torchzero 0.3.14__py3-none-any.whl → 0.3.15__py3-none-any.whl

torchzero/core/module.py

@@ -3,7 +3,7 @@ from abc import ABC, abstractmethod
 from collections import ChainMap, defaultdict
 from collections.abc import Callable, Iterable, MutableMapping, Sequence
 from operator import itemgetter
-from typing import Any, final, overload, Literal, cast
+from typing import Any, Literal, cast, final, overload
 
 import torch
 
@@ -13,259 +13,14 @@ from ..utils import (
     Params,
     _make_param_groups,
     get_state_vals,
+    vec_to_tensors,
 )
-from ..utils.derivatives import hvp, hvp_fd_central, hvp_fd_forward
-from ..utils.python_tools import flatten
+from ..utils.derivatives import flatten_jacobian, hvp, hvp_fd_central, hvp_fd_forward
 from ..utils.linalg.linear_operator import LinearOperator
+from ..utils.python_tools import flatten
+from .var import Var
 
 
-def _closure_backward(closure, params, retain_graph, create_graph):
-    with torch.enable_grad():
-        if not (retain_graph or create_graph):
-            return closure()
-
-        for p in params: p.grad = None
-        loss = closure(False)
-        grad = torch.autograd.grad(loss, params, retain_graph=retain_graph, create_graph=create_graph)
-        for p,g in zip(params,grad): p.grad = g
-        return loss
-
-# region Vars
-# ----------------------------------- var ----------------------------------- #
-class Var:
-    """
-    Holds parameters, gradient, update, objective function (closure) if supplied, loss, and some other info.
-    Modules take in a ``Var`` object, modify and it is passed to the next module.
-
-    """
-    def __init__(
-        self,
-        params: list[torch.Tensor],
-        closure: Callable | None,
-        model: torch.nn.Module | None,
-        current_step: int,
-        parent: "Var | None" = None,
-        modular: "Modular | None" = None,
-        loss: torch.Tensor | None = None,
-        storage: dict | None = None,
-    ):
-        self.params: list[torch.Tensor] = params
-        """List of all parameters with requires_grad = True."""
-
-        self.closure = closure
-        """A closure that reevaluates the model and returns the loss, None if it wasn't specified"""
-
-        self.model = model
-        """torch.nn.Module object of the model, None if it wasn't specified."""
-
-        self.current_step: int = current_step
-        """global current step, starts at 0. This may not correspond to module current step,
-        for example a module may step every 10 global steps."""
-
-        self.parent: "Var | None" = parent
-        """parent ``Var`` object. When ``self.get_grad()`` is called, it will also set ``parent.grad``.
-        Same with ``self.get_loss()``. This is useful when ``self.params`` are different from ``parent.params``,
-        e.g. when projecting."""
-
-        self.modular: "Modular" = cast(Modular, modular)
-        """Modular optimizer object that created this ``Var``."""
-
-        self.update: list[torch.Tensor] | None = None
-        """
-        current update. Update is assumed to be a transformed gradient, therefore it is subtracted.
-
-        If closure is None, this is initially set to cloned gradient. Otherwise this is set to None.
-
-        At the end ``var.get_update()`` is subtracted from parameters. Therefore if ``var.update`` is ``None``,
-        gradient will be used and calculated if needed.
-        """
-
-        self.grad: list[torch.Tensor] | None = None
-        """gradient with current parameters. If closure is not ``None``, this is set to ``None`` and can be calculated if needed."""
-
-        self.loss: torch.Tensor | Any | None = loss
-        """loss with current parameters."""
-
-        self.loss_approx: torch.Tensor | Any | None = None
-        """loss at a point near current point. This can be useful as some modules only calculate loss at perturbed points,
-        whereas some other modules require loss strictly at current point."""
-
-        self.post_step_hooks: list[Callable[[Modular, Var]]] = []
-        """list of functions to be called after optimizer step.
-
-        This attribute should always be modified in-place (using ``append`` or ``extend``).
-
-        The signature is:
-
-        ```python
-        def hook(optimizer: Modular, var: Vars): ...
-        ```
-        """
-
-        self.is_last: bool = False
