Configs

RunConfig

class dacapo.experiments.RunConfig(task_config, architecture_config, trainer_config, datasplit_config, name, repetition, num_iterations, validation_interval=1000, start_config=None)

DataSplitConfigs

class dacapo.experiments.datasplits.TrainValidateDataSplitConfig(name, train_configs, validate_configs)

This is the standard Train/Validate DataSplit config.

datasplit_type

alias of TrainValidateDataSplit

ArchitectureConfigs

class dacapo.experiments.architectures.CNNectomeUNetConfig(name, input_shape, fmaps_out, fmaps_in, num_fmaps, fmap_inc_factor, downsample_factors, kernel_size_down=None, kernel_size_up=None, eval_shape_increase=None, upsample_factors=None, constant_upsample=True, padding='valid')

This class configures the CNNectomeUNet based on https://github.com/saalfeldlab/CNNectome/blob/master/CNNectome/networks/unet_class.py

Includes support for super resolution via the upsampling factors.

architecture_type

alias of CNNectomeUNet

TaskConfigs

class dacapo.experiments.tasks.OneHotTaskConfig(name, classes)

This is a One Hot prediction task that outputs a probability vector of length c for each voxel where c is the number of classes. Each voxel prediction has all positive values an l1 norm equal to 1.

Post processing is extremely easy, the class of each voxel is simply the argmax over the vector of output probabilities.

task_type

alias of OneHotTask

class dacapo.experiments.tasks.AffinitiesTaskConfig(name, neighborhood, lsds=False)

This is a Affinities task config used for generating and evaluating voxel affinities for instance segmentations.

task_type

alias of AffinitiesTask

class dacapo.experiments.tasks.DistanceTaskConfig(name, channels, clip_distance, tol_distance, scale_factor=1, mask_distances=False)

This is a Distance task config used for generating and evaluating signed distance transforms as a way of generating segmentations.

The advantage of generating distance transforms over regular affinities is you can get a denser signal, i.e. 1 misclassified pixel in an affinity prediction could merge 2 otherwise very distinct objects, this cannot happen with distances.

task_type

alias of DistanceTask

TrainerConfigs

class dacapo.experiments.trainers.GunpowderTrainerConfig(name, batch_size, learning_rate, num_data_fetchers=5, augments=_Nothing.NOTHING, snapshot_interval=None, min_masked=0.15, clip_raw=True)

trainer_type

alias of GunpowderTrainer