invrs-opt 0.3.2__py3-none-any.whl → 0.5.0__py3-none-any.whl

invrs_opt/__init__.py

@@ -3,7 +3,7 @@
 Copyright (c) 2023 The INVRS-IO authors.
 """
 
-__version__ = "v0.3.2"
+__version__ = "v0.5.0"
 __author__ = "Martin F. Schubert <mfschubert@gmail.com>"
 
 from invrs_opt.lbfgsb.lbfgsb import density_lbfgsb as density_lbfgsb

invrs_opt/experimental/client.py

@@ -0,0 +1,152 @@
+"""Defines basic client optimizers for use with an optimization service.
+
+Copyright (c) 2023 The INVRS-IO authors.
+"""
+
+import json
+import requests
+import time
+from typing import Any, Dict, Optional
+
+from totypes import json_utils
+
+from invrs_opt import base
+from invrs_opt.experimental import labels
+
+
+PyTree = Any
+StateToken = str
+
+SESSION = None
+SERVER_ADDRESS = None
+
+
+def login(server_address: str) -> None:
+    """Set the global server address and create a requests session."""
+    global SESSION
+    global SERVER_ADDRESS
+    SESSION = requests.Session()
+    SERVER_ADDRESS = server_address
+
+
+def optimizer_client(
+    algorithm: str,
+    hparams: Dict[str, Any],
+    server_address: Optional[str],
+    session: Optional[requests.Session],
+) -> base.Optimizer:
+    """Generic optimizer class."""
+
+    if server_address is None:
+        if SERVER_ADDRESS is None:
+            raise ValueError(
+                "Argument `server_address` and the global `SERVER_ADDRESS` cannot "
+                "both be `None`. Use the `login` method to set the global, or "
+                "explicitly provide a value."
+            )
+    if session is None:
+        if SESSION is None:
+            raise ValueError(
+                "Argument `session` and the global `SESSION` cannot "
+                "both be `None`. Use the `login` method to set the global, or "
+                "explicitly provide a value."
+            )
+        session = SESSION
+
+    opt_config = {
+        labels.ALGORITHM: algorithm,
+        labels.HPARAMS: hparams,
+    }
+
+    def init_fn(params: PyTree) -> StateToken:
+        """Handles 'init' requests."""
+        serialized_data = json_utils.json_from_pytree(
+            dict(opt_config=opt_config, data={"params": params})
+        )
+        post_response = session.post(
+            f"{SERVER_ADDRESS}/{labels.ROUTE_INIT}/", data=serialized_data
+        )
+
+        if not post_response.status_code == 200:
+            raise requests.RequestException(post_response.text)
+        response = json.loads(post_response.text)
+        new_state_token: str = response[labels.STATE_TOKEN]
+        return new_state_token
+
+    def update_fn(
+        *,
+        grad: PyTree,
+        value: float,
+        params: PyTree,
+        state: StateToken,
+    ) -> StateToken:
+        """Handles 'update' requests."""
+        state_token = state
+        del state
+        serialized_data = json_utils.json_from_pytree(
+            {
+                labels.OPT_CONFIG: opt_config,
+                labels.DATA: {
+                    labels.PARAMS: params,
+                    labels.VALUE: value,
+                    labels.GRAD: grad,
+                    labels.STATE_TOKEN: state_token,
+                },
+            }
+        )
+        post_response = session.post(
+            f"{SERVER_ADDRESS}/{labels.ROUTE_UPDATE}/{state_token}/",
+            data=serialized_data,
+        )
+
+        if not post_response.status_code == 200:
+            raise requests.RequestException(post_response.text)
+        response = json.loads(post_response.text)
+        new_state_token: str = response[labels.STATE_TOKEN]
+        return new_state_token
+
+    def params_fn(
+        state: StateToken,
+        timeout: float = 60.0,
+        poll_interval: float = 0.1,
+    ) -> PyTree:
+        """Handles 'params' requests."""
+        state_token = state
+        del state
+        assert timeout >= poll_interval
+        start_time = time.time()
+        while time.time() < start_time + timeout:
+            get_response = session.get(
+                f"{SERVER_ADDRESS}/{labels.ROUTE_PARAMS}/{state_token}"
+            )
+            if get_response.status_code == 200:
+                break
+            elif get_response.status_code == 404 and get_response.text.endswith(
+                labels.MESSAGE_STATE_NOT_READY.format(state_token)
+            ):
+                time.sleep(poll_interval)
+            else:
+                raise requests.RequestException(get_response.text)
+
+        if not get_response.status_code == 200:
+            raise requests.Timeout("Timed out while waiting for params.")
+        response = json.loads(get_response.text)
+        return json_utils.pytree_from_json(response[labels.PARAMS])
+
+    return base.Optimizer(init=init_fn, update=update_fn, params=params_fn)
+
+
+# -----------------------------------------------------------------------------
+# Specific optimizers implemented here.
+# -----------------------------------------------------------------------------
+
+
+def lbfgsb(maxcor: int = 20, line_search_max_steps: int = 100) -> base.Optimizer:
+    """Optimizer implementing the L-BFGS-B scheme."""
+    hparams = {
+        "maxcor": maxcor,
+        "line_search_max_steps": line_search_max_steps,
+    }
+    return optimizer_client(
+        algorithm="lbfgsb", hparams=hparams, server_address=None, session=None
+    )