-        """
-        Indicates that current module is either last or next-to-last before a learning rate module.
-        This is always False if current module has children or is a child.
-        This is because otherwise the ``is_last`` would be passed to child modules, even though they aren't last.
-        """
-
-        self.nested_is_last: bool = False
-        """
-        Indicates that current module is either last or next-to-last before a learning rate module, for modules
-        that have children. This will be passed to the children unless ``var.clone()`` is used, therefore
-        a child of a last module may also receive ``var.nested_is_last=True``.
-        """
-
-        self.last_module_lrs: list[float] | None = None
-        """
-        List of per-parameter learning rates if current module is next-to-last before a
-        learning rate module, otherwise this is set to None. Ignore this unless you are manually applying
-        update to parameters.
-        """
-
-        self.stop: bool = False
-        """if True, all following modules will be skipped.
-        If this module is a child, it only affects modules at the same level (in the same Chain)."""
-
-        self.skip_update: bool = False
-        """if True, the parameters will not be updated."""
-
-        # self.storage: dict = {}
-        # """Storage for any other data, such as hessian estimates, etc."""
-
-        self.attrs: dict = {}
-        """attributes, Modular.attrs is updated with this after each step. This attribute should always be modified in-place"""
-
-        if storage is None: storage = {}
-        self.storage: dict = storage
-        """additional kwargs passed to closure will end up in this dict. This attribute should always be modified in-place"""
-
-        self.should_terminate: bool | None = None
-        """termination criteria, Modular.should_terminate is set to this after each step if not None"""
-
-    def get_loss(self, backward: bool, retain_graph = None, create_graph: bool = False) -> torch.Tensor | float:
-        """Returns the loss at current parameters, computing it if it hasn't been computed already and assigning ``var.loss``.
-        Do not call this at perturbed parameters. Backward always sets grads to None before recomputing."""
-        if self.loss is None:
-
-            if self.closure is None: raise RuntimeError("closure is None")
-            if backward:
-                with torch.enable_grad():
-                    self.loss = self.loss_approx = _closure_backward(
-                        closure=self.closure, params=self.params, retain_graph=retain_graph, create_graph=create_graph
-                    )
-
-                # initializing to zeros_like is equivalent to using zero_grad with set_to_none = False.
-                # it is technically a more correct approach for when some parameters conditionally receive gradients
-                # and in this case it shouldn't be slower.
-
-                # next time closure() is called, it will set grad to None.
-                # zero_grad(set_to_none=False) shouldn't be used (I should add a warning)
-                self.grad = [p.grad if p.grad  is not None else torch.zeros_like(p) for p in self.params]
-            else:
-                self.loss = self.loss_approx = self.closure(False)
-
-        # if self.loss was not None, above branch wasn't executed because loss has already been evaluated, but without backward since self.grad is None.
-        # and now it is requested to be evaluated with backward.
-        if backward and self.grad is None:
-            warnings.warn('get_loss was called with backward=False, and then with backward=True so it had to be re-evaluated, so the closure was evaluated twice where it could have been evaluated once.')
-            if self.closure is None: raise RuntimeError("closure is None")
-
-            with torch.enable_grad():
-                self.loss = self.loss_approx = _closure_backward(
-                    closure=self.closure, params=self.params, retain_graph=retain_graph, create_graph=create_graph
-                )
-            self.grad = [p.grad if p.grad is not None else torch.zeros_like(p) for p in self.params]
-
-        # set parent grad
-        if self.parent is not None:
-            # the way projections/split work, they make a new closure which evaluates original
