deepblink.models package

Submodules

• deepblink.models.spots module

Module contents

Models module with the training loop and logic to handle data which feeds into the loop.

class deepblink.models.Model(augmentation_args: Dict[KT, VT], dataset_args: Dict[KT, VT], dataset_cls: deepblink.datasets._datasets.Dataset, network_args: Dict[KT, VT], network_fn: Callable, loss_fn: Callable, optimizer_fn: Callable, train_args: Dict[KT, VT], pre_model: keras.engine.training.Model = None, **kwargs)

Bases: object

Base class, to be subclassed by predictors for specific type of data, e.g. spots.

Parameters:

• dataset_args – Dataset arguments containing - version, cell_size, flip, illuminate, rotate, gaussian_noise, and translate.
• dataset_cls – Specific dataset class.
• network_args – Network arguments containing - n_channels.
• network_fn – Network function returning a built model.
• loss_fn – Loss function.
• optimizer_fn – Optimizer function.
• train_args – Training arguments containing - batch_size, epochs, learning_rate.
• pre_model – Loaded, pre-trained model to bypass a new network creation.

Kwargs:

batch_format_fn: Formatting function added in the specific model, e.g. spots. batch_augment_fn: Same as batch_format_fn for augmentation.

evaluate(x: numpy.ndarray, y: numpy.ndarray) → List[float]

Evaluate on images / masks and return l2 norm and f1 score.

fit(dataset: deepblink.datasets._datasets.Dataset, augment_val: bool = True, callbacks: list = None) → None

Training loop.

metrics

Return metrics.

class deepblink.models.SpotsModel(**kwargs)

Bases: deepblink.models._models.Model

Class to predict spot localization; see base class.

metrics

List of all metrics recorded during training.

predict_on_image(image: numpy.ndarray) → numpy.ndarray

Predict on a single input image.