Changelog#
All notable changes to this project will be documented in this file.
The format is based on Keep a Changelog (+ the Migration Guide), and this project adheres to Semantic Versioning.
[0.21.0] - 2023-07-25#
Changed#
BREAKING CHANGE: Fixing the multiprocessing issues for the Optuna wrapper in 0.20.0, has uncovered another issue with the implementation. Namely, that the current implementation would reuse the same storage when called multiple times. This leads to issues when calling the optimization in a loop (e.g. within a cross validation), as the final optimized parameters leak from one iteration to the next. To solve this issue,
get_study_params
now gets aunique_id
parameter, that should be used to create a unique storage location for each call. We also throw an error, if the selected storage already contains a study. (#81)Below you can find the updated migration guide:
OLD (0.19.0)
def create_study(): return optuna.create_study(sampler=RandomSampler(seed=42), storage="sqlite:///example.db") OptunaSearch(..., create_study=create_study, ...)
NEW (0.21.0)
def get_study_params(seed: int, unique_id: str): return dict(sampler=RandomSampler(seed=seed), storage=f"sqlite:///example_{unique_id}.db") OptunaSearch(..., get_study_params=get_study_params, random_seed=42, ...)
Note: If you were using in-memory storage, you don’t need to use the
unique_id
parameter.
[0.20.0] - 2023-07-24#
Changed#
BREAKING CHANGE: The way how all Optuna based optimizer work has been changed. Instead of passing a function, that returns a study, you now need to pass a function that returns the parameters of a study. Creating the study is now handled by tpcp internally to avoid issues with multiprocessing. This results in two changes. The parameter name for all optuna pipelines has changed from
create_study
toget_study_params
. Further, the expected call signature changed, asget_study_params
now gets a seed as argument. This seed should be used to initialize the random number generator of the sampler and pruner of a study to ensure that each process gets a different seed and sampling process. (#80)To migrate your code, you need to change the following:
OLD:
def create_study(): return optuna.create_study(sampler=RandomSampler(seed=42)) OptunaSearch(..., create_study=create_study, ...)
NEW:
def get_study_params(seed: int): return dict(sampler=RandomSampler(seed=seed)) OptunaSearch(..., get_study_params=get_study_params, random_seed=42, ...)
[0.19.0] - 2023-07-06#
Added#
All optimization methods that do complicated loops (over parameters or CV-Folds) now raise new custom error messages (OptimizationError and TestError) if they encounter an error. These new errors have further information in which iteration of the loop the error occurred and should make it easier to debug issues.
When a scorer fails, we now print the name (i.e. the group) of the datapoint that caused the error. This should make it easier to debug issues with the scorer.
Changed#
We dropped support for joblib<0.13.0. due to some changes in the API. We only support the new API now, which allowed us to simplify some of the multiprocessing code.
[0.18.0] - 2023-04-13#
Fixed#
When
super().__init__()
is called before all parameters of the child class are initialized, we don’t get an error anymore. Now all classes remember their parameters when they are defined and don’t try to access parameters that are not defined in their own init. (#69)
Changed#
Validation is now performed recursively on all subclasses. Note like before validation is still only performed once per class. But with this change, we can also validate base classes that are not used directly. (#70)
Added#
We validate now, if a child class implements all the parameters of its parent class. While not strictly necessary, this is a sign of bad design, if not done. It could also lead to issues with tpcps validation logic. (#70)
It is now possible to hook into the validation and perform custom validation of classes. (#70)
The dataset class now activly triggers validation and checks if the dataset subclass implements
groupby_cols
andsubset_index
.
