NEWWsnc3suCebnTfeb4lVMhy-H8LGOZbGnwd8A,7902
+torchzero/linalg/eigh.py,sha256=l1fX_7hL-DFk8gu20-NuSKDJcRpz58KxUKQHeBhCcHE,9035
 torchzero/linalg/linalg_utils.py,sha256=1RBpQevb25_hbRieONw6CgnoWOJFXXv8oWTMugEivEk,385
-torchzero/linalg/linear_operator.py,sha256=mVEOvu6yY7TYhUdmZm1IAc6_pWnTaykKDgZu_-J-atk,16653
+torchzero/linalg/linear_operator.py,sha256=MWTY7DS8B8IkR28kVA9nmoM-OU-1eBsP22iYXkDrj9A,16654
 torchzero/linalg/matrix_power.py,sha256=gEWGvh3atc7745dwNcxNg0RtUrVgeKD6KxyRckKkkdQ,1255
-torchzero/linalg/orthogonalize.py,sha256=Fv6zv1JvS9AVwjiMVed55J8-pEbVZv7vqoEo5g0Zrv0,3270
+torchzero/linalg/orthogonalize.py,sha256=GSvDZA9evTpu3obqCkEocgpDp_91sRexoAwH2q0zTEY,5345
 torchzero/linalg/qr.py,sha256=KykXhSlye0vhyP5JjX6pkPnheHKLLbAKmDff8Hogxyo,2857
+torchzero/linalg/sketch.py,sha256=dKD9t7I7stv089cCvZyPAOZ0D9wzVG1TmV3297w0tk4,1261
 torchzero/linalg/solve.py,sha256=kING1WCioof8_EKgHeyr53dlft_9KtlJnwOWega3DnA,14355
 torchzero/linalg/svd.py,sha256=jmunSxM-twR5VCUGI_gmV3j7QxMJIe1aBoBlJf5i2fo,1432
 torchzero/linalg/torch_linalg.py,sha256=brhMXUZyYuxuEV-FyQerep1iL7auutW5kmgJpOzUROw,6001
-torchzero/modules/__init__.py,sha256=dsOalCw-OVkD8rhpQdcODc3Hsd_sQ2_2xVC-J8mlSuk,632
+torchzero/modules/__init__.py,sha256=ZN20E2ES6zDf5DuFbZpuCKFinFc5eGR1h00iYZ_XBGU,652
 torchzero/modules/opt_utils.py,sha256=aj7xqHmeze4izxG9k3L6ziG-K_yj8n8fkFpIv-X8V78,8141
 torchzero/modules/adaptive/__init__.py,sha256=X8w2Dal3k0WpLQN-WolnWBBgUyIiZF5RnqBlN0dcAYw,1081
-torchzero/modules/adaptive/adagrad.py,sha256=hMT-Al-vtD6tzPUpQ79LCNko97D7rJN5ji9JOfBqR3k,12015
+torchzero/modules/adaptive/adagrad.py,sha256=NDwmUZaEk0lWnbgYxN23yTWK5A5dQ9BtoKzRTFSKozY,12131
 torchzero/modules/adaptive/adahessian.py,sha256=ucf8loS_lU9VjCb_M42WwXESjPJ_KFChLGkIMFWXO5o,8734
-torchzero/modules/adaptive/adam.py,sha256=Okm7Sc9fMArQAZ7Ph4Etq68uL-IXKY4YNqHWpTzPoTY,3767
-torchzero/modules/adaptive/adan.py,sha256=965tBUwKy6uDiY2la6fVcGcsvGMs90Zg-ZHPtozJGe4,4110
+torchzero/modules/adaptive/adam.py,sha256=RDHYyIAJdi1Pxny8HOHiCFgvPztNwlJlCtzE_ZE-138,3896
+torchzero/modules/adaptive/adan.py,sha256=tmQHiJ5MNwOGP3fp479goHh0xXlhnzULhHxKcVZOkvM,4219
 torchzero/modules/adaptive/adaptive_heavyball.py,sha256=iDiZqke6z6FOR9mhoHMLMm7jvxjzHIQANTe0FBwNj1Q,2230
 torchzero/modules/adaptive/aegd.py,sha256=WLN6vvbSRhQ1P753M3bx_becSF-3cTbu37nhz3NvdGM,1903
 torchzero/modules/adaptive/esgd.py,sha256=gnah-7zk_fMsn7yIWivqDgnaaSdDFXpxg33ywF6TMZg,6173
