abstract_result_packages

class aitoolbox.experiment.result_package.abstract_result_packages.AbstractResultPackage(pkg_name=None, strict_content_check=False, np_array=True, **kwargs)

Bases: ABC

Base Result package used to derive specific result packages from

Functions which the user should potentially override in a specific result package:

• prepare_results_dict()

• list_additional_results_dump_paths()

• set_experiment_dir_path_for_additional_results()

Parameters:

• pkg_name (str or None) – result package name used just for clarity

• strict_content_check (bool) – should just print warning or raise the error and crash

• np_array (bool or str) –

  how the inputs should be handled. Should the package try to automatically guess, or you want to manually decide whether to leave the inputs as they are or convert them to np.array. Possible options: True, False, ‘auto’

  Be slightly careful with ‘auto’ as it sometimes doesn’t work, so it is preferable to explicitly use True/False

• **kwargs (dict) – additional package_metadata for the result package

abstract prepare_results_dict()

Perform result package building

Mostly this consists of executing calculation of selected performance metrics and returning their result dicts. If you want to use multiple performance metrics you have to combine them in the single self.results_dict at the end by doing this:

return {**metric_dict_1, **metric_dict_2}

Returns:

calculated result dict

Return type:

dict

prepare_result_package(y_true, y_predicted, hyperparameters=None, **kwargs)

Prepares the result package by taking labels and running them through the specified metrics

This function is automatically called from the torchtrain callbacks to evaluate the provided callback. The main feature of this function is the call to the user-derived prepare_results_dict() function of the implemented result package where the metrics evaluation logic is implemented.

Parameters:

• y_true (numpy.ndarray or list) – ground truth targets

• y_predicted (numpy.ndarray or list) – predicted targets

• hyperparameters (dict or None) – dictionary filled with the set hyperparameters

• **kwargs (dict) – additional results for the result package

Returns:

None

static auto_y_input_array_convert(y_array)

Try to automatically decide if array should be left as it is or convert to np.array

Not working in all the situations so relying on it at all times is not recommended. Especially for costly experiments rely rather on your own judgement and explicitly define if np.array conversion is needed.

TODO: make it smarter so ‘auto’ option can be used more often

Parameters:

y_array (list) –

Return type:

list or numpy.array

get_results()

Get calculated results dict

Returns:

results dict

Return type:

dict

get_hyperparameters()

Get hyperparameters in a dict form

Returns:

hyperparameters dict

Return type:

dict

get_additional_results_dump_paths()

Return paths to the additional results which are stored to local drive when the package is evaluated

For example if package plots attention heatmaps and saves pictures to disk, this function will return paths to these picture files. This is achieved via the call to the user-implemented function list_additional_results_dump_paths().

Returns:

list of lists of string paths if it is not None. Each element of the list should be list of: [[results_file_name, results_file_local_path], … [,]]

Return type:

list or None

list_additional_results_dump_paths()

Specify the list of metadata files you also want to save & upload to s3 during the experiment saving procedure

By default, there are no additional files that are saved as the return is None. If you want to save your specific additional files produced during the training procedure, then override this method specifying the file paths.

If you want to save a whole folder of files, use zip_additional_results_dump() function to zip it into a single file and save this zip instead.

The specified files are any additional data you would want to include into the experiment folder in addition to the model save files and performance evaluation report files. For example a zip of attention heatmap pictures in the machine translation projects.

Returns:

list of lists of string paths if it is not None. Each element of the list should be list of: [[results_file_name, results_file_local_path], … [,]]

Return type:

list or None

set_experiment_dir_path_for_additional_results(project_name, experiment_name, experiment_timestamp, local_model_result_folder_path)

Set experiment folder path after potential timestamps have already been generated.

Experiment folder setting for additional metadata results output is needed only in certain result packages, for example in QuestionAnswerResultPackage where the self.output_text_dir initially has only the name of the folder where the results text predictions for each example should be stored. This function when implemented reforms the folder name so that it becomes a full path placing the folder inside the experiment folder (for which the timestamp at the start of train loop is needed).

Another use of this function is in MachineTranslationResultPackage where the attention heatmap pictures are stored as additional metadata results.

As can be seen from the fact that the train loop mechanism is mentioned, this method’s functionality is primarily used for PyTorch experiments.

Parameters:

• project_name (str) – root name of the project

• experiment_name (str) – name of the particular experiment

• experiment_timestamp (str) – time stamp at the start of training

• local_model_result_folder_path (str) – root local path where project folder will be created

Returns:

None

qa_check_hyperparameters_dict()

Quality check the hyperparams dict

Returns:

None

qa_check_additional_results_dump_paths()

Quality check the additional results path

Returns:

None

warn_about_result_data_problem(msg)

Generic function for writing out warnings

Either just printing out the warning or throw the error exception.