[0.17.0] - 2023-03-24#
Added#
We now have a workaround for global configuration that should be passed to worker processes when using multiprocessing. This is a workaround to a joblib issue and is quite hacky. If you want to use this feature with your own configs you can use
tpcp.parallel.register_global_parallel_callback
. If you need to write your own parallel loop using joblib, you need to usetpcp.parallel.delayed
instead ofjoblib.delayed
. (#65)
[0.16.0] - 2023-03-21#
Changed#
We are now raising an explicit ValidationError, if any of the parameters of a class have a trailing underscore, as this syntax is reserved for result objects. (#63)
Added#
The Optuna search methods have new parameter called
eval_str_paras
that allows to automatically turn categorical string parameters into python objects. This can be usefull, if you want to select between complex objects and not just strings in your parameter search. To use this in your subclasses, you need to wrap the use oftrial.params
withself.sanitize_params(trial.params)
. (#64)
[0.15.0] - 2023-02-07#
Added#
GridSearch and GridSearchCV now have the option to pick the parameters with the lowest score if desired. This is useful, if your metric represents an error and you want to pick the parameters that minimize the error. To do that, you can set the
return_optimized
parameter of these classes to the name of metric prefixed with a-
. (e.g.return_optimized="-rmse"
). (#61)A new Optimization Algorithm called
OptunaSearch
. This is a (nearly) drop-in replacement forGridSearch
using Optuna under the hood. It can be used to quickly implement parameter searches with different samplers for non-optimizable algorithms. (#57)
Changed#
In this release we added multiple safe guards against edge cases related to non-deterministic dataset indices. Most of these changes are internal and should not require any changes to your code. Still, they don’t solve all edge cases. Make sure your index is deterministic ;) (#62)
The index of datasets objects are now cached The first time
create_index
is called, the index is stored insubset_index
and used for subsequent calls. This should avoid the overhead of creating the index every time (in particular if the index creation requires IO). It should also help to avoid edge cases, wherecreate_index
is called multiple times and returns different results.When
create_index
of a dataset is called, we actually call it twice now, to check if the index is deterministic. Having a non-deterministic index can lead to hard to debug issues, so we want to make sure that this is not the case. It could still be that the index changes when using a different machine/OS (which is not ideal for reproducibility), but this should prevent most cases leading to strange issues.Internally, the
_optimize_and_score
method now directly gets the subset of the dataset, instead of the indices of the train and test set. This should again help to avoid issues, where the index of the dataset changes between calculating the splits and actually retrieving the data.
[0.14.0] - 2023-02-01#
Added#
Custom Aggregators can now use the
RETURN_RAW_SCORES
class variable to specify, if their raw input scores should be returned. (#58)
Fixed#
GridSearch and GridSearchCV now correctly handle custom aggregators that return scores with new names. (#58)
When using the
create_group_labels
method on dataset with multiple groupby columns, the method returned a list of tuples. This caused issues withGroupKFold
, as the method internally flattens the list of tuples. To avoid this, the method now return a list of strings. The respective string is simply the string representation of the tuple that was returned before. (#59)The fix provided in 0.12.1 to fix hashing of objects defined in the
__main__
module was only partially working. When the object in question was nested in another object, the hashing would still fail. This is hopefully now fixed for good. (#60)
[0.13.0] - 2023-01-11#
Changed#
Some improvements to the documentation
Added#
Added an option to the optuna search to use multiprocessing using the suggestions made in https://github.com/optuna/optuna/issues/2862 . This has not been extensively tested in real projects. Therefore, use with care and please report any issues you encounter.
Deprecated#
Fully deprecated the
_skip_validation
parameter for base classes, which was briefly used for some old versions.
[0.12.2] - 2022-12-14#
Fixed#
The previous for fixing hashing of objects defined in the
__main__
module was not working This should now be fixed.
[0.12.1] - 2022-12-14#
Changed#
The
safe_run
method did unintentionally double-wrap the run method, if it already had amake_action_safe
decorator. This is now fixed.
Fixed#
Under certain conditions hashing of an object defined in the
__main__
module failed. This release implements a workaround for this issue, that should hopefully resolve most cases.
