Skip to content

MAP Optimizers and solvers

Optimizers module.

AdamOptimizer

Bases: BaseOptimizer

Adam optimizer for NumPyro models.

This class implements the Adam optimization algorithm.

Parameters:

Name Type Description Default
step_size float

The step size (learning rate) for the optimizer. Default is 0.001.

 0.001
Source code in src/prophetverse/engine/optimizer/optimizer.py 
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
class AdamOptimizer(BaseOptimizer):
    """
    Adam optimizer for NumPyro models.

    This class implements the Adam optimization algorithm.

    Parameters
    ----------
    step_size : float, optional
        The step size (learning rate) for the optimizer. Default is 0.001.
    """

    def __init__(self, step_size=0.001):
        """Initialize the AdamOptimizer."""
        self.step_size = step_size
        super().__init__()

    def create_optimizer(self) -> _NumPyroOptim:
        """
        Create and return a NumPyro Adam optimizer instance.

        Returns
        -------
        _NumPyroOptim
            An instance of NumPyro's Adam optimizer.
        """
        return numpyro.optim.Adam(step_size=self.step_size)

__init__(step_size=0.001)

Initialize the AdamOptimizer.

Source code in src/prophetverse/engine/optimizer/optimizer.py 
75
76
77
78
def __init__(self, step_size=0.001):
    """Initialize the AdamOptimizer."""
    self.step_size = step_size
    super().__init__()

create_optimizer()

Create and return a NumPyro Adam optimizer instance.

Returns:

Type Description
_NumPyroOptim

An instance of NumPyro's Adam optimizer.
Source code in src/prophetverse/engine/optimizer/optimizer.py 
80
81
82
83
84
85
86
87
88
89
def create_optimizer(self) -> _NumPyroOptim:
    """
    Create and return a NumPyro Adam optimizer instance.

    Returns
    -------
    _NumPyroOptim
        An instance of NumPyro's Adam optimizer.
    """
    return numpyro.optim.Adam(step_size=self.step_size)

BaseOptimizer

Bases: BaseObject

Base class for optimizers in NumPyro.

This abstract base class defines the interface that all optimizers must implement.

Source code in src/prophetverse/engine/optimizer/optimizer.py 
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
class BaseOptimizer(BaseObject):
    """
    Base class for optimizers in NumPyro.

    This abstract base class defines the interface that all optimizers must implement.
    """

    _tags = {"object_type": "optimizer", "is_solver": False}

    def __init__(self):
        """
        Initialize the BaseOptimizer.

        Since this is an abstract base class, initialization does not do anything.
        """
        ...

    def create_optimizer(self) -> _NumPyroOptim:
        """
        Create and return a NumPyro optimizer instance.

        Returns
        -------
        _NumPyroOptim
            An instance of a NumPyro optimizer.

        Raises
        ------
        NotImplementedError
            This method must be implemented in subclasses.
        """
        raise NotImplementedError(
            "create_optimizer method must be implemented in subclass"
        )

__init__()

Initialize the BaseOptimizer.

Since this is an abstract base class, initialization does not do anything.

Source code in src/prophetverse/engine/optimizer/optimizer.py 
36
37
38
39
40
41
42
def __init__(self):
    """
    Initialize the BaseOptimizer.

    Since this is an abstract base class, initialization does not do anything.
    """
    ...

create_optimizer()

Create and return a NumPyro optimizer instance.

Returns:

Type Description
_NumPyroOptim

An instance of a NumPyro optimizer.

Raises:

Type Description
NotImplementedError

This method must be implemented in subclasses.
Source code in src/prophetverse/engine/optimizer/optimizer.py 
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
def create_optimizer(self) -> _NumPyroOptim:
    """
    Create and return a NumPyro optimizer instance.

    Returns
    -------
    _NumPyroOptim
        An instance of a NumPyro optimizer.

    Raises
    ------
    NotImplementedError
        This method must be implemented in subclasses.
    """
    raise NotImplementedError(
        "create_optimizer method must be implemented in subclass"
    )

CosineScheduleAdamOptimizer

Bases: BaseOptimizer

Adam optimizer with cosine decay learning rate schedule.

This optimizer combines the Adam optimizer with a cosine decay schedule for the learning rate.

Parameters:

Name Type Description Default
init_value float

Initial learning rate. Default is 0.001.

 0.001
decay_steps int

Number of steps over which the learning rate decays. Default is 100_000.

 100000
alpha float

Final multiplier for the learning rate. Default is 0.0.