invrs_opt/experimental/labels.py

@@ -0,0 +1,23 @@
+"""Defines labels and messages used in the context of an optimization service.
+
+Copyright (c) 2023 The INVRS-IO authors.
+"""
+
+VALUE = "value"
+GRAD = "grad"
+PARAMS = "params"
+STATE_TOKEN = "state_token"
+MESSAGE = "message"
+
+OPT_CONFIG = "opt_config"
+ALGORITHM = "algorithm"
+HPARAMS = "hparams"
+DATA = "data"
+
+ROUTE_INIT = "init"
+ROUTE_UPDATE = "update"
+ROUTE_PARAMS = "params"
+
+MESSAGE_STATE_NOT_KNOWN = "State token {} was not recognized."
+MESSAGE_STATE_NOT_READY = "State for token {} is not ready."
+MESSAGE_STATE_NOT_VALID = "State for token {} is not valid."

invrs_opt/lbfgsb/lbfgsb.py

@@ -29,16 +29,16 @@ LbfgsbState = Tuple[PyTree, PyTree, JaxLbfgsbDict]
 # Task message prefixes for the underlying L-BFGS-B implementation.
 TASK_START = b"START"
 TASK_FG = b"FG"
+TASK_CONVERGED = b"CONVERGENCE"
 
-# Parameters which configure the state update step.
 UPDATE_IPRINT = -1
-UPDATE_PGTOL = 0.0
-UPDATE_FACTR = 0.0
 
 # Maximum value for the `maxcor` parameter in the L-BFGS-B scheme.
 MAXCOR_MAX_VALUE = 100
 MAXCOR_DEFAULT = 20
 LINE_SEARCH_MAX_STEPS_DEFAULT = 100
+FTOL_DEFAULT = 0.0
+GTOL_DEFAULT = 0.0
 
 # Maps bound scenarios to integers.
 BOUNDS_MAP: Dict[Tuple[bool, bool], int] = {
@@ -54,6 +54,8 @@ FORTRAN_INT = scipy_lbfgsb.types.intvar.dtype
 def lbfgsb(
     maxcor: int = MAXCOR_DEFAULT,
     line_search_max_steps: int = LINE_SEARCH_MAX_STEPS_DEFAULT,
+    ftol: float = FTOL_DEFAULT,
+    gtol: float = GTOL_DEFAULT,
 ) -> base.Optimizer:
     """Return an optimizer implementing the standard L-BFGS-B algorithm.
 
@@ -85,10 +87,17 @@ def lbfgsb(
     While the algorithm can work with pytrees of jax arrays, numpy arrays can
     also be used. Thus, e.g. the optimizer can directly be used with autograd.
 
