CHANGELOG.md

v1.13.3 (2024-08-13)

Fix

• added f1 score to list of output statistics

v1.13.2 (2024-08-09)

Fix

• changed results output SD on the same row as param
• update model patience config option

v1.13.1 (2024-07-11)

Fix

• added optimal acceptance time recording to results

Refactor

• refactored run.py

v1.13.0 (2024-06-19)

Feat

• add ones init type to Weight

Fix

• remove watermark dependencies
• make watermark package optional
• changed precision to 4
• fix incorrect text display for model.config
• added assert checks for incorrect choice set numbering in dataset

Refactor

• fix argument name inconsistency

v1.12.0 (2024-03-06)

Feat

• functions.py: added gelu function

Fix

• basic.py: fix bug in null_log_likelihood_fn for givens
• dataset.py: fixed slice call when no batch_size argument is used
• dataset.py: fixed the split method to accept None as argument

Refactor

• functions.py: refactored method to covert list of utilities to tensor variables
• dataset.py: accept mixed str and TensorVariable types in _make_tensor
• basic.py: moved build_gh_fn() to BaseModel

v1.11.1 (2024-02-29)

v1.11.0 (2024-02-29)

Feat

• run: added is_training flag

Fix

• pyproject.toml: fix incorrect bump message
• added null_log_likelihood_fn to basic.py
• results: add argument for max cutoff in show_training_plot
• results.py: fixed legend placement on training plot
• run.py: add option for acceptance_method argument

Refactor

• minor refactorization

v1.10.0 (2024-02-15)

Feat

• layers.py: added new layer type LayerNorm
• expression.py: added new parameter type Gamma
• results.py: add argument offset to show_training_plot
• run: moved train and compute to pycmtensor.run for cleaniness

Fix

• TasteNet.py: changed instance check type from .DenseLayer to .Layer
• run.py: add limit for verbosity

v1.9.2 (2024-02-08)

Fix

• dataset: added split via n-count and added option for pre-shuffle
• dataset: updated scale_variable method to handle both single variables and lists of variables

v1.9.1 (2023-12-20)

Fix

• workflow: remove poetry from tests
• add missing poetry dependencies

v1.9.0 (2023-12-20)

Feat

• overhaul dependencies for conda feedstock

v1.8.1 (2023-11-28)

Fix

• basic: add lr_scheduler to compute() function

Refactor

• clean up and enhance code with codiumAI

v1.8.0 (2023-11-24)

Feat

• utils: add new function to display numbers as human readable format
• results: add new function to plot training statistics

Fix

• scheduler: fix missing arguments
• basic: renamed variable to avoid confusion

v1.7.1 (2023-09-28)

Fix

• scheduler: fixed calculation erro in cyclic lr
• condition for stdout

v1.7.0 (2023-09-22)

Feat

• new option for selection of model acceptance pattern
• calculating distributed Beta now uses percentile instead of mean
• models: add new func compute to run the model on a manual set of params
• regularizers: new method include_regularization_terms() in BaseModel for adding regularizers

Fix

• dependencies: prevent breakage of numpy 1.26
• import compute from pycmtensor.models
• fix l2 regularizer formula
• set init_types as a module level list and add tests
• add minmax clipping on neural net layer outputs
• temporary function for negative relu
• expression fix for random draw variables

Refactor

• relax the condition for stdout during training
• config: refactor config.add into .add() and .update() methods

v1.6.4 (2023-08-14)

Refactor

• changed basic model functions into staticmethod
• expressions.py: make ExpressionParser.parse a staticmethod
• remove depreciated modules

v1.6.3 (2023-08-11)

Fix

• expressions.py: use predefined draw types for random draws
• basic.py: include choice label(s) as a dictionary for predict()
• moved elasticities from statistics to model

v1.6.1 (2023-08-10)

Fix

• scheduler.py: add learning rate lower bounds for decaying functions
• basic.py: add placeholder arguments *args

Refactor

• basic.py: improve efficency of hessian matrix calculation over the sum of log likelihood over observation
• basic.py: refactoring common model functions into BaseModel
• syntax and naming changes

v1.5.0 (2023-07-26)