[0.12.0] - 2022-11-15#
Added#
Added the concept of the
self_optimize_with_info
method that can be implemented instead or in addition to theself_optimize
method. This method should be used when an optimize method requires to return/output additional information besides the main result and is supported by theOptimize
wrapper. (#49)Added a new method called
__clone_param__
that gives a class control over how params are cloned. This can be helpful, if for some reason objects don’t behave well with deepcopy.Added a new method called
__repr_parameters__
that gives a class control over how params are represented. This can be used to customize the representation of individual parameters in the__repr__
method.Add proper repr for
CloneFactory
[0.11.0] - 2022-10-17#
Added#
Support for Optuna >3.0
Example on how to use
attrs
anddataclass
with tpcpAdded versions for
Dataset
andCustomOptunaOptimize
that work with dataclasses and attrs.Added first class support for composite objects (e.g. objects that need a list of other objects as parameters). This is basically sklearn pipelines with fewer restrictions (#48).
Changed#
CustomOptunaOptimize
now expects a callable to define the study, instead of taking a study object itself. This ensures that the study objects can be independent when the class is called as part ofcross_validate
.Parameters are only validated when
get_params
is called. This reduces the reliance on__init_subclass__
and that we correctly wrap the init. This makes it possible to easier supportattrs
anddataclass
[0.10.0] - 2022-09-09#
Changed#
Reworked once again when and how annotations for tpcp classes are processed. Processing is now delayed until you are actually using the annotations (i.e. as part of the “safe wrappers”). The only user facing change is that the chance of running into edge cases is lower and that
__field_annotations__
is now only available on class instances and not the class itself anymore.
[0.9.1] - 2022-09-08#
Fixed#
Classes without init can now pass the tpcp checks
Added#
You can nest parameter annotations into
ClassVar
and they will still be processed. This is helpful when using dataclasses and annotating nested parameter values.
[0.9.0] - 2022-08-11#
This release drops Python 3.7 support!
Added#
Bunch new high-level documentation
Added submission version of JOSS paper
Changed#
The
aggregate
methods of custom aggregators now gets the list of datapoints in additions to the scores. Both parameters are now passed as keyword only arguments.
[0.8.0] - 2022-08-09#
Added#
An example on how to use the
dataclass
decorator with tpcp classes. (#41)In case you need complex aggregations of scores across data points, you can now wrap the return values of score functions in custom
Aggregators
. The best plac eto learn about this feature is the new “Custom Scorer” example. (#42)All cross_validation based methods now have a new parameter called
mock_labels
. This can be used to provide a “y” value to the split method of a sklearn-cv splitter. This is required e.g. for Stratified KFold splitters. (#43)
Changed#
Most of the class proccesing and sanity checks now happens in the init (or rather a post init hook) instead of during class initialisation. This increases the chance for some edge cases, but allows to post-process classes, before tpcp checks are run. Most importantly, it allows the use of the
dataclass
decorator in combination with tpcp classes. For the “enduser”, this change will have minimal impact. Only, if you relied on accessing special tpcp class parameters before the class (e.g.__field_annotations__
) was initialised, you will get an error now. Other than that, you will only notice a very slight overhead on class initialisation, as we know need to run some basic checks when you call the init orget_params
. (#41)The API of the Scorer class was modified. In case you used custom Scorer before, they will likely not work anymore. Further, we removed the
error_score
parameter from the Scorer and all related methods, that forwarded this parameter (e.g.GridSearch
). Error that occur in the score function will now always be raised! If you need special handling of error cases, handle them in your error function yourself (i.e. using try-except). This gives more granular control and makes the implementation of the expected score function returns much easier on thetpcp
side. (#42)
[0.7.0] - 2022-06-23#
Added#
The
Dataset
class now has a new parametergroup
, which will return the group/row information, if there is only a single group/row left in the dataset. This parameter returns either a string or a namedtuple to make it easy to access the group/row information.The
Dataset.groups
parameter now returns a list of namedtuples when it previously returned a list of normal tuples.New
is_single_group
andassert_is_single_group
methods for theDataset
class are added. They are shortcuts for callingself.is_single(groupby_cols=self.groupby_cols)
andself.assert_is_single(groupby_cols=self.groupby_cols)
.
Removed#
We removed the
OptimizableAlgorithm
base class, as it is not really useful. We recommend implementing your own base class or mixin if you are implementing a set of algorithms that need a normal and an optimizable version.