-torchzero/modules/adaptive/ggt.py,sha256=eYCeV3GArdLv9WuWeim0V3CHJYl3FVKtrtsGshkqwWg,6608
-torchzero/modules/adaptive/lion.py,sha256=H3aI2qnrMtmkvXcoddzjjxdkoD5cq_QwIkLmd_bVPso,1085
+torchzero/modules/adaptive/ggt.py,sha256=7G0Hh8lWy4o73VYVHcZ1JJyDqeKcXi2Y6Qp3qIHosOY,6512
+torchzero/modules/adaptive/lion.py,sha256=yeKUt3WIITtWx97IQzudgbdai77MCfnL_cu90vRkTmA,1141
 torchzero/modules/adaptive/lre_optimizers.py,sha256=AwWUIwnBrozR2HFYLfJnMCBHAWWMKzkS63xFKstRgc0,9760
-torchzero/modules/adaptive/mars.py,sha256=w-cK-1tFuR74SY01xS5jsg1b9qs3l8eOptGrUyQ2m80,2261
+torchzero/modules/adaptive/mars.py,sha256=WquKzTnCZcxzslcvSBMFJVz_kjuCuAzlesw1bHnKqOg,2325
 torchzero/modules/adaptive/matrix_momentum.py,sha256=YefF2k746ke7qiiabdhCPCUFB1_fRddAfGCyIOwV3Ok,6789
-torchzero/modules/adaptive/msam.py,sha256=nqwjuhBMX2UO-omUIeOcD5ti6PIKfKs-RVCn7ourkKA,6946
+torchzero/modules/adaptive/msam.py,sha256=cHfdNkk3Joy2aENwUZXGf3N0P7zcxYGKuySf699OTfM,7051
 torchzero/modules/adaptive/muon.py,sha256=jQ6jlfM4vVRidGJ7FrLtgPnZeuIfW_zU72o7LvOKqh8,8023
 torchzero/modules/adaptive/natural_gradient.py,sha256=8UzacvvIMbYVVE2q0HQ9DLLHYlm1eu6cAiRsOv5XRzQ,7078
 torchzero/modules/adaptive/orthograd.py,sha256=0u2sfGZJjlJItLX2WRP5fLAD8Wd9SgJzQYAUpARJ64A,1813
-torchzero/modules/adaptive/rmsprop.py,sha256=qWVkRmUQ3dui9yBVYtAEll7OlXZDKNT_m70FakTOrTY,4529
-torchzero/modules/adaptive/rprop.py,sha256=a4_UkWse5u2JFAEIlxQqDBUwvUfxh1kNs2ZIhtccnWE,11540
+torchzero/modules/adaptive/rmsprop.py,sha256=sb709Smpkm8H3vYOsh7BzWni5hAf3nBQevhagyOt7mo,4655
+torchzero/modules/adaptive/rprop.py,sha256=vw-Rufa-gpHgq1gDarmNQexrFr13lPLq_mj3c3pNB_Q,11593
 torchzero/modules/adaptive/sam.py,sha256=CTMCqaH9s5EmKQyj1GpqSeTO1weyfsNWPYFN1xaSm_o,5709
-torchzero/modules/adaptive/shampoo.py,sha256=C_Mo7UFQtDxW4McWJjT731FNAp3g9MqF0Hka54Yi3xQ,9847
-torchzero/modules/adaptive/soap.py,sha256=hz2N6-jUSWU93RNViIS1c-Ue2uKmQx6BxyYg6mEa2fo,12408
+torchzero/modules/adaptive/shampoo.py,sha256=1WpjroFS37HmDLV51iK4d8vtnJWFrGCsDkoQav0p47E,10048
+torchzero/modules/adaptive/soap.py,sha256=jyS6F2o4bMKzMU8H2dDggFQEqMqw4W1rX78u8p3uaV4,12619
 torchzero/modules/adaptive/sophia_h.py,sha256=O_izgGlUgUlpH3Oi5PdCKTyxus4yO1PaJUFhGXuGG9k,7063
 torchzero/modules/adaptive/psgd/__init__.py,sha256=g73mAkWEutwU6jzjiwdbYk5Yxgs4i6QVWefFKkm8cDw,223
 torchzero/modules/adaptive/psgd/_psgd_utils.py,sha256=YtwbUKyVWITZPmpwCBJBC42XQP9HcxNx_znEaIv3hsI,1096
@@ -58,21 +62,20 @@ torchzero/modules/adaptive/psgd/psgd_kron_newton.py,sha256=oH-oI1pvbR-z6H6ma1O2G