Feat

• optimizers.py: include new optimizer RProp
• functions.py: speed up computation and compilation by using static indexing in log_likelihood function
• functions.py: add relu() function (source taken from Theano 0.7.1)
• basic.py: new function include_params_for_convergence
• dataset.py: use as_tensor_variable to construct tensor vector from dataset[[item1, item2,...]]
• dataset.py: added list(str) as tensor arguments for train_dataset() and valid_dataset()

Fix

• init.py: fix init circular imports
• statistics.py: update t_test and varcovar matrix calculations for vector parameters
• layers.py: various fixes to neural net layers
• optimizers.py: fix SQNBFGS algorithm
• functions.py: include output as params for 1st and 2nd order derivatives
• expressions.py: fix base class inheritence and use config.seed for seed value
• expressions.py: include self in overloaded operators if other instance is of similar type as self
• basic.py: fix incorrect saved params in train()
• expressions.py: clip lower and upper bounds when updating Betas
• expressions.py: fixed Weights and Bias mathematical operations
• config.py: renamed config.py to defaultconfig.py to avoid name conflicts

Refactor

• results.py: updates results calculations for beta vectors
• expressions.py: add set_value in Param base class
• optimizers.py: removed unused imports

v1.4.0 (2023-07-16)

Feat

• optimizers.py: added new BFGS algorithm
• dataset.py: moved dataset initialization to dataset.py
• pycmtensor.py: Implemented early stopping on coefficient convergence in training loop
• functions.py: logit method now takes uneven dimensioned-utilities
• expression.py: Added RandomDraws expression for sampling in mixed logit
• get_train_data optional argument numpy_out to return numpy arrays rather than pandas arrays
• BHHH algorithm for calculating var-covar matrix applies to each data row

Fix

• results.py: fixed instance when params are not Betas
• expressions.py: include RandomDraws as an option in expression evaluations
• Update tests workflow file conda packages
• pycmtensor.py: Added missing configuration for stepLR drop_every
• tests.yml: Update tests workflow file conda packages
• optimizers.py: Fixed name typo in __all__
• results.py: Corrected calculation of hessian and bhhh matrices
• scheduler.py: Moved class function calls to parent class
• statistics.py: Fixed rob varcovar calculation error
• MNL.py: Moved aesara function to parent class
• data.py: Streamlined class function calls and removed unnecessary code
• removed package import clashes with config.py
• removed gnorm calculation
• update hessian matrix and bhhh algorithm functions

Refactor

• pycmtensor.py: temporarily removed pycmtensor.py
• MNL.py: replaced function constructors from pycmtensor.py as a function call inside the model Class object
• basic.py: moved model functionality to from pycmtensor.py to models/basic.py
• utils.py: Removed unused code

v1.3.2 (2023-06-23)

Fix

• make arguments in MNL as optional keyword arguments
• moved learning rate variable to PyCMTensorModel class

Refactor

• make model variables as property
• update __all__ package variables
• added train_data and valid_data property to Data class

v1.3.1 (2022-11-17)

Fix

• fix utility dimensions for asc only cases

v1.3.0 (2022-11-10)

Feat

• optimizers: added Nadam optimizer
• layers.py: added DenseLayer BatchNormLayer ResidualLayer
• added pycmtensor.about() to output package metadata
• added EMA function functions.exp_mov_average()

Fix

• renamed depreceated instances of aesara modules
• data.py: defaults batch_size argument to 0 if batch_size is None
• updated syntax for expressions.py class objects
• added init_type property to Weights class
• moved model aesara compile functions from models.MNL to pycmtensor.PyCMTensorModel
• added argument type hints in function.py

Refactor

• data: added import dataset cleaning step as arguments in Data()
• moved ResidualLayer to pycmtensor.models.layers
• updated timing to perf_counter
• pycmtensor: refactoring model_loglikelihood

v1.2.1 (2022-10-25)

Feat

• added pycmtensor.about() to output package metadata
• added EMA function functions.exp_mov_average()

Fix

• updated syntax for expressions.py class objects
• added init_type property to Weights class
• moved model aesara compile functions from models.MNL to pycmtensor.PyCMTensorModel

v1.2.0 (2022-10-14)

Feat

• expressions: added Weights class object (#59)
• functions: added rmse and mae objective functions (#58)
• batch shuffle for training
• function: added KL divergence loss function (#50)

