Fine tuning API

fine_tune(images_dir, masks_dir, output_dir, *, mode='2D', patch_size=None, initial_weights=None, epochs=10, batch_size=None, steps_per_epoch=None, learning_rate=0.0001, validation_fraction=0.2, validation_images_dir=None, validation_masks_dir=None, seed=42)

Fine tune published weights with paired grayscale TIFFs and instance masks.

Images and masks have matching filename stems in separate directories and the same shape. Mask value 0 is background; each positive integer identifies one cell. Training writes best weights, epoch history, and run metadata into output_dir.

Parameters:

Name	Type	Description	Default
images_dir	path - like	Directories of paired training images and instance masks.	required
masks_dir	path - like	Directories of paired training images and instance masks.	required
output_dir	path - like	Directory for best.weights.h5, history.csv, and run.json.	required
mode	('2D', '3D')	Dimension of the image data.	"2D"
patch_size	int	Published model size; defaults to 256 for 2D and 128 for 3D.	None
initial_weights	path - like	Starting weights. If omitted, use the published model cache.	None
epochs	int	Maximum number of training epochs.	10
batch_size	int	Samples per batch; defaults to 4 for 2D and 1 for 3D.	None
steps_per_epoch	int	Training batches per epoch; inferred from pair count when omitted.	None
learning_rate	float	Adam optimizer learning rate.	1e-4
validation_fraction	float	Fraction of pairs held out when no explicit validation set is given.	0.2
validation_images_dir	path - like	Separate paired validation directories. Supply both or neither.	None
validation_masks_dir	path - like	Separate paired validation directories. Supply both or neither.	None
seed	int	Seed for splitting, sampling, and TensorFlow initialization.	42

Returns:

Type	Description
FineTuneResult	Paths to the saved weights and history, and split sizes.

FineTuneResult dataclass

Paths to the best weights and history, with training split sizes.

Attributes:

Name	Type	Description
weights_path	Path	Best validation weights saved by training.
history_path	Path	CSV containing metrics for each epoch.
train_count	int	Number of training image/mask pairs.
validation_count	int	Number of validation image/mask pairs.

fine_tune is also available as from morphodeep.training import fine_tune.