[0.6.3] - 2022-05-31#
It is now possible to use namedtuples as parameters and they are correctly cloned (#39)
[0.6.2] - 2022-04-21#
The poetry lockfiles are now committed to the repository.
Some dependencies are updated
[0.6.1] - 2022-04-05#
Changed#
[0.6.0] - 2022-04-04#
Added#
A new class to wrap the optimization framework Optuna.
CustomOptunaOptimize
can be used to create custom wrapper classes for various Optuna optimizations, that play nicely withtpcp
and can be nested within tpcp operations. (#27)A new example for the
CustomOptunaOptimize
wrapper that explains how to create complex custom optimizers usingOptuna
and the new Scorer callbacks (see below) (#27)Scorer
now supports an optional callback function, which will be called after each datapoint is scored. (#29)Pipelines, Optimize objects, and
Scorer
are nowGeneric
. This improves typing (in particular with VsCode), but means a little bit more typing (pun intended), when creating new Pipelines and Optimizers (#29)Added option for scoring function to return arbitrary additional information using the
NoAgg
wrapper (#31)(experimental) Torch compatibility for hash based comparisons (e.g. in the
safe_run
wrapper). Before the wrapper would fail, with torch module subclasses, as their pickle based hashes where not consistent. We implemented a custom hash function that should solve this. For now, we will consider this feature experimental, as we are not sure if it breaks in certain use-cases. (#33)tpcp.types
now exposes a bunch of internal types that might be helpful to type custom Pipelines and Optimizers. (#34)
Changed#
The return type for the individual values in the
Scorer
class is notList[float]
instead ofnp.ndarray
. This also effects the output ofcross_validate
,GridSearch.gs_results_
andGridSearchCV.cv_results_
(#29)cf
now has “faked” return type, so that type checkers in the user code, do not complain anymore. (#29)All TypeVar Variables are now called
SomethingT
instead ofSomething_
(#34)
[0.5.0] - 2022-03-15#
Added#
The
make_optimize_safe
decorator (and hence, theOptimize
method) make use of the parameter annotations to check that only parameters marked asOptimizableParameter
are changed by theself_optimize
method. This check also supports nested parameters, in case the optimization involves optimizing nested objects. (#9)All tpcp objects now have a basic representation that is automatically generated based on their parameters (#13)
Added algo optimization and evaluation guide and improved docs overall (#26)
Added examples for all fundamental concepts (#23)
[0.4.0] - 2021-12-13#
In this release the entire core of tcpc was rewritten and a lot of the api was reimagened. This release also has no proper deprecation. This repo will the wild west until the 1.0 release
There are a multitude of changes, that I do not remember all. Here is the list of most important things:
A lot of the import paths have changed. It is likely that you need to update some imports
A new way to resolve the mutable default issue was introduced. It uses a explicit factory. To make that more clear, the name for mutable wrapper has changed from
mdf
->cf
or fromdefault
->CloneFactory
The overall class API got a lot slimmer. A bunch of methods like
get_attributes
, are now functions that can be called with the class or instance as argument. Most methods were further renamed to represent there meaning in the context oftpcp
. For example,get_attributes
->get_results
The new concept of Parameter Annotation was added. You can add type annotations with the types
Parameter
,HyperParameter
,OptimizableParameter
, orPureParameter
to a class parameter to indicate its role for the algorithm. This is in particularly helpful for optimizable Pipelines and algorithms. Wrapper methods can then use these annotations to run implementation checks and optimizations. The only method that uses these at the momement isGridSearchCV
. It now allows to setpure_parameters
toTrue
, which will collect the list of all attributes annotated withPureParameter
and use this information for performance optimization. Before, you needed to specify the list of pure parameters manually.Some core classes where renamed:
BaseAlgorithm
->Algorithm
,SimplePipeline
->Pipeline
The decorators to make
run
andself_optimize
safe are now reimagened and renamed tomake_run_safe
andmake_optimize_safe