 torchzero/modules/adaptive/psgd/psgd_kron_whiten.py,sha256=vmhkY6cKaRE5qzy_4tUkIJp6qC3L6ESZMuiU_ih5tR4,7299
 torchzero/modules/adaptive/psgd/psgd_lra_newton.py,sha256=JL8JmqHgcFqfkX7VeD3sRvNj0xeCuDTHxjNyQ_HigBw,4709
 torchzero/modules/adaptive/psgd/psgd_lra_whiten.py,sha256=SaNYtE4_2tV29CbVaTHi8A6RxmhoMaucF5NoMRg6QaA,4197
+torchzero/modules/basis/__init__.py,sha256=MeXoykwqqmWt-Gx8YWMycVL7m5N4j7Ob_L0GbcwLOfM,65
+torchzero/modules/basis/ggt_basis.py,sha256=wVNFN-9a0xGszudMDi_04mqPSschACF7kiftLkMyqYA,7749
+torchzero/modules/basis/soap_basis.py,sha256=pwlxIa9lW9V1NcLPmhm--LVbyq7ALSfkV_4b6ki1hO8,10479
 torchzero/modules/clipping/__init__.py,sha256=ZaffMF7mIRK6hZSfuZadgjNTX6hF5ANiLBny2w3S7I8,250
 torchzero/modules/clipping/clipping.py,sha256=C2dMt0rpuiLMsKq2EWi8qhISSxfCU0nKKGgjWEk2Yxc,14198
-torchzero/modules/clipping/ema_clipping.py,sha256=D4NgXzXYMjK_SKQU3rVoOKzaCd9igGQg_7sXiGMgMqI,6750
-torchzero/modules/clipping/growth_clipping.py,sha256=I1nk5xXBjk0BzWYzMC58LZHouY44myZNIUjM-duv7zc,6508
+torchzero/modules/clipping/ema_clipping.py,sha256=7lFkQWVkchxlZynYXS4JDjhxB8T5tbE0qsP3GXK6mrA,6916
+torchzero/modules/clipping/growth_clipping.py,sha256=VAmUUeIsSGWrGmZiFAngWUBBsxj4d0QAMf36oAMZL8A,6556
 torchzero/modules/conjugate_gradient/__init__.py,sha256=G5WcVoiQYupRBeqjI4lCraGeXNSvWT-_-ynpcE6NQS8,184
 torchzero/modules/conjugate_gradient/cg.py,sha256=fcmP77_v_RPpb0sDV2B_90FvFY67FdJt54KHdccY5YU,14540
-torchzero/modules/experimental/__init__.py,sha256=YbBrWu2vkXHiBcDXmim-Yte4ZxfmQCs_0fCeIArvtnM,942
-torchzero/modules/experimental/adanystrom.py,sha256=fUWPxxi1aJhWme_d31dBG0XxEZY1hJr6AEiFHdFDxCQ,8970
-torchzero/modules/experimental/common_directions_whiten.py,sha256=R_1fQKlvMD99oFrflJLgxl6ObV8jyPc7-NxAUFQeoYA,4941
-torchzero/modules/experimental/coordinate_momentum.py,sha256=HzKy8X5qEvud-xKHJYHpzH6ObxzvYcMcdgodsCw4Bbk,1099
-torchzero/modules/experimental/cubic_adam.py,sha256=RhcHajUfUAcXZDks0X0doR18YtMItQYPmxuEihud4bo,5137
+torchzero/modules/experimental/__init__.py,sha256=najUDh01Av6gEeMYRV9X9lWAr4ZrC6ZgJcPtNpon7ZQ,734
+torchzero/modules/experimental/coordinate_momentum.py,sha256=4BMmgooPysYlX7QOaTUjBn6MNfBAMujM5TCm72vSexw,1152
+torchzero/modules/experimental/cubic_adam.py,sha256=97sgbtkqG1ziXOMxlCor-L-UzzqgSumz8shVOgYL4oQ,5303
 torchzero/modules/experimental/curveball.py,sha256=beHGD1Wh9GxYqMBh1k9Ru6TG3U9eZR6_l8ZUQcZzYxw,2765
 torchzero/modules/experimental/dct.py,sha256=CW-Y2gcjlHlxtIx7SekUOfw2EzujA6v0LcjDYGAfh6M,2433