 0.0
exponent int

Exponent for the cosine decay schedule. Default is 1.

 1
Source code in src/prophetverse/engine/optimizer/optimizer.py 
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
class CosineScheduleAdamOptimizer(BaseOptimizer):
    """
    Adam optimizer with cosine decay learning rate schedule.

    This optimizer combines the Adam optimizer with a cosine decay schedule
    for the learning rate.

    Parameters
    ----------
    init_value : float, optional
        Initial learning rate. Default is 0.001.
    decay_steps : int, optional
        Number of steps over which the learning rate decays. Default is 100_000.
    alpha : float, optional
        Final multiplier for the learning rate. Default is 0.0.
    exponent : int, optional
        Exponent for the cosine decay schedule. Default is 1.
    """

    def __init__(self, init_value=0.001, decay_steps=100_000, alpha=0.0, exponent=1):
        self.init_value = init_value
        self.decay_steps = decay_steps
        self.alpha = alpha
        self.exponent = exponent
        super().__init__()

    def create_optimizer(self) -> _NumPyroOptim:
        """
        Create and return a NumPyro optimizer with cosine decay schedule.

        Returns
        -------
        _NumPyroOptim
            An instance of a NumPyro optimizer with cosine decay schedule.
        """
        scheduler = optax.cosine_decay_schedule(
            init_value=self.init_value,
            decay_steps=self.decay_steps,
            alpha=self.alpha,
            exponent=self.exponent,
        )

        opt = optax_to_numpyro(
            optax.chain(
                optax.scale_by_adam(),
                optax.scale_by_schedule(scheduler),
                optax.scale(-1.0),
            )
        )
        return opt

create_optimizer()

Create and return a NumPyro optimizer with cosine decay schedule.

Returns:

Type Description
_NumPyroOptim

An instance of a NumPyro optimizer with cosine decay schedule.
Source code in src/prophetverse/engine/optimizer/optimizer.py 
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
def create_optimizer(self) -> _NumPyroOptim:
    """
    Create and return a NumPyro optimizer with cosine decay schedule.

    Returns
    -------
    _NumPyroOptim
        An instance of a NumPyro optimizer with cosine decay schedule.
    """
    scheduler = optax.cosine_decay_schedule(
        init_value=self.init_value,
        decay_steps=self.decay_steps,
        alpha=self.alpha,
        exponent=self.exponent,
    )

    opt = optax_to_numpyro(
        optax.chain(
            optax.scale_by_adam(),
            optax.scale_by_schedule(scheduler),
            optax.scale(-1.0),
        )
    )
    return opt

LBFGSSolver

Bases: BaseOptimizer

L-BFGS solver.

This solver is more practical than other optimizers since it usually does not require tuning of hyperparameters to get better estimates.

If your model does not converge with the default hyperparameters, you can try increasing memory_size, max_linesearch_steps, or setting a larger number of steps.

Parameters:

Name Type Description Default
gtol float

Gradient tolerance for stopping criterion.

 1e-6
tol float

Function value tolerance for stopping criterion.

 1e-6
learning_rate float

Initial learning rate.

 1e-3
memory_size int

Memory size for L-BFGS updates.

 10
scale_init_precond bool

Whether to scale the initial preconditioner.

 True
max_linesearch_steps int

Maximum number of line search steps.

 20
initial_guess_strategy str

Strategy for the initial line search step size guess.

 "one"
max_learning_rate float

Maximum allowed learning rate during line search.

 None
linesearch_tol float

Tolerance parameter for line search.

 0
increase_factor float

Factor by which to increase step size during line search when conditions are met.

 2
slope_rtol float

Relative tolerance for slope in the line search.

 0.0001
curv_rtol float

Curvature condition tolerance for line search.

 0.9
approx_dec_rtol float

Approximate decrease tolerance for line search.

 0.000001
stepsize_precision float

Stepsize precision tolerance.

 1e5
Source code in src/prophetverse/engine/optimizer/optimizer.py 
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
class LBFGSSolver(BaseOptimizer):
    """
    L-BFGS solver.

    This solver is more practical than other optimizers since it usually does not
    require tuning of hyperparameters to get better estimates.

    If your model does not converge with the default hyperparameters, you can try
    increasing `memory_size`, `max_linesearch_steps`, or setting a larger number
    of steps.

