baby.training.flattener_trainer.FlattenerTrainer¶
- class baby.training.flattener_trainer.FlattenerTrainer(save_dir: Path, stats_file: str, flattener_file: str)¶
Bases:
object- Attributes
Methods
fit([nbins, min_size, pad_frac, bud_max])Optimise the parameters of the SegmentationFlattener based on previously computed statistics.
generate_flattener_stats(train_gen, val_gen, ...)Generate overlap and erosion statistics for augmented data in input.
plot_stats([nbins])Plot a histogram of cell overlap statistics of the training set.
- __init__(save_dir: Path, stats_file: str, flattener_file: str)¶
Optimises the hyper-parameters for the SegmentationFlattener.
#TODO describe method for optimisation :param save_dir: the base directory in which to save outputs :param stats_file: the name of the file in which the stats are saved :param flattener_file: the name of the file defining the flattener
Methods
__init__(save_dir, stats_file, flattener_file)Optimises the hyper-parameters for the SegmentationFlattener.
fit([nbins, min_size, pad_frac, bud_max])Optimise the parameters of the SegmentationFlattener based on previously computed statistics.
generate_flattener_stats(train_gen, val_gen, ...)Generate overlap and erosion statistics for augmented data in input.
plot_stats([nbins])Plot a histogram of cell overlap statistics of the training set.
Attributes
The last flattener saved to file.
The last statistics computed, loaded from self.stats_file
- fit(nbins=30, min_size=10, pad_frac=0.03, bud_max=200)¶
Optimise the parameters of the SegmentationFlattener based on previously computed statistics.
# TODO define optimisation method
# TODO missing some unknown parameters :param nbins: :param min_size: the minimum size of a segmented object for it to be considered a cell. :param pad_frac: :param bud_max: the maximum size (in pixels) of cells considered buds. :return: None, saves the optimal parameters to the flattener file :raises: BadProcess if no statistics have been computed yet.
- property flattener: SegmentationFlattening¶
The last flattener saved to file. :return: a segmentation flattener :raise: BadProcess, if no flattener has been assigned.
- generate_flattener_stats(train_gen: ImageLabel, val_gen: ImageLabel, train_aug: Augmenter, val_aug: Augmenter, max_erode: int = 5)¶
Generate overlap and erosion statistics for augmented data in input.
- Parameters
train_gen – the generator of training images and their labels
val_gen – the generator of validation images and their labels
train_aug – the augmenter to use for training images
val_aug – the augmenter to use for validation images
max_erode – the maximum allowed number of erosions used to
generate erosion values :return: None, saves results to self.stats_file
- plot_stats(nbins=30)¶
Plot a histogram of cell overlap statistics of the training set.
# TODO describe what the plot means # TODO add an image as an example
- Parameters
nbins – binning of data, passed to matplotlib.pyplot.hist2d
- Returns
None, saves the resulting figure under `self.save_dir /
“flattener_stats.png”`
- property stats: TrainValProperty¶
The last statistics computed, loaded from self.stats_file
- Returns
The last training and validation statistics computed.
- Raises
BadProcess error if the file does not exist.