-torchzero/modules/experimental/eigen_sr1.py,sha256=rCcWVplTWQh91xpgDap35CGEex41C19irUfDlq9lviU,6865
-torchzero/modules/experimental/eigengrad.py,sha256=UPuyo-OmCmu3XLAPclIfsnMN4qcNwX83m7S_55syukA,8455
 torchzero/modules/experimental/fft.py,sha256=s95EzvK4-ZJdwZbVhtqwirY9eVy7v6mFDRMgoLY9wjo,3020
 torchzero/modules/experimental/gradmin.py,sha256=LajM0GU1fB6PsGDg8k0KjKI73RvyZYqPvzcdoVYDq-c,3752
 torchzero/modules/experimental/higher_order_newton.py,sha256=qLSCbkmd7dw0lAhOJGpvvOesZfCMNt2Vz_mc7HknCMQ,12131
@@ -82,15 +85,15 @@ torchzero/modules/experimental/newton_solver.py,sha256=aHZh8EA-QQop3iGz7Ge37KTNg
 torchzero/modules/experimental/newtonnewton.py,sha256=TYUuQwHu8bom08czU9lP7MQq5qFBq_JYZTH_Wmm4g-o,3269
 torchzero/modules/experimental/reduce_outward_lr.py,sha256=ehctg5zLEOHPfiQQUq5ShMj3pDhtxqdNUEneMR9l7Bs,1275
 torchzero/modules/experimental/scipy_newton_cg.py,sha256=psllNtDwUbkVAXBDKwWEueatOmDNPFy-pMwBkqF3_r0,3902
-torchzero/modules/experimental/spsa1.py,sha256=DiQ_nHAC8gnqoNNK7oe6djOiwpwvI5aPtpKA43F7jrQ,3607
 torchzero/modules/experimental/structural_projections.py,sha256=IwpgibNDO0slzMyi6djQXRhQO6IagUgUUCr_-7US1IE,4104
-torchzero/modules/grad_approximation/__init__.py,sha256=_mQ2sWvnMfqc3RQcVmZuBlphtLZCO7z819abGY6kYuM,196
+torchzero/modules/grad_approximation/__init__.py,sha256=BAFXc73_ORySVDyXiyZxpusXWn7K66KFT9LZEMwVKes,221
 torchzero/modules/grad_approximation/fdm.py,sha256=hq7U8UkzCfc7z0J1ZmZo9xOLzHHY0uRjebcwZQrBCzA,4376
 torchzero/modules/grad_approximation/forward_gradient.py,sha256=7fKZoKetYzgD85L3W0x1oG56SdWHj5MDWwmWpV7bpr4,3949
 torchzero/modules/grad_approximation/grad_approximator.py,sha256=hX4nqa0yw1OkA2UKmzZ3HhvMfL0Wwv1yQePxrgAueS8,4782
 torchzero/modules/grad_approximation/rfdm.py,sha256=-5zqMB98YNNa1aQXXtf6UNGSJxySO7mn1NksWyPzp3o,19607
+torchzero/modules/grad_approximation/spsa1.py,sha256=DiQ_nHAC8gnqoNNK7oe6djOiwpwvI5aPtpKA43F7jrQ,3607
 torchzero/modules/least_squares/__init__.py,sha256=mJwE2IXVB3mn_7BzsmDNKhfyViCV8GOrqHJJjz04HR4,41
-torchzero/modules/least_squares/gn.py,sha256=3RQ_7e35Ql9uVUUPi34nef9eQNeZ09fldi964V61Tgg,7889
+torchzero/modules/least_squares/gn.py,sha256=hufsWNq_UdEPFDFKNGgCiM4R9739Xu8JqYWSwKkdSZ8,8087
 torchzero/modules/line_search/__init__.py,sha256=_QjxUJmNC8OqtUuyTJp9wDfHNFKZBZqj6lttWKhG-cI,217
 torchzero/modules/line_search/_polyinterp.py,sha256=i3sNl6SFAUJi4oxhhjBlcxJY9KRunIZjJ8sGdaJOVjc,10990
