lsst.scarlet.lite.parameters.AdaproxParameter Class Reference

Inheritance diagram for lsst.scarlet.lite.parameters.AdaproxParameter:

Public Member Functions
	__init__ (self, np.ndarray x, Callable|float step, Callable|None grad=None, Callable|None prox=None, float b1=0.9, float b2=0.999, float eps=1e-8, float p=0.25, np.ndarray|None m0=None, np.ndarray|None v0=None, np.ndarray|None vhat0=None, str scheme="amsgrad", float prox_e_rel=1e-6)
	update (self, int it, np.ndarray input_grad, *args)

Public Attributes
	b1
	b2
	eps
	p
	phi_psi
	e_rel
	x

Detailed Description

Operator updated using te Proximal ADAM algorithm

Uses multiple variants of adaptive quasi-Newton gradient descent
    * Adam (Kingma & Ba 2015)
    * NAdam (Dozat 2016)
    * AMSGrad (Reddi, Kale & Kumar 2018)
    * PAdam (Chen & Gu 2018)
    * AdamX (Phuong & Phong 2019)
    * RAdam (Liu et al. 2019)
See details of the algorithms in the respective papers.

Definition at line 356 of file parameters.py.

Constructor & Destructor Documentation

__init__()

lsst.scarlet.lite.parameters.AdaproxParameter.__init__	(		self,
		np.ndarray	x,
		Callable | float	step,
		Callable | None	grad = None,
		Callable | None	prox = None,
		float	b1 = 0.9,
		float	b2 = 0.999,
		float	eps = 1e-8,
		float	p = 0.25,
		np.ndarray | None	m0 = None,
		np.ndarray | None	v0 = None,
		np.ndarray | None	vhat0 = None,
		str	scheme = "amsgrad",
		float	prox_e_rel = 1e-6 )

Reimplemented from lsst.scarlet.lite.parameters.Parameter.

Definition at line 369 of file parameters.py.

    ):
        shape = x.shape
        dtype = x.dtype
        if m0 is None:
            m0 = np.zeros(shape, dtype=dtype)
 
        if v0 is None:
            v0 = np.zeros(shape, dtype=dtype)
 
        if vhat0 is None:
            vhat0 = np.ones(shape, dtype=dtype) * -np.inf
 
        super().__init__(
            x,
            {
                "m": m0,
                "v": v0,
                "vhat": vhat0,
            },
            step,
            grad,
            prox,
        )
 
        if isinstance(b1, float):
            _b1 = SingleItemArray(b1)
        else:
            _b1 = b1
 
        self.b1 = _b1
        self.b2 = b2
        self.eps = eps
        self.p = p
 
        self.phi_psi = phi_psi[scheme]
        self.e_rel = prox_e_rel
 

Member Function Documentation

update()

lsst.scarlet.lite.parameters.AdaproxParameter.update	(		self,
		int	it,
		np.ndarray	input_grad,
		*	args )

Update the parameter and meta-parameters using the PGM

See `~Parameter` for more.

Reimplemented from lsst.scarlet.lite.parameters.Parameter.

Definition at line 421 of file parameters.py.

    def update(self, it: int, input_grad: np.ndarray, *args):
        """Update the parameter and meta-parameters using the PGM
 
        See `~Parameter` for more.
        """
        _x = self.x
        # Calculate the gradient
        grad = cast(Callable, self.grad)(input_grad, _x, *args)
        # Get the update for the parameter
        phi, psi = self.phi_psi(
            it,
            grad,
            self.helpers["m"],
            self.helpers["v"],
            self.helpers["vhat"],
            self.b1,
            self.b2,
            self.eps,
            self.p,
        )
        # Calculate the step size
        step = self.step
        if it > 0:
            _x += -step * phi / psi
        else:
            # This is a scheme that Peter Melchior and I came up with to
            # dampen the known affect of ADAM, where the first iteration
            # is often much larger than desired.
            _x += -step * phi / psi / 10
 
        self.x = cast(Callable, self.prox)(_x)
 
 

Member Data Documentation

b1

lsst.scarlet.lite.parameters.AdaproxParameter.b1

Definition at line 413 of file parameters.py.

b2

lsst.scarlet.lite.parameters.AdaproxParameter.b2

Definition at line 414 of file parameters.py.

e_rel

lsst.scarlet.lite.parameters.AdaproxParameter.e_rel

Definition at line 419 of file parameters.py.

eps

lsst.scarlet.lite.parameters.AdaproxParameter.eps

Definition at line 415 of file parameters.py.

p

lsst.scarlet.lite.parameters.AdaproxParameter.p

Definition at line 416 of file parameters.py.

phi_psi

lsst.scarlet.lite.parameters.AdaproxParameter.phi_psi

Definition at line 418 of file parameters.py.

x

lsst.scarlet.lite.parameters.AdaproxParameter.x

Definition at line 451 of file parameters.py.

---

The documentation for this class was generated from the following file:

• /j/snowflake/release/lsstsw/stack/lsst-scipipe-8.0.0/Linux64/scarlet_lite/gae0086650b+585e252eca/python/lsst/scarlet/lite/parameters.py