-            # closure and projects the gradient, and set it as their var.closure.
-            # then on `get_loss(backward=True)` it is called, so it also sets original parameters gradient.
-            # and we set it to parent var here.
-            if self.parent.loss is None: self.parent.loss = self.loss
-            if self.parent.grad is None and backward:
-                if all(p.grad is None for p in self.parent.params):
-                    warnings.warn("Parent grad is None after backward.")
-                self.parent.grad = [p.grad if p.grad is not None else torch.zeros_like(p) for p in self.parent.params]
-
-        return self.loss # type:ignore
-
-    def get_grad(self, retain_graph: bool | None = None, create_graph: bool = False) -> list[torch.Tensor]:
-        """Returns the gradient at initial parameters, computing it if it hasn't been computed already and assigning
-        ``var.grad`` and potentially ``var.loss``. Do not call this at perturbed parameters."""
-        if self.grad is None:
-            if self.closure is None: raise RuntimeError("closure is None")
-            self.get_loss(backward=True, retain_graph=retain_graph, create_graph=create_graph) # evaluate and set self.loss and self.grad
-
-        assert self.grad is not None
-        return self.grad
-
-    def get_update(self) -> list[torch.Tensor]:
-        """Returns the update. If update is None, it is initialized by cloning the gradients and assigning to ``var.update``.
-        Computing the gradients may assign ``var.grad`` and ``var.loss`` if they haven't been computed.
-        Do not call this at perturbed parameters."""
-        if self.update is None: self.update = [g.clone() for g in self.get_grad()]
-        return self.update
-
-    def clone(self, clone_update: bool, parent: "Var | None" = None):
-        """Creates a shallow copy of the Vars object, update can optionally be deep-copied (via ``torch.clone``).
-
-        Doesn't copy ``is_last``, ``nested_is_last`` and ``last_module_lrs``. They will always be ``False``/``None``.
-
-        Setting ``parent`` is only if clone's parameters are something different,
-        while clone's closure referes to the same objective but with a "view" on parameters.
-        """
-        copy = Var(params = self.params, closure=self.closure, model=self.model, current_step=self.current_step, parent=parent)
-
-        if clone_update and self.update is not None:
-            copy.update = [u.clone() for u in self.update]
-        else:
-            copy.update = self.update
-
-        copy.grad = self.grad
-        copy.loss = self.loss
-        copy.loss_approx = self.loss_approx
-        copy.closure = self.closure
-        copy.post_step_hooks = self.post_step_hooks
-        copy.stop = self.stop
-        copy.skip_update = self.skip_update
-
-        copy.modular = self.modular
-        copy.attrs = self.attrs
-        copy.storage = self.storage
-        copy.should_terminate = self.should_terminate
-
-        return copy
-
-    def update_attrs_from_clone_(self, var: "Var"):
-        """Updates attributes of this `Vars` instance from a cloned instance.
-        Typically called after a child module has processed a cloned `Vars`
-        object. This propagates any newly computed loss or gradient values
-        from the child's context back to the parent `Vars` if the parent
-        didn't have them computed already.
-
-        Also, as long as ``post_step_hooks`` and ``attrs`` are modified in-place,
-        if the child updates them, the update will affect the parent too.
-        """
-        if self.loss is None: self.loss = var.loss
-        if self.loss_approx is None: self.loss_approx = var.loss_approx
-        if self.grad is None: self.grad = var.grad
-
-        if var.should_terminate is not None: self.should_terminate = var.should_terminate
-
-    def zero_grad(self, set_to_none=True):
-        if set_to_none:
-            for p in self.params: p.grad = None
-        else:
-            grads = [p.grad for p in self.params if p.grad is not None]
-            if len(grads) != 0: torch._foreach_zero_(grads)
-
-# endregion
-
-
-# region Module
-# ---------------------------------- module ---------------------------------- #
 class Module(ABC):
     """Abstract base class for an optimizer modules.
 