 torchzero/modules/line_search/adaptive.py,sha256=YNabP6-01dhAUDAOuHRPZCwiV5xTRdHmkN667HQ6V3w,3798
@@ -102,21 +105,21 @@ torchzero/modules/line_search/strong_wolfe.py,sha256=9jGjxebuXHbl8wEFpvV0s4mMX4J
 torchzero/modules/misc/__init__.py,sha256=UYY9CeNepnC8H1LnFa829ux5MEjtGZ9zql624IbCFX8,825
 torchzero/modules/misc/debug.py,sha256=wFt9wB6IdRSsOGLhQjdjmGt4KdB0V5IT0iBFMj97R3Y,1617
 torchzero/modules/misc/escape.py,sha256=c_OMf2jQ7MbxkrXWNmgIpZrBe28N9f89tnzuCQ3fu3A,1930
-torchzero/modules/misc/gradient_accumulation.py,sha256=Xzjt_ulm6Z3mpmtagoUqoefhoeSDVnmX__tVbcI_RQE,2271
+torchzero/modules/misc/gradient_accumulation.py,sha256=1BVqGXwv1YPg7DRJWP0XY6s-vzxrvyXLdruM1Y5KJ5s,2326
 torchzero/modules/misc/homotopy.py,sha256=oa0YFYfv8kkg9v7nukdjTwinuyQa4Nt7kTpddUVCSKg,2257
-torchzero/modules/misc/misc.py,sha256=f-3qxBq1KYI3iGYJXzv1cHEJHc0ScEp-vCLCgiaEgJQ,15002
+torchzero/modules/misc/misc.py,sha256=eWVyYSYiQxcS7G7aVM4nFYiF0csE9qcztTaP4id5CbE,15306
 torchzero/modules/misc/multistep.py,sha256=twdE-lU9Wa0b_uquH9kZ-1OwP0gqWfFMJkdjVWJRwe4,6599
 torchzero/modules/misc/regularization.py,sha256=MCd_tnBYfFnx0b3sM1vHNQ_WbTVfo7l8pxmxGVgWcc0,5935
 torchzero/modules/misc/split.py,sha256=rmi9PgMgiqddrr8fY8Dbdcl2dgwTn9YBAve_bg5Zd08,4288
 torchzero/modules/misc/switch.py,sha256=_ycuD23gR0ZvIUmX3feYBr0_WTX22Pfhu3whpiSCMv4,3678
 torchzero/modules/momentum/__init__.py,sha256=AKWC4HIkN9ZJwN38dJvVJkFEhiP9r93G-kMDokBfsj8,281
-torchzero/modules/momentum/averaging.py,sha256=Q6WLwCJwgNY96YIfQXWpsX-2kDR7n0IOMDfZMvNVc9U,3035
+torchzero/modules/momentum/averaging.py,sha256=OTO_LRNiAhbcKTXrWI-uENqIOH_3DX5_1uYJ3eMVcJY,3202
 torchzero/modules/momentum/cautious.py,sha256=1hD2H08OQaNZG52sheRADBsuf9uJsaoLV4n-UVGUH3Y,8379
-torchzero/modules/momentum/momentum.py,sha256=MPHd4TU1bSlEKLGfueNdmaZ13V5J1suW6agBc3SvrTs,4389
-torchzero/modules/ops/__init__.py,sha256=xUYzWWLlSwaT8sw3dWywkALqI6YGCZgptWQJVy83HhM,1249
-torchzero/modules/ops/accumulate.py,sha256=f-Uutg7gNFRobTc5YI9JlfFiSacXmg0gDhIwQNwZSZg,3439
+torchzero/modules/momentum/momentum.py,sha256=D6Rfy_Ha5jd9uEk3cwCXfGH1dMiP4k4w08SHiE-hChc,4494
+torchzero/modules/ops/__init__.py,sha256=p5hwECuODOv6E4H0lETQHweSsUtMlsGE0d8bfTv2Rwc,1225
+torchzero/modules/ops/accumulate.py,sha256=mbJFwykU2fa6IIfsHVXdhmRp7QX1czpCWjw6AYkNn1k,3636
 torchzero/modules/ops/binary.py,sha256=eB6zwz5ZSSyeWvwVfuOFMjem93oMB7hCo4kNF705jn8,12219
-torchzero/modules/ops/higher_level.py,sha256=cUh-908S0GWVGekmUN5c_Vx0HP3P2tQoKN3COQM5TaQ,8965