+    When the optimization has converged (according to `ftol` or `gtol` criteria), the
+    optimizer simply returns the parameters which obtained the converged result. The
+    convergence can be queried by `is_converged(state)`.
+
     Args:
         maxcor: The maximum number of variable metric corrections used to define
             the limited memory matrix, in the L-BFGS-B scheme.
         line_search_max_steps: The maximum number of steps in the line search.
+        ftol: Tolerance for stopping criteria based on function values. See scipy
+            documentation for details.
+        gtol: Tolerance for stopping criteria based on gradient.
 
     Returns:
         The `base.Optimizer`.
@@ -96,6 +105,8 @@ def lbfgsb(
     return transformed_lbfgsb(
         maxcor=maxcor,
         line_search_max_steps=line_search_max_steps,
+        ftol=ftol,
+        gtol=gtol,
         transform_fn=lambda x: x,
         initialize_latent_fn=lambda x: x,
     )
@@ -105,6 +116,8 @@ def density_lbfgsb(
     beta: float,
     maxcor: int = MAXCOR_DEFAULT,
     line_search_max_steps: int = LINE_SEARCH_MAX_STEPS_DEFAULT,
+    ftol: float = FTOL_DEFAULT,
+    gtol: float = GTOL_DEFAULT,
 ) -> base.Optimizer:
     """Return an L-BFGS-B optimizer with additional transforms for density arrays.
 
@@ -117,11 +130,18 @@ def density_lbfgsb(
     and spacing parameters of the `DensityArray2D`. Where the bounds differ, the
     density is scaled before the transform is applied, and then unscaled afterwards.
 
+    When the optimization has converged (according to `ftol` or `gtol` criteria), the
+    optimizer simply returns the parameters which obtained the converged result. The
+    convergence can be queried by `is_converged(state)`.
+
     Args:
         beta: Determines the steepness of the thresholding.
         maxcor: The maximum number of variable metric corrections used to define
             the limited memory matrix, in the L-BFGS-B scheme.
         line_search_max_steps: The maximum number of steps in the line search.
+        ftol: Tolerance for stopping criteria based on function values. See scipy
+            documentation for details.
+        gtol: Tolerance for stopping criteria based on gradient.
 
     Returns:
         The `base.Optimizer`.
@@ -164,6 +184,8 @@ def density_lbfgsb(
     return transformed_lbfgsb(
         maxcor=maxcor,
         line_search_max_steps=line_search_max_steps,
+        ftol=ftol,
+        gtol=gtol,
         transform_fn=transform_fn,
         initialize_latent_fn=initialize_latent_fn,
     )
@@ -172,6 +194,8 @@ def density_lbfgsb(
 def transformed_lbfgsb(
     maxcor: int,
     line_search_max_steps: int,
+    ftol: float,
+    gtol: float,
     transform_fn: Callable[[PyTree], PyTree],
     initialize_latent_fn: Callable[[PyTree], PyTree],
 ) -> base.Optimizer:
@@ -182,10 +206,17 @@ def transformed_lbfgsb(
     `transform_fn`. In the simple case where this is just `lambda x: x` (i.e.
     the identity), this is equivalent to the standard L-BFGS-B algorithm.
 
+    When the optimization has converged (according to `ftol` or `gtol` criteria), the
+    optimizer simply returns the parameters which obtained the converged result. The
+    convergence can be queried by `is_converged(state)`.
+
     Args:
         maxcor: The maximum number of variable metric corrections used to define
             the limited memory matrix, in the L-BFGS-B scheme.
         line_search_max_steps: The maximum number of steps in the line search.
+        ftol: Tolerance for stopping criteria based on function values. See scipy
+            documentation for details.
+        gtol: Tolerance for stopping criteria based on gradient.
         transform_fn: Function which transforms the internal latent parameters to
             the parameters returned by the optimizer.
         initialize_latent_fn: Function which computes the initial latent parameters
@@ -218,6 +249,8 @@ def transformed_lbfgsb(
                 upper_bound=_bound_for_params(upper_bound, params),
                 maxcor=maxcor,
                 line_search_max_steps=line_search_max_steps,
+                ftol=ftol,
+                gtol=gtol,
             )
             latent_params = _to_pytree(scipy_lbfgsb_state.x, params)
             return latent_params, scipy_lbfgsb_state.to_jax()
@@ -261,10 +294,12 @@ def transformed_lbfgsb(
             flat_latent_params = jnp.asarray(scipy_lbfgsb_state.x)
             return flat_latent_params, scipy_lbfgsb_state.to_jax()
 