Parameters:

msg (str) – warning message either printed or written in the raised error

Raises:

ValueError –

Returns:

None

static zip_additional_results_dump(source_dir_path, zip_path)

Utility function for zipping a folder into .zip archive

Parameters:

• source_dir_path (str) – path to the folder that is going to be zipped

• zip_path (str) – specify the path of the zip file which will be created

Returns:

the full path to the produced zip file (with the .zip extension appended)

Return type:

str

__add__(other)

Concatenate result packages

Combine results from both result packages into a single one.

Parameters:

other (AbstractResultPackage or dict) – another result package to be concatenated

Returns:

merged result package

Return type:

MultipleResultPackageWrapper

__radd__(other)

Concatenate result package

Parameters:

other (AbstractResultPackage or dict) – another result package to be concatenated

Returns:

merged result package

Return type:

MultipleResultPackageWrapper

add_merge_multi_pkg_wrap(other_object)

Result package merge

Parameters:

other_object (AbstractResultPackage or dict) – another result package to be merged with the current package

Returns:

merged result package

Return type:

MultipleResultPackageWrapper

static _create_other_object_pkg(other_object)

Util to deep copy and wrap results into the simple result package

Parameters:

other_object (AbstractResultPackage or dict) – results package or results dict

Returns:

deep copy of results wrapped in the simple result package

Return type:

AbstractResultPackage or MultipleResultPackageWrapper

__iadd__(other)

Append result package

Parameters:

other (AbstractResultPackage or dict) – another result package to be appended to the current package

Returns:

merged result package

Return type:

AbstractResultPackage

add_merge_dicts(other)

Append result package to the current one

Parameters:

other (AbstractResultPackage or dict) – another result package to be appended to the current package

Returns:

merged result package

Return type:

AbstractResultPackage

_merge_dicts(other_results_dict)

Results dict merge util

Parameters:

other_results_dict (dict) – another results dict to be added to the results dict in the current result package

Returns:

merged result package

Return type:

AbstractResultPackage

warn_if_results_dict_not_defined()

class aitoolbox.experiment.result_package.abstract_result_packages.PreCalculatedResultPackage(results_dict, strict_content_check=False, **kwargs)

Bases: AbstractResultPackage

Result package which doesn’t have any evaluation logic but just accepts pre-calculated results dict

Parameters:

• results_dict (dict) – pre-calculated results dict

• strict_content_check (bool) – should just print warning or raise the error and crash

• **kwargs (dict) – result package additional meta-data

prepare_results_dict()

Perform result package building

Mostly this consists of executing calculation of selected performance metrics and returning their result dicts. If you want to use multiple performance metrics you have to combine them in the single self.results_dict at the end by doing this:

return {**metric_dict_1, **metric_dict_2}

Returns:

calculated result dict

Return type:

dict

class aitoolbox.experiment.result_package.abstract_result_packages.MultipleResultPackageWrapper(strict_content_check=False, **kwargs)

Bases: AbstractResultPackage

Wrapper result package which combines multiple evaluated result packages into a single result package

Parameters:

• strict_content_check (bool) – should just print warning or raise the error and crash

• **kwargs (dict) – result package additional meta-data

prepare_result_package(result_packages, hyperparameters=None, **kwargs)

Prepares the multiple result package by merging the results from both result packages

Parameters:

• result_packages (list) – list of result packages where each of them is object inherited from aitoolbox.experiment.result_package.abstract_result_packages.AbstractResultPackage. If you want to add raw results in dict form, this dict first needs to be wrapped into PreCalculatedResultPackage to satisfy the result package object requirement.

• hyperparameters (dict or None) – hyperparameters dict

• **kwargs – result package additional meta-data

Returns:

None

prepare_results_dict()

Perform result package building

Mostly this consists of executing calculation of selected performance metrics and returning their result dicts. If you want to use multiple performance metrics you have to combine them in the single self.results_dict at the end by doing this:

return {**metric_dict_1, **metric_dict_2}

Returns:

calculated result dict

Return type:

dict

get_additional_results_dump_paths()

Return paths to the additional results which are stored to local drive when the package is evaluated

For example if package plots attention heatmaps and saves pictures to disk, this function will return paths to these picture files. This is achieved via the call to the user-implemented function list_additional_results_dump_paths().

Returns:

list of lists of string paths if it is not None. Each element of the list should be list of: [[results_file_name, results_file_local_path], … [,]]

Return type:

list or None

__len__()

Get number of result packages inside the multi result package wrapper

Returns:

number of result packages inside this multi package wrapper

Return type:

int

add_merge_multi_pkg_wrap(other_object)

Result package merge

Parameters:

other_object (AbstractResultPackage or dict) – another result package to be merged with the current package

Returns:

merged result package

Return type:

MultipleResultPackageWrapper