+torchzero/modules/ops/higher_level.py,sha256=f9DFNI9rnxc-rShAJOfsiwvyGsWu8FsJwJf5yg_V4eg,9366
 torchzero/modules/ops/multi.py,sha256=WzNK07_wL7z0Gb2pmv5a15Oss6tW9IG79x1c4ZPmOqQ,8643
 torchzero/modules/ops/reduce.py,sha256=SzpkNV5NTsVFp-61a1m8lDKJ1ivJmfQofolFWxbbAe4,6526
 torchzero/modules/ops/unary.py,sha256=vXvWfDFo2CBFwb1ej_WV-fGg61lQRbwN4HklAik8tJY,4844
@@ -136,12 +139,12 @@ torchzero/modules/restarts/__init__.py,sha256=7282ePwN_I0vSeLPYS4TTclE9ZU7pL6Upy
 torchzero/modules/restarts/restars.py,sha256=gcRZ8VHGg60cFVzsk0TWa6-EXoqEFbEeP1p7fs2Av0Q,9348
 torchzero/modules/second_order/__init__.py,sha256=42HeVA3Azl_tXV0_injU-q4QOu7lXzt6AVUcwnPy4Ag,313
 torchzero/modules/second_order/ifn.py,sha256=oAjfFVjLzG6L4n_ELXAWGZSicWizilQy_hQf4hmOoL0,2019
-torchzero/modules/second_order/inm.py,sha256=OddoZHQfSuFnlx_7Zj2qiVcC2A_9yMVn_0Gy1A7hNAg,3420
+torchzero/modules/second_order/inm.py,sha256=_FnaUHKLl46AtI_XYwF52wtOUbAaO5EMUNRJspX5FEM,3574
 torchzero/modules/second_order/multipoint.py,sha256=mHG1SFLsILELIspxZ8U_hxJBlkGwzvUWg96bOIrQsIY,7500
-torchzero/modules/second_order/newton.py,sha256=QcLXsglvf4zJEwR4cldsGVZCABQtxb6U5qVmU3spN_A,11061
-torchzero/modules/second_order/newton_cg.py,sha256=k8G8CSmeIQZObkWVURFnbF_4g2UvJiwh3xToxn7sFJE,14816
-torchzero/modules/second_order/nystrom.py,sha256=WQFfJj0DOfWXyyx36C54m0WqZPIvTTK7n8U7khLhGLg,13359
-torchzero/modules/second_order/rsn.py,sha256=9s-JyJNNeDlIFv8YVGn7y8DGPnP93WJEjpUQXehX3uY,9980
+torchzero/modules/second_order/newton.py,sha256=W37_ePdAB1wnlRrNRd2ovNgkbodK1JV8J4SJytVuF_M,11456
+torchzero/modules/second_order/newton_cg.py,sha256=gHmpLRQ2FRr0750gYkFQ7XweJVZmYI6yG9H2vrKvAdA,14925
+torchzero/modules/second_order/nystrom.py,sha256=lGLjtzq2WAWcaT3E6Say82ySZ1yp9I2ASuOqyNTUmiQ,13361
+torchzero/modules/second_order/rsn.py,sha256=13t42cUvY8JQMC4zf4UsqKvpnTXuXZUZJDECCxRYWjg,11286
 torchzero/modules/smoothing/__init__.py,sha256=RYxCLLfG2onBbMUToaoedsr20rXaayyBt7Ov8OxULrU,80
 torchzero/modules/smoothing/laplacian.py,sha256=1cewdvnneKn51bbIBqKij0bkveKE7wOYCZ-aGlqzK5M,5201
 torchzero/modules/smoothing/sampling.py,sha256=bCH7wlTYZ_vtKUKSkI6znORxQ5Z6DGcpo10F-GYvFlE,12880
@@ -155,7 +158,7 @@ torchzero/modules/trust_region/cubic_regularization.py,sha256=QJjLRkfERvOzV5dTdy
 torchzero/modules/trust_region/dogleg.py,sha256=zwFR49gghxztVGEETF2D4AkeGgHkQRbHGGelav3GuFg,3619
 torchzero/modules/trust_region/levenberg_marquardt.py,sha256=-qbeEW3qRKou48bBdZ-u4Nv43TMt475XV6P_aWfxtqE,5039