-        params, latent_params, jax_lbfgsb_state = state
+        _, latent_params, jax_lbfgsb_state = state
         _, vjp_fn = jax.vjp(transform_fn, latent_params)
         (latent_grad,) = vjp_fn(grad)
-        flat_latent_grad, unflatten_fn = flatten_util.ravel_pytree(latent_grad)
+        flat_latent_grad, unflatten_fn = flatten_util.ravel_pytree(
+            latent_grad
+        )  # type: ignore[no-untyped-call]
 
         (
             flat_latent_params,
@@ -286,6 +321,11 @@ def transformed_lbfgsb(
     )
 
 
+def is_converged(state: LbfgsbState) -> jnp.ndarray:
+    """Returns `True` if the optimization has converged."""
+    return state[2]["converged"]
+
+
 # ------------------------------------------------------------------------------
 # Helper functions.
 # ------------------------------------------------------------------------------
@@ -390,8 +430,11 @@ def _example_state(params: PyTree, maxcor: int) -> PyTree:
     float_params = tree_util.tree_map(lambda x: jnp.asarray(x, dtype=float), params)
     example_jax_lbfgsb_state = dict(
         x=jnp.zeros(n, dtype=float),
+        converged=jnp.asarray(False),
         _maxcor=jnp.zeros((), dtype=int),
         _line_search_max_steps=jnp.zeros((), dtype=int),
+        _ftol=jnp.zeros((), dtype=float),
+        _gtol=jnp.zeros((), dtype=float),
         _wa=jnp.ones(_wa_size(n=n, maxcor=maxcor), dtype=float),
         _iwa=jnp.ones(n * 3, dtype=jnp.int32),  # Fortran int
         _task=jnp.zeros(59, dtype=int),
@@ -441,10 +484,13 @@ class ScipyLbfgsbState:
     """
 
     x: NDArray
+    converged: NDArray
     # Private attributes correspond to internal variables in the `scipy.optimize.
     # lbfgsb._minimize_lbfgsb` function.
     _maxcor: int
     _line_search_max_steps: int
+    _ftol: NDArray
+    _gtol: NDArray
     _wa: NDArray
     _iwa: NDArray
     _task: NDArray
@@ -474,8 +520,11 @@ class ScipyLbfgsbState:
         """Generates a dictionary of jax arrays defining the state."""
         return dict(
             x=jnp.asarray(self.x),
+            converged=jnp.asarray(self.converged),
             _maxcor=jnp.asarray(self._maxcor),
             _line_search_max_steps=jnp.asarray(self._line_search_max_steps),
+            _ftol=jnp.asarray(self._ftol),
+            _gtol=jnp.asarray(self._gtol),
             _wa=jnp.asarray(self._wa),
             _iwa=jnp.asarray(self._iwa),
             _task=_array_from_s60_str(self._task),
@@ -494,8 +543,11 @@ class ScipyLbfgsbState:
         state_dict = copy.deepcopy(state_dict)
         return ScipyLbfgsbState(
             x=onp.asarray(state_dict["x"], dtype=onp.float64),
+            converged=onp.asarray(state_dict["converged"], dtype=bool),
             _maxcor=int(state_dict["_maxcor"]),
             _line_search_max_steps=int(state_dict["_line_search_max_steps"]),
+            _ftol=onp.asarray(state_dict["_ftol"], dtype=onp.float64),
+            _gtol=onp.asarray(state_dict["_gtol"], dtype=onp.float64),
             _wa=onp.asarray(state_dict["_wa"], onp.float64),
             _iwa=onp.asarray(state_dict["_iwa"], dtype=FORTRAN_INT),
             _task=_s60_str_from_array(state_dict["_task"]),
@@ -516,6 +568,8 @@ class ScipyLbfgsbState:
         upper_bound: ElementwiseBound,
         maxcor: int,
         line_search_max_steps: int,
+        ftol: float,
+        gtol: float,
     ) -> "ScipyLbfgsbState":
         """Initializes the `ScipyLbfgsbState` for `x0`.
 