@@ -317,9 +72,12 @@ class Module(ABC):
         return self
 
     def set_child(self, key: str, module: "Module | Sequence[Module]"):
+        from .chain import maybe_chain
         self.children[key] = maybe_chain(module)
 
     def set_children_sequence(self, modules: "Iterable[Module | Sequence[Module]]", prefix = 'module_'):
+        from .chain import maybe_chain
+
         modules = list(modules)
         for i, m in enumerate(modules):
             self.set_child(f'{prefix}{i}', maybe_chain(m))
@@ -558,82 +316,6 @@ class Module(ABC):
         """``_extra_pack`` return will be passed to this method when loading state_dict.
         This method is called after loading the rest of the state dict"""
 
-
-
-    # ------------------------------ HELPER METHODS ------------------------------ #
-    @torch.no_grad
-    def Hvp(
-        self,
-        v: Sequence[torch.Tensor],
-        at_x0: bool,
-        var: Var,
-        rgrad: Sequence[torch.Tensor] | None,
-        hvp_method: Literal['autograd', 'forward', 'central'],
-        h: float,
-        normalize: bool,
-        retain_grad: bool,
-    ) -> tuple[Sequence[torch.Tensor], Sequence[torch.Tensor] | None]:
-        """
-        Returns ``(Hvp, rgrad)``, where ``rgrad`` is gradient at current parameters,
-        possibly with ``create_graph=True``, or it may be None with ``hvp_method="central"``.
-        Gradient is set to vars automatically if ``at_x0``, you can always access it with ``vars.get_grad()``
-
-        Single sample example:
-
-        ```python
-        Hvp, _ = self.hvp(v, at_x0=True, rgrad=None, ..., retain_graph=False)
-        ```
-
-        Multiple samples example:
-
-        ```python
-        D = None
-        rgrad = None
-        for i in range(n_samples):
-            v = [torch.randn_like(p) for p in params]
-            Hvp, rgrad = self.hvp(v, at_x0=True, rgrad=rgrad, ..., retain_graph=i < n_samples-1)
-
-            if D is None: D = Hvp
-            else: torch._foreach_add_(D, Hvp)
-
-        if n_samples > 1: torch._foreach_div_(D, n_samples)
-        ```
-
-        Args:
-            v (Sequence[torch.Tensor]): vector in hessian-vector product
-            at_x0 (bool): whether this is being called at original or perturbed parameters.
-            var (Var): Var
-            rgrad (Sequence[torch.Tensor] | None): pass None initially, then pass what this returns.
-            hvp_method (str): hvp method.
-            h (float): finite difference step size
-            normalize (bool): whether to normalize v for finite difference
-            retain_grad (bool): retain grad
-        """
-        # get grad
-        if rgrad is None and hvp_method in ('autograd', 'forward'):
-            if at_x0: rgrad = var.get_grad(create_graph = hvp_method=='autograd')
-            else:
-                if var.closure is None: raise RuntimeError("Closure is required to calculate HVp")
-                with torch.enable_grad():
-                    loss = var.closure()
-                    rgrad = torch.autograd.grad(loss, var.params, create_graph = hvp_method=='autograd')
-
-        if hvp_method == 'autograd':
-            assert rgrad is not None
-            Hvp = hvp(var.params, rgrad, v, retain_graph=retain_grad)
-
-        elif hvp_method == 'forward':
-            assert rgrad is not None
-            loss, Hvp = hvp_fd_forward(var.closure, var.params, v, h=h, g_0=rgrad, normalize=normalize)
-
-        elif hvp_method == 'central':
-            loss, Hvp = hvp_fd_central(var.closure, var.params, v, h=h, normalize=normalize)
-
-        else:
-            raise ValueError(hvp_method)
-
-        return Hvp, rgrad
-
     def get_generator(self, device: torch.types.Device, seed: int | None):
         if seed is None: return None
 
@@ -642,277 +324,4 @@ class Module(ABC):
 
         return self.global_state['generator']
 
