campa.tl.Experiment

class Experiment(config)

Experiment stored on disk with neural network.

Initialised with config dictionary with keys:

• experiment: where to save experiment

  • dir: experiment folder

  • name: name of the experiment

  • save_config: (bool), whether to save this config in the folder

• data: which dataset to use for training

  • data_config: name of the data config to use, should be registered in campa.ini

  • dataset_name: name of the dataset, relative to DATA_DIR

  • output_channels: Channels that should be predicted by the neural network. Defaults to all input channels.

• model: model definition

  • model_cls: instance or value of ModelEnum

  • model_kwargs: keyword arguments passed to the model class

  • init_with_weights: if true, looks for saved weights in experiment_dir. if a path, loads these weights

• training: training hyper-parameters

  • learning_rate: learning rate to use

  • epochs: number of epochs to train

  • batch_size: number of samples per batch

  • loss: mapping of model output names to values of LossEnum. Possible names are decoder and latent.

  • metrics: mapping of model output names to values of LossEnum.

  • save_model_weights: (bool) whether or not to save the model.

  • save_history: (bool) save csv with losses and metrics at each epoch.

  • overwrite_history: overwrite existing history csv file. Otherwise concatenate to it.

• evaluation: evaluation on val/test split

  • split: train, val, or test

  • predict_reps: (list) model output that should be predicted. Possible values: decoder, latent.

  • img_ids: number of images to predict, or list of image ids.

  • predict_imgs: (bool) whether to predict reconstructed images.

  • predict_cluster_imgs: (bool) whether to predict clustered images.

• cluster: clustering on val/test split

  • cluster_name: name of the clustering, used to save npy file.

  • cluster_rep: model output name to use for clustering, or “mpp”.

  • cluster_method: leiden or kmeans.

  • leiden_resolution: resolution parameter for leiden clustering.

  • subsample: None or “subsample”, whether or not to subsample data before clustering.

  • subsample_kwargs: passed to campa.data.MPPData.subsample() for creating the subsample data for clustering.

  • umap: (bool) predict UMAP of cluster_rep.

Parameters

config (MutableMapping[str, Any]) – Experiment config.

Attributes

config	Experiment config, see Experiment.
dir	Experiment directory.
epoch	Last epoch for which there is a trained model.
estimator_config	Config dictionary to initialise campa.tl.Estimator.
evaluate_config	Config dictionary to initialise campa.tl.Predictor.
full_path	Full path to Experiment.
is_trainable	Return false, if this is not a trainable experiment.
name	Experiment name.

Methods

from_dir(exp_path)	Initialise experiment from trained experiment in exp_path.
get_cluster_annotation([cluster_name, ...])	Read cluster_annotation file for full data from disk.
get_experiments_from_config(config_fname[, ...])	Initialise and return experiments from configs in config file.
get_experiments_from_dir(exp_dir[, ...])	Initialise and return experiments from experiment directory.
get_history()	Training history.
get_split_cluster_annotation([cluster_name])	Read cluster_annotation file for evaluation split from disk.
get_split_imgs_mpp_data()	Val_imgs / test_imgs MPPData read from results_epoch{self.epoch}`.
get_split_mpp_data()	Val or test MPPData read from results_epoch{self.epoch}.
set_to_evaluate()	Prepare Experiment for evaluation.