GUI Training: SLEAP does not use non-GUI parameters when using training config as template

[roomrys]

Bug description

Loading a training config in the GUI does not use the loaded training config's hidden parameters, i.e. parameters not displayed in the GUI.

Expected behaviour

The loaded training config's parameters will be used as defaults for parameters which are not set in the GUI.

Actual behaviour

Parameters set in GUI are used, but parameters not set in GUI have defaults determined by TrainingJobConfig class in code.

Your personal set up

Screenshots

How to reproduce

1. Go to Training Pipeline in GUI
2. Click on Model Configuration tab
3. Select model from drop-down.
4. Run

[roomrys]

Relevant Code

• run calls get_ever_head_config_data
  

  sleap/sleap/gui/learning/dialog.py

  Lines 581 to 617 in 5093f69

  	def run(self):
  	"""Run with current dialog settings."""
  	pipeline_form_data = self.pipeline_form_widget.get_form_data()
  	items_for_inference = self.get_items_for_inference(pipeline_form_data)
  	config_info_list = self.get_every_head_config_data(pipeline_form_data)
  	# Close the dialog now that we have the data from it
  	self.accept()
  	# Run training/learning pipeline using the TrainingJobs
  	new_counts = runners.run_learning_pipeline(
  	labels_filename=self.labels_filename,
  	labels=self.labels,
  	config_info_list=config_info_list,
  	inference_params=pipeline_form_data,
  	items_for_inference=items_for_inference,
  	)
  	self._handle_learning_finished.emit(new_counts)
  	# count < 0 means there was an error and we didn't get any results.
  	if new_counts is not None and new_counts >= 0:
  	total_count = items_for_inference.total_frame_count
  	no_result_count = total_count - new_counts
  	message = (
  	f"Inference ran on {total_count} frames."
  	f"\n\nInstances were predicted on {new_counts} frames "
  	f"({no_result_count} frame{'s' if no_result_count != 1 else ''} with "
  	"no instances found)."
  	)
  	win = QtWidgets.QMessageBox(text=message)
  	win.setWindowTitle("Inference Results")
  	win.exec_()

• get_every_head_config_data calls make_training_config_from_key_val_dict
  

  sleap/sleap/gui/learning/dialog.py

  Lines 415 to 446 in 5093f69

  	def get_every_head_config_data(
  	self, pipeline_form_data
  	) -> List[configs.ConfigFileInfo]:
  	cfg_info_list = []
  	# Copy relevant data into linked fields (i.e., anchor part).
  	self.adjust_data_to_update_other_tabs(pipeline_form_data)
  	for tab_name in self.shown_tab_names:
  	trained_cfg_info = self.tabs[tab_name].trained_config_info_to_use
  	if trained_cfg_info:
  	trained_cfg_info.dont_retrain = trained_cfg_info
  	cfg_info_list.append(trained_cfg_info)
  	else:
  	tab_cfg_key_val_dict = self.tabs[tab_name].get_all_form_data()
  	self.merge_pipeline_and_head_config_data(
  	head_name=tab_name,
  	head_data=tab_cfg_key_val_dict,
  	pipeline_data=pipeline_form_data,
  	)
  	cfg = scopedkeydict.make_training_config_from_key_val_dict(
  	tab_cfg_key_val_dict
  	)
  	cfg_info = configs.ConfigFileInfo(config=cfg, head_name=tab_name)
  	cfg_info_list.append(cfg_info)
  	return cfg_info_list

• make_training_config_from_key_val_dict creates TrainingJobConfig
  

  sleap/sleap/gui/learning/scopedkeydict.py

  Lines 177 to 191 in 5093f69

  	def make_training_config_from_key_val_dict(key_val_dict: dict) -> TrainingJobConfig:
  	"""
  	Make :py:class:`TrainingJobConfig` object from flat dictionary.
  	Arguments:
  	key_val_dict: Flat dictionary from :py:class:`TrainingEditorWidget`.
  	Returns:
  	The :py:class:`TrainingJobConfig` object.
  	"""
  	apply_cfg_transforms_to_key_val_dict(key_val_dict)
  	cfg_dict = ScopedKeyDict(key_val_dict).to_hierarchical_dict()
  	cfg = cattr.structure(cfg_dict, TrainingJobConfig)
  	return cfg

• TrainingJobConfig has defaults for values not specified in GUI (outputs = OutputsConfig)
  

  sleap/sleap/nn/config/training_job.py

  Lines 41 to 62 in 5093f69

  	class TrainingJobConfig:
  	"""Configuration of a training job.
  	Attributes:
  	data: Configuration options related to the training data.
  	model: Configuration options related to the model architecture.
  	optimization: Configuration options related to the training.
  	outputs: Configuration options related to outputs during training.
  	name: Optional name for this configuration profile.
  	description: Optional description of the configuration.
  	sleap_version: Version of SLEAP that generated this configuration.
  	filename: Path to this config file if it was loaded from disk.
  	"""
  	data: DataConfig = attr.ib(factory=DataConfig)
  	model: ModelConfig = attr.ib(factory=ModelConfig)
  	optimization: OptimizationConfig = attr.ib(factory=OptimizationConfig)
  	outputs: OutputsConfig = attr.ib(factory=OutputsConfig)
  	name: Optional[Text] = ""
  	description: Optional[Text] = ""
  	sleap_version: Optional[Text] = sleap.__version__
  	filename: Optional[Text] = ""

[roomrys]

This feature is now available in the (pre) release 1.3.0a0, to install, first uninstall and then:
conda (Windows/Linux/GPU):

conda create -y -n sleap -c sleap -c sleap/label/dev -c nvidia -c conda-forge sleap=1.3.0a0

pip (any OS except Apple Silicon):

pip install sleap==1.3.0a0

Warning: This is a pre-release! Expect bugs and strange behavior when testing.