 torchzero/modules/trust_region/trust_cg.py,sha256=X9rCJQWvptjZVH2H16iekvAYmleKQAYZKRKC3V0JjFY,4455
-torchzero/modules/trust_region/trust_region.py,sha256=oXMNIvboz0R_1J0Gfd4IvbnwZFl32csNVv-lTYGB0zk,12913
+torchzero/modules/trust_region/trust_region.py,sha256=ax1pJDr3NPLfojUXRMb-hsxD4MpQL1bPAOwozAVTCJI,12930
 torchzero/modules/variance_reduction/__init__.py,sha256=3pwPWZpjgz1btfLJ3rEaK7Wl8B1pDh0HIf0kvD_NJH8,22
 torchzero/modules/variance_reduction/svrg.py,sha256=hXEJ0PUYSksHV0ws3t3cE_4MUTTEn1Htu37iZdDdJCs,8746
 torchzero/modules/weight_decay/__init__.py,sha256=zQrjSujD0c-rKfKjUpuutfAODljsz1hS3zUNJW7zbh4,132
@@ -196,14 +199,14 @@ torchzero/utils/metrics.py,sha256=XPpOvY257tb4mN3Sje1AVNlQkOXiW24_lXXdtd0JYok,31
 torchzero/utils/numberlist.py,sha256=iMoqz4IzXy-aE9bqVYJ21GV6pl0z-NeTsXR-LaI8C24,6229
 torchzero/utils/optimizer.py,sha256=G741IvE57RaVYowr9FEqfRm_opPAeu4UWKU5iPKDMFA,8415
 torchzero/utils/optuna_tools.py,sha256=F-1Xg0n_29MVEb6lqgUFFNIl9BNJ6MOdIJPduoNH4JU,1325
-torchzero/utils/params.py,sha256=nQo270aOURU7rJ_D102y2pSXbzhJPK0Z_ehx4mZBMes,5784
+torchzero/utils/params.py,sha256=-amJs518rpI0zzYavTlWrl60JNrgsk1xxdGvIrSw1ZI,6406
 torchzero/utils/python_tools.py,sha256=HATghTNijlQxmw8rzJfZPPGj1CjcnRxEwogmrgqnARU,4577
-torchzero/utils/tensorlist.py,sha256=4rN8gm967pPmtO5kotXqIX7Mal0ps-IHkGBybfeWY4M,56357
+torchzero/utils/tensorlist.py,sha256=wpzBJvIAmw9VXsg1UF8gZtq-eh7GlvdM6WL_7NyPYlY,56363
 torchzero/utils/thoad_tools.py,sha256=G8k-z0vireEUtI3A_YAR6dtwYjSnN49e_GadcHwwQKc,2319
 torchzero/utils/torch_tools.py,sha256=DsHaSRGZ3-IuySZJTrkojTbaMMlttJFe0hFvB2xnl2U,5069
 torchzero/utils/benchmarks/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
-torchzero/utils/benchmarks/logistic.py,sha256=RHsjHEWkPqaag0kt3wfmdddh4DhftcyW9r70tj9OGp4,4382
-torchzero-0.4.1.dist-info/METADATA,sha256=hB0rFqXnaRbwVkFRwTwjXpKnIFLi8MBvLXbgXTuUGWk,564
-torchzero-0.4.1.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
-torchzero-0.4.1.dist-info/top_level.txt,sha256=ETW_iE2ubg0oMyef_h-ayB5i1OOZZd4SNdR3ltIbHe0,16
-torchzero-0.4.1.dist-info/RECORD,,
+torchzero/utils/benchmarks/logistic.py,sha256=1c9kB6tDaKsSNlQn44_Lso2_g-85fQK45RvwLZOcJOo,4587
+torchzero-0.4.2.dist-info/METADATA,sha256=nApA6WdQrTYR0c5TXCORxOktKgVwxlyMqgnfkKNPHLk,564
+torchzero-0.4.2.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
+torchzero-0.4.2.dist-info/top_level.txt,sha256=ETW_iE2ubg0oMyef_h-ayB5i1OOZZd4SNdR3ltIbHe0,16
+torchzero-0.4.2.dist-info/RECORD,,