@@ -526,6 +580,9 @@ class ScipyLbfgsbState:
             maxcor: The maximum number of variable metric corrections used to define
                 the limited memory matrix, in the L-BFGS-B scheme.
             line_search_max_steps: The maximum number of steps in the line search.
+            ftol: Tolerance for stopping criteria based on function values. See scipy
+                documentation for details.
+            gtol: Tolerance for stopping criteria based on gradient.
 
         Returns:
             The `ScipyLbfgsbState`.
@@ -556,8 +613,11 @@ class ScipyLbfgsbState:
         wa_size = _wa_size(n=n, maxcor=maxcor)
         state = ScipyLbfgsbState(
             x=onp.array(x0, onp.float64),
+            converged=onp.asarray(False),
             _maxcor=maxcor,
             _line_search_max_steps=line_search_max_steps,
+            _ftol=onp.asarray(ftol, onp.float64),
+            _gtol=onp.asarray(gtol, onp.float64),
             _wa=onp.zeros(wa_size, onp.float64),
             _iwa=onp.zeros(3 * n, FORTRAN_INT),
             _task=task,
@@ -580,12 +640,14 @@ class ScipyLbfgsbState:
         grad: NDArray,
         value: NDArray,
     ) -> None:
-        """Performs an in-place update of the `ScipyLbfgsbState`.
+        """Performs an in-place update of the `ScipyLbfgsbState` if not converged.
 
         Args:
             grad: The function gradient for the current `x`.
             value: The scalar function value for the current `x`.
         """
+        if self.converged:
+            return
         if grad.shape != self.x.shape:
             raise ValueError(
                 f"`grad` must have the same shape as attribute `x`, but got shapes "
@@ -606,8 +668,8 @@ class ScipyLbfgsbState:
                 nbd=self._bound_type,
                 f=value,
                 g=grad,
-                factr=UPDATE_FACTR,
-                pgtol=UPDATE_PGTOL,
+                factr=self._ftol / onp.finfo(float).eps,
+                pgtol=self._gtol,
                 wa=self._wa,
                 iwa=self._iwa,
                 task=self._task,
@@ -619,6 +681,8 @@ class ScipyLbfgsbState:
                 maxls=self._line_search_max_steps,
             )
             task_str = self._task.tobytes()
+            if task_str.startswith(TASK_CONVERGED):
+                self.converged = onp.asarray(True)
             if task_str.startswith(TASK_FG):
                 break
 

{invrs_opt-0.3.2.dist-info → invrs_opt-0.5.0.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: invrs_opt
-Version: 0.3.2
+Version: 0.5.0
 Summary: Algorithms for inverse design
 Author-email: "Martin F. Schubert" <mfschubert@gmail.com>
 Maintainer-email: "Martin F. Schubert" <mfschubert@gmail.com>
@@ -33,8 +33,10 @@ License-File: LICENSE
 Requires-Dist: jax
 Requires-Dist: jaxlib
 Requires-Dist: numpy
+Requires-Dist: requests
 Requires-Dist: scipy
 Requires-Dist: totypes
+Requires-Dist: types-requests
 Provides-Extra: dev
 Requires-Dist: bump-my-version ; extra == 'dev'
 Requires-Dist: darglint ; extra == 'dev'
@@ -47,7 +49,7 @@ Requires-Dist: pytest-cov ; extra == 'tests'
 Requires-Dist: pytest-subtests ; extra == 'tests'
 