-# endregion
-
 Chainable = Module | Sequence[Module]
-
-
-def unroll_modules(*modules: Chainable) -> list[Module]:
-    unrolled = []
-
-    for m in modules:
-        if isinstance(m, Module):
-            unrolled.append(m)
-            unrolled.extend(unroll_modules(list(m.children.values())))
-        else:
-            unrolled.extend(unroll_modules(*m))
-
-    return unrolled
-
-
-# region Modular
-# ---------------------------------- Modular --------------------------------- #
-
-class _EvalCounterClosure:
-    """keeps track of how many times closure has been evaluated, and sets closure return"""
-    __slots__ = ("modular", "closure")
-    def __init__(self, modular: "Modular", closure):
-        self.modular = modular
-        self.closure = closure
-
-    def __call__(self, *args, **kwargs):
-        if self.closure is None:
-            raise RuntimeError("One of the modules requires closure to be passed to the step method")
-
-        v = self.closure(*args, **kwargs)
-
-        # set closure return on 1st evaluation
-        if self.modular._closure_return is None:
-            self.modular._closure_return = v
-
-        self.modular.num_evaluations += 1
-        return v
-
-# have to inherit from Modular to support lr schedulers
-# although Accelerate doesn't work due to converting param_groups to a dict
-class Modular(torch.optim.Optimizer):
-    """Chains multiple modules into an optimizer.
-
-    Args:
-        params (Params | torch.nn.Module): An iterable of parameters to optimize
-            (typically `model.parameters()`), an iterable of parameter group dicts,
-            or a `torch.nn.Module` instance.
-        *modules (Module): A sequence of `Module` instances that define the
-            optimization algorithm steps.
-    """
-    # this is specifically for lr schedulers
-    param_groups: list[ChainMap[str, Any]] # pyright:ignore[reportIncompatibleVariableOverride]
-
-    def __init__(self, params: Params | torch.nn.Module, *modules: Module):
-        if len(modules) == 0: raise RuntimeError("Empty list of modules passed to `Modular`")
-        self.model: torch.nn.Module | None = None
-        """The model whose parameters are being optimized, if a model instance was passed to `__init__`."""
-        if isinstance(params, torch.nn.Module):
-            self.model = params
-            params = params.parameters()
-
-        self.modules = modules
-        """Top-level modules providedduring initialization."""
-
-        self.unrolled_modules = unroll_modules(self.modules)
-        """A flattened list of all modules including all children."""
-
-        param_groups = _make_param_groups(params, differentiable=False)
-        self._per_parameter_global_settings: dict[torch.Tensor, list[MutableMapping[str, Any]]] = {}
-
-        # make sure there is no more than a single learning rate module
-        lr_modules = [m for m in self.unrolled_modules if 'lr' in m.defaults]
-        if len(lr_modules) > 1:
-            warnings.warn(f'multiple learning rate modules detected: {lr_modules}. This may lead to componding of learning rate multiplication with per-parameter learning rates and schedulers.')
-
-        # iterate over all per-parameter settings overrides and check if they are applied at most once
-        for group in param_groups:
-            for k in group:
-                if k in ('params', 'lr'): continue
-                modules_with_k = [m for m in self.unrolled_modules if k in m.defaults and k not in m._overridden_keys]
-                if len(modules_with_k) > 1:
-                    warnings.warn(f'`params` has a `{k}` key, and multiple modules have that key: {modules_with_k}. If you intended to only set `{k}` to one of them, use `module.set_param_groups(params)`')
-
-        # defaults for schedulers
-        defaults = {}
-        for m in self.unrolled_modules: defaults.update(m.defaults)
-        super().__init__(param_groups, defaults=defaults)
-
-        # note - this is what super().__init__(param_groups, defaults=defaults) does:
-
-        # self.defaults = defaults
-        # for param_group in param_groups:
-        #     self.add_param_group(param_group)
-
-        # add_param_group adds a ChainMap where defaults are lowest priority,
-        # and entries specifed in param_groups or scheduler are higher priority.
-        # pytorch schedulers do group["lr"] = new_lr, which sets higher priority key.
-        # in each module, settings passed to that module by calling set_param_groups are highest priority
-
-        self.current_step = 0
-        """global step counter for the optimizer."""
-
-        self.num_evaluations = 0
-        """number of times the objective has been evaluated (number of closure calls or number of steps if closure is None)."""
-
-        # reformulations will change the closure to return a different loss (e.g. a sqrt homotopy, gaussian homotopy)
-        # we want to return original loss so this attribute is used
-        self._closure_return = None
-        """on each step, first time a closure is evaluated, this attribute is set to the returned value. `step` method returns this."""
-
-        self.attrs = {}
-        """custom attributes that can be set by modules, for example EMA of weights or best so far"""
-
-        self.should_terminate = False
-        """is set to True by termination criteria modules."""
-
-    def add_param_group(self, param_group: dict[str, Any]):
-        proc_param_group = _make_param_groups([param_group], differentiable=False)[0]
-        self.param_groups.append(ChainMap(proc_param_group, self.defaults))
-
-        for p in proc_param_group['params']:
-            # updates global per-parameter setting overrides (medium priority)
-            self._per_parameter_global_settings[p] = [m.settings[p].maps[1] for m in self.unrolled_modules]
-
-    def state_dict(self):
-        all_params = [p for g in self.param_groups for p in g['params']]