    Parameters
    ----------
    gtol : float, default=1e-6
        Gradient tolerance for stopping criterion.
    tol : float, default=1e-6
        Function value tolerance for stopping criterion.
    learning_rate : float, default=1e-3
        Initial learning rate.
    memory_size : int, default=10
        Memory size for L-BFGS updates.
    scale_init_precond : bool, default=True
        Whether to scale the initial preconditioner.
    max_linesearch_steps : int, default=20
        Maximum number of line search steps.
    initial_guess_strategy : str, default="one"
        Strategy for the initial line search step size guess.
    max_learning_rate : float, optional
        Maximum allowed learning rate during line search.
    linesearch_tol : float, default=0
        Tolerance parameter for line search.
    increase_factor : float, default=2
        Factor by which to increase step size during line search when conditions are
        met.
    slope_rtol : float, default=0.0001
        Relative tolerance for slope in the line search.
    curv_rtol : float, default=0.9
        Curvature condition tolerance for line search.
    approx_dec_rtol : float, default=0.000001
        Approximate decrease tolerance for line search.
    stepsize_precision : float, default=1e5
        Stepsize precision tolerance.
    """

    _tags = {
        "is_solver": True,
    }

    def __init__(
        self,
        gtol: float = 1e-6,
        tol: float = -jnp.inf,
        learning_rate=1e-3,
        memory_size=50,
        scale_init_precond=True,
        # linesearch
        max_linesearch_steps=50,
        initial_guess_strategy="one",
        max_learning_rate=None,
        linesearch_tol=0,
        increase_factor=2,
        slope_rtol=0.0001,
        curv_rtol=0.9,
        approx_dec_rtol=0.000001,
        stepsize_precision=1e5,
    ):
        self.gtol = gtol
        self.tol = tol
        self.learning_rate = learning_rate
        self.memory_size = memory_size
        self.scale_init_precond = scale_init_precond

        # Linesearch
        self.max_linesearch_steps = max_linesearch_steps
        self.initial_guess_strategy = initial_guess_strategy
        self.max_learning_rate = max_learning_rate
        self.linesearch_tol = linesearch_tol
        self.increase_factor = increase_factor
        self.slope_rtol = slope_rtol
        self.curv_rtol = curv_rtol
        self.approx_dec_rtol = approx_dec_rtol
        self.stepsize_precision = stepsize_precision

        super().__init__()

        self.max_iter = 1000

    def set_max_iter(self, max_iter: int):
        """Set the maximum number of iterations for the solver."""
        new_obj = self.clone()
        new_obj.max_iter = max_iter
        return new_obj

    def create_optimizer(self):
        """Create and return a NumPyro L-BFGS solver instance."""
        return LBFGS(
            max_iter=self.max_iter,
            tol=self.tol,
            gtol=self.gtol,
            learning_rate=self.learning_rate,
            memory_size=self.memory_size,
            scale_init_precond=self.scale_init_precond,
            # linesearch
            max_linesearch_steps=self.max_linesearch_steps,
            initial_guess_strategy="one",
            max_learning_rate=self.max_learning_rate,
            linesearch_tol=self.linesearch_tol,
            increase_factor=self.increase_factor,
            slope_rtol=self.slope_rtol,
            curv_rtol=self.curv_rtol,
            approx_dec_rtol=self.approx_dec_rtol,
            stepsize_precision=self.stepsize_precision,
        )

create_optimizer()

Create and return a NumPyro L-BFGS solver instance.

Source code in src/prophetverse/engine/optimizer/optimizer.py 
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
def create_optimizer(self):
    """Create and return a NumPyro L-BFGS solver instance."""
    return LBFGS(
        max_iter=self.max_iter,
        tol=self.tol,
        gtol=self.gtol,
        learning_rate=self.learning_rate,
        memory_size=self.memory_size,
        scale_init_precond=self.scale_init_precond,
        # linesearch
        max_linesearch_steps=self.max_linesearch_steps,
        initial_guess_strategy="one",
        max_learning_rate=self.max_learning_rate,
        linesearch_tol=self.linesearch_tol,
        increase_factor=self.increase_factor,
        slope_rtol=self.slope_rtol,
        curv_rtol=self.curv_rtol,
        approx_dec_rtol=self.approx_dec_rtol,
        stepsize_precision=self.stepsize_precision,
    )

set_max_iter(max_iter)

Set the maximum number of iterations for the solver.

Source code in src/prophetverse/engine/optimizer/optimizer.py 
231
232
233
234
235
def set_max_iter(self, max_iter: int):
    """Set the maximum number of iterations for the solver."""
    new_obj = self.clone()
    new_obj.max_iter = max_iter
    return new_obj