 # invrs-opt - Optimization algorithms for inverse design
-`v0.3.2`
+`v0.5.0`
 
 ## Overview
 

invrs_opt-0.5.0.dist-info/RECORD

@@ -0,0 +1,14 @@
+invrs_opt/__init__.py,sha256=SK2nHKhwT60br9D6W8gfJ5Kj11BR3fIdGAEVAJBgH1I,309
+invrs_opt/base.py,sha256=dSX9QkMPzI8ROxy2cFNmMwk_89eQbk0rw94CzvLPQoY,907
+invrs_opt/py.typed,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
+invrs_opt/experimental/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
+invrs_opt/experimental/client.py,sha256=td5o_YqqbcSypDrWCVrHGSoF8UxEdOLtKU0z9Dth9lA,4842
+invrs_opt/experimental/labels.py,sha256=dQDAMPyFMV6mXnMy295z8Ap205DRdVzysXny_Be8FmY,562
+invrs_opt/lbfgsb/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
+invrs_opt/lbfgsb/lbfgsb.py,sha256=59OpEYmGH3ofDv6m8KQQob5mQJei2vTcm1COh7SAbgA,27896
+invrs_opt/lbfgsb/transform.py,sha256=a_Saj9Wq4lvqCJBrg5L2Z9DZ2NVs1xqrHLqha90a9Ws,5971
+invrs_opt-0.5.0.dist-info/LICENSE,sha256=ty6jHPvpyjHy6dbhnu6aDSY05bbl2jQTjnq9u1sBCfM,1078
+invrs_opt-0.5.0.dist-info/METADATA,sha256=kV8T1uAmMnr0IDaNjq1SK8btKa_XlyHjapKOfti_MbY,3326
+invrs_opt-0.5.0.dist-info/WHEEL,sha256=GJ7t_kWBFywbagK5eo9IoUwLW6oyOeTKmQ-9iHFVNxQ,92
+invrs_opt-0.5.0.dist-info/top_level.txt,sha256=hOziS2uJ_NgwaW9yhtOfeuYnm1X7vobPBcp_6eVWKfM,10
+invrs_opt-0.5.0.dist-info/RECORD,,

{invrs_opt-0.3.2.dist-info → invrs_opt-0.5.0.dist-info}/WHEEL

@@ -1,5 +1,5 @@
 Wheel-Version: 1.0
-Generator: bdist_wheel (0.42.0)
+Generator: bdist_wheel (0.43.0)
 Root-Is-Purelib: true
 Tag: py3-none-any
 

invrs_opt-0.3.2.dist-info/RECORD

@@ -1,11 +0,0 @@
-invrs_opt/__init__.py,sha256=JG9h93Vq0AA2jX6zcqwxjA-LXOws8CJQeEbY9VQvQWQ,309
-invrs_opt/base.py,sha256=dSX9QkMPzI8ROxy2cFNmMwk_89eQbk0rw94CzvLPQoY,907
-invrs_opt/py.typed,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
-invrs_opt/lbfgsb/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
-invrs_opt/lbfgsb/lbfgsb.py,sha256=2X0GVQwMCj2beYHLiAdwoAnBIdJLPThH0Jjduz3GzHA,25080
-invrs_opt/lbfgsb/transform.py,sha256=a_Saj9Wq4lvqCJBrg5L2Z9DZ2NVs1xqrHLqha90a9Ws,5971
-invrs_opt-0.3.2.dist-info/LICENSE,sha256=ty6jHPvpyjHy6dbhnu6aDSY05bbl2jQTjnq9u1sBCfM,1078
-invrs_opt-0.3.2.dist-info/METADATA,sha256=pwrRENMlb_s1bNAeROw-cJ-nIVuqR9815r-jPNM3cEg,3272
-invrs_opt-0.3.2.dist-info/WHEEL,sha256=oiQVh_5PnQM0E3gPdiz09WCNmwiHDMaGer_elqB3coM,92
-invrs_opt-0.3.2.dist-info/top_level.txt,sha256=hOziS2uJ_NgwaW9yhtOfeuYnm1X7vobPBcp_6eVWKfM,10
-invrs_opt-0.3.2.dist-info/RECORD,,