-        id_to_idx = {id(p): i for i,p in enumerate(all_params)}
-
-        groups = []
-        for g in self.param_groups:
-            g = g.copy()
-            g['params'] = [id_to_idx[id(p)] for p in g['params']]
-            groups.append(g)
-
-        state_dict = {
-            "idx_to_id": {v:k for k,v in id_to_idx.items()},
-            "params": all_params,
-            "groups": groups,
-            "defaults": self.defaults,
-            "modules": {i: m.state_dict() for i, m in enumerate(self.unrolled_modules)}
-        }
-        return state_dict
-
-    def load_state_dict(self, state_dict: dict):
-        self.defaults.clear()
-        self.defaults.update(state_dict['defaults'])
-
-        idx_to_param = dict(enumerate(state_dict['params']))
-        groups = []
-        for g in state_dict['groups']:
-            g = g.copy()
-            g['params'] = [idx_to_param[p] for p in g['params']]
-            groups.append(g)
-
-        self.param_groups.clear()
-        for group in groups:
-            self.add_param_group(group)
-
-        id_to_tensor = {state_dict['idx_to_id'][i]: p for i,p in enumerate(state_dict['params'])}
-        for m, sd in zip(self.unrolled_modules, state_dict['modules'].values()):
-            m._load_state_dict(sd, id_to_tensor)
-
-
-    def step(self, closure=None, loss=None, **kwargs): # pyright: ignore[reportIncompatibleMethodOverride]
-        # clear closure return from previous step
-        self._closure_return = None
-
-        # propagate global per-parameter setting overrides
-        for g in self.param_groups:
-            settings = dict(g.maps[0]) # ignore defaults
-            params = settings.pop('params')
-            if not settings: continue
-
-            for p in params:
-                if not p.requires_grad: continue
-                for map in self._per_parameter_global_settings[p]: map.update(settings)
-
-        # create var
-        params = [p for g in self.param_groups for p in g['params'] if p.requires_grad]
-        var = Var(params=params, closure=_EvalCounterClosure(self, closure), model=self.model, current_step=self.current_step, modular=self, loss=loss, storage=kwargs)
-
-        # if closure is None, assume backward has been called and gather grads
-        if closure is None:
-            var.grad = [p.grad if p.grad is not None else torch.zeros_like(p) for p in params]
-            self.num_evaluations += 1
-
-        n_modules = len(self.modules)
-        if n_modules == 0: raise RuntimeError("There are no modules in this `Modular` optimizer")
-        last_module = self.modules[-1]
-        last_lr = last_module.defaults.get('lr', None)
-
-        # step
-        for i, module in enumerate(self.modules):
-            if i!=0: var = var.clone(clone_update=False)
-
-            # last module, or next to last module before lr
-            if (i == n_modules - 1) or ((i == n_modules - 2) and (last_lr is not None)):
-                if module.children: var.nested_is_last = True
-                else: var.is_last = True
-                if last_lr is not None: var.last_module_lrs = [last_module.settings[p]['lr'] for p in var.params]
-
-            var = module.step(var)
-            if var.stop: break
-
-        # apply update
-        if not var.skip_update:
-            with torch.no_grad():
-                torch._foreach_sub_(params, var.get_update())
-
-        # update attributes
-        self.attrs.update(var.attrs)
-        if var.should_terminate is not None: self.should_terminate = var.should_terminate
-
-        # hooks
-        for hook in var.post_step_hooks:
-            hook(self, var)
-
-        self.current_step += 1
-        #return var.loss if var.loss is not None else var.loss_approx
-        return self._closure_return
-
-    def __repr__(self):
-        return f'Modular({", ".join(str(m) for m in self.modules)})'
-# endregion
-
-# region Chain
-# ----------------------------------- Chain ---------------------------------- #
-class Chain(Module):
-    """Chain of modules, mostly used internally"""
-    def __init__(self, *modules: Module | Iterable[Module]):
-        super().__init__()
-        flat_modules: list[Module] = flatten(modules)
-        for i, module in enumerate(flat_modules):
-            self.set_child(f'module_{i}', module)
-
-    def update(self, var):
-        # note here that `update` and `apply` shouldn't be used directly
-        # as it will update all modules, and then apply all modules
-        # it is used in specific cases like Chain as trust region hessian module
-        for i in range(len(self.children)):
-            self.children[f'module_{i}'].update(var)
-            if var.stop: break
-        return var
-
-    def apply(self, var):
-        for i in range(len(self.children)):
-            var = self.children[f'module_{i}'].apply(var)
-            if var.stop: break
-        return var
-
-    def step(self, var):
-        for i in range(len(self.children)):
-            var = self.children[f'module_{i}'].step(var)
-            if var.stop: break
-        return var
-
-    def __repr__(self):
-        s = self.__class__.__name__
-        if self.children:
-            if s == 'Chain': s = 'C' # to shorten it
-            s = f'{s}({", ".join(str(m) for m in self.children.values())})'
-        return s
-
-def maybe_chain(*modules: Chainable) -> Module:
-    """Returns a single module directly if only one is provided, otherwise wraps them in a :code:`Chain`."""
-    flat_modules: list[Module] = flatten(modules)
-    if len(flat_modules) == 1:
-        return flat_modules[0]
-    return Chain(*flat_modules)
-# endregion
-

torchzero/core/reformulation.py

@@ -3,7 +3,9 @@ from collections.abc import Callable, Sequence
 
 import torch
 
-from .module import Chainable, Modular, Module, Var
+from .chain import Chain
+from .module import Chainable, Module
+from .var import Var
 
 
 class Reformulation(Module, ABC):