To start the training, run the following command:

CUDA_VISIBLE_DEVICES=0 ipython experiments/cremi/train_model.py -- <yourExperimentName> --DATA_HOMEDIR <path/to/the/cremi/data/you/downloaded> --inherit main_config.yml

The experiment data are saved by default in the experiments/cremi/runs folder.

To start a new training and loading a previous model, run for example the following command:

CUDA_VISIBLE_DEVICES=0 ipython experiments/cremi/train_model.py -- name_new_experiment --DATA_HOMEDIR <path/to/the/cremi/data/you/downloaded> --inherit main_config.yml --update0 new_experiment_config.yml --config.model.model_kwargs.loadfrom PATH_TO_OLD_CHECKPOINT.pytorch --config.trainer.optimizer.Adam.lr 6e-5

CUDA_VISIBLE_DEVICES=0 ipython experiments/cremi/infer.py -- test_infer --DATA_HOMEDIR <path/to/the/cremi/data/you/downloaded> --inherit main_config.yml --update0 infer_config.yml --config.model.model_kwargs.loadfrom RUNS__HOME/model_name/checkpoint.pytorch --config.name_experiment your_infer_experiment_name --config.loaders.infer.loader_config.batch_size 1 --config.export_path <directory-where-to-save-affinities>

More specific infer-parameters are found in configs/infer_config.yml. Some examples:

• Predict only a small part of the data using the parameter loaders.infer.volume_config.data_slice
• You may need to adjust the parameter model.slicing_config.window_size according to the memory available on your GPU.

After predicting the affinities, have a look at the example script postprocess_affinities.py showing how to convert the affinities into an instance segmentation using GASP or Mutex Watershed.