Fix

• added expand_dims into logit function
• replace class function Beta.Beta with Beta.beta
• removed flatten() from logit function

v1.1.0 (2022-09-23)

Feat

• scheduler: added learning rate scheduling to train()
• code: overhaul and cleanup

Fix

• environment: update project deps and pre-commit routine
• config: remove unnecessary cxx flags from macos builds

Perf

• config: misc optimization changes

v1.0.6 (2022-08-12)

Fix

• config: added optimizing speedups to config
• config: set default cyclic_lr_mode and cyclic_lr_step_size to None
• pre-commit-config: update black to 22.6.0 in pre-commit check

Refactor

• models: refactored build_functions() into models.py
• database: refactor set_choice(choiceVar)

v1.0.5 (2022-07-27)

Fix

• tests: removed depreciated tests
• routine: remove depreciated tqdm module

v1.0.4 (2022-07-27)

Fix

• pycmtensor.py: update training method
• config.py: new config option verbosity: "high", "low"
• pycmtensor.py: remove warnings for max_iter<patience

v1.0.3 (2022-05-12)

v1.0.1 (2022-05-12)

Fix

• scheduler: fix missing args in input parameters
• scheduler: fix constantLR missing input paramerer

v1.0.0 (2022-05-10)

Feat

• python: update to python 3.10

Fix

• tests: update tests files to reflect changes in biogeme removal

v0.8.0 (2022-05-10)

Feat

• deps: remove Biogeme dependencies

v0.7.1 (2022-05-10)

Fix

• expressions: remove Biogeme dependencies
• database: remove dependencies of Biogeme
• debug: remove debug handler after each run to prevent duplication
• models: add function to return layer output -> get_layer_outputs()
• debug: disables tqdm if debug mode is on and activates debug_log

Refactor

• move elasticites from models to statistics for consistency

v0.7.0 (2022-03-17)

Feat

• models: add functionality to compute elasticities of choice vs attribute in models.py

Fix

• results: remove unnessary show_weights option in Results
• set default max_epoch on training run to adaptive rule
• print valid config options when invalid options are given as args to train()
• scheduler: modified cyclic_lr config loading sequence to fix unboundError
• train: turn saving model off for now
• config: generate os dependent ld_flags

Refactor

• utils: refactored save_to_pickle and disables it

Perf

• IterationTracker: use numpy array to store iteration data

v0.6.5 (2022-03-14)

Feat

• models: Implement the ResLogit layer

Fix

• config: set default learning schedule to ConstantLR
• config: set default seed to a random number on init

v0.6.4 (2022-03-13)

Feat

• scheduler.py: add new scheduler (CyclicLR) for adaptive LR

Fix

• project: fix project metadata and ci
• config: loadout config from train() to configparser
• utils: fix TypeError check

v0.5.0 (2022-03-02)

Feat

• config: add PyCMTensorConfig class to store config settings
• expressions: add magic methods lt le gt le ne eq
• config.py: enable pre-writing of .aesararc config file on module load
• models: add method prob() to MNLogit to output prob slices
• time_format: enable logging of build and estimation time
• results: add Predict class to output probs or discrete choices
• optimizers: add AdaGram algorithm
• Database: add getattr build-in type to Database
• pycmtensor.py: add model.output_choices to generate choices

Fix

• statistics: add small value to stderror calculation to address sqrt(0)
• dependencies: move ipywidgets and pydot to dependencies
• renamed .rst to .md fix FileNotFoundError
• result: print more verbose results and options
• Database: add name to shared_data
• train: model instance now load initiated model class (not input Class as argument)
• Database: set choiceVar to mandatory argument
• PyCMTensor: rename append_to_params to add_params for consistency
• PyCMTensor: new method to add regularizers to cost function
• Expressions: invokes different operator for Beta Beta maths
• show excluded data in model est. output
• results: standardized naming conventions in modules db->database
• tqdm: add arg in train() to enable notebook progressbar
• swissmetro_test.ipynb: update swissmetro example

Refactor

• PyCMTensor: refactoring models from pycmtensor.py
• Database: refactor(Database): refactoring database.py from pycmtensor.py
• optimizers: refactor base Optimizer class
• moved Beta Weights to expressions.py

Perf

• shared_data: improve iteration speed by implementing shared() on input data