Develop

CHANGELOG.md

@@ -1,5 +1,11 @@
 # CHANGELOG
 
+## v1.1.1 (2026-04-17)
+
+### Bug fix: LBFGS default optimizer regression with PyTorch >= 2.10
+
+The default LBFGS optimizer now explicitly sets `max_eval=25`. This restores correct training behavior after [pytorch/pytorch#161488](https://github.com/pytorch/pytorch/pull/161488) (shipped in PyTorch 2.10) fixed a bug where `max_eval` was silently ignored by the strong Wolfe line search. Prior to that fix, `max_eval` defaulted to `2` (from `max_iter * 1.25 + 1` with `max_iter=1`), which was effectively ignored — the line search ran freely. After the fix, the cap was correctly enforced, causing the optimizer to converge after only a handful of epochs with a poor loss.
+
 ## v1.1.0 (2026-04-06)
 
 ### Refactor of `Process` API and internals

README.rst

@@ -20,6 +20,6 @@ Getting Started
 
     pip install git+https://github.com/onesixsolutions/torchcast.git#egg=torchcast
 
-``torchcast`` requires Python >= 3.8 and PyTorch >= 1.8.
+``torchcast`` requires Python >= 3.9 and PyTorch >= 1.12.
 
 See the `Quick Start <https://docs.strong.io/torchcast/quick_start.html>`_ for a simple example that will get you up to speed, or delve into the `examples <https://docs.strong.io/torchcast/examples/examples.html>`_ or the `API <https://docs.strong.io/torchcast/api/api.html>`_.

docs/api/utils.rst

@@ -4,4 +4,4 @@ Utils
 .. include:: ../macros.hrst
 
 .. automodule:: torchcast.utils
-   :members: TimeSeriesDataset, TimeSeriesDataLoader, add_season_features, complete_times, make_baseline, SimpleTrainer, StateSpaceTrainer, SeasonalEmbeddingsTrainer, Stopping
+   :members: TimeSeriesDataset, TimeSeriesDataLoader, add_season_features, complete_times, make_baseline, SimpleTrainer, StateSpaceTrainer, ModelMatEmbeddingsTrainer, Stopping

torchcast/__init__.py

@@ -1 +1 @@
-__version__ = '1.1.0'
+__version__ = '1.1.1'

torchcast/process/regression.py

@@ -107,7 +107,7 @@ def __init__(self,
     def _init_state_elements(self,
                              predictors: Sequence[str],
                              fixed: Sequence[str]) -> Sequence[StateElement]:
-        assert 'ceiling' not in predictors, f"`ceiling` is a reserved name for {type(self).__name__}"
+        assert '_ceiling' not in predictors, f"`_ceiling` is a reserved name for {type(self).__name__}"
         coefs = [
             StateElement(name=p, measure_multi=None, has_process_variance=p not in fixed)
             for p in predictors
@@ -137,7 +137,7 @@ def get_measured_mean(self, mean: torch.Tensor, time: int, cache: dict) -> torch
         coefs = mean[:, :self.num_predictors]
         ceiling = mean[:, self.num_predictors]
         cache['yhat'] = (X * coefs).sum(-1)
-        return cache['yhat'] - torch.log1p(torch.exp(cache['yhat'] - ceiling))
+        return cache['yhat'] - torch.nn.functional.softplus(cache['yhat'] - ceiling)
 
     def get_measurement_jacobian(self, mean: torch.Tensor, time: int, cache: dict) -> torch.Tensor:
         # TODO: reparameterize

torchcast/state_space/predictions.py

@@ -368,6 +368,12 @@ def _to_dataframe(self,
                     if m not in by_measure:
                         continue
                     actuals[m] = tens[..., mgroup.index(m)]
+            missing = set(by_measure) - set(dataset.all_measures)
+            if missing:
+                warn(
+                    f"The following measures in your model are not present in your dataset, please double-check that "
+                    f"the names you passed to the dataset match the `measures` you passed to the model:\n{missing}"
+                )
         out = []
         times = TimeSeriesDataset.get_dataset_times(
             dataset.start_offsets, num_timesteps=self.state_means.shape[1], dt_unit=dataset.dt_unit

torchcast/state_space/state_space.py

@@ -379,9 +379,10 @@ def fit(self,
 
         :param y: A tensor containing the batch of time-series(es), see :func:`StateSpaceModel.forward()`.
         :param optimizer: The optimizer to use. Can also pass a function which takes the parameters and returns an
-         optimizer instance. Default is :class:`torch.optim.LBFGS` with ``(line_search_fn='strong_wolfe', max_iter=1)``.
+         optimizer instance. Default is :class:`torch.optim.LBFGS` with
+         ``(line_search_fn='strong_wolfe', max_iter=1, max_eval=25)``.
         :param stopping: Controls stopping/convergence rules; should be a :class:`torchcast.utils.Stopping` instance, or
-         a dict of keyword-args to one. Example: ``stopping={'abstol' : .001, 'monitor' : 'params'}``
+         a dict of keyword-args to one. Example: ``stopping={'abstol' : .001, 'monitor_params' : True}``
         :param verbose: If True (default) will print the loss and epoch.
         :param callbacks: A list of functions that will be called at the end of each epoch, which take the current
          epoch's loss value.
@@ -402,9 +403,12 @@ def fit(self,
             optimizer = optimizer([p for p in self.parameters() if p.requires_grad])
         elif optimizer is None:
             optimizer = torch.optim.LBFGS(
+                # only pass params that require grad:
                 [p for p in self.parameters() if p.requires_grad],
-                # https://discuss.pytorch.org/t/unclear-purpose-of-max-iter-kwarg-in-the-lbfgs-optimizer/65695/4
+                # see https://discuss.pytorch.org/t/unclear-purpose-of-max-iter-kwarg-in-the-lbfgs-optimizer/65695/4
                 max_iter=1,
+                # see https://github.com/pytorch/pytorch/pull/161488
+                max_eval=25,
                 line_search_fn='strong_wolfe'
             )