--- ops: [train, evaluate] model: { # TODO: update this line with the absolute path to the file. path: /deeperdeepsea.py, class: DeeperDeepSEA, class_args: { sequence_length: 1000, n_targets: 9, }, non_strand_specific: mean } sampler: !obj:selene_sdk.samplers.IntervalsSampler { reference_sequence: !obj:selene_sdk.sequences.Genome { # we include relative paths here, but we recommend using absolute # paths for future configuration files input_path: ./genome/Bdistachyon.fasta }, features: !obj:selene_sdk.utils.load_features_list { input_path: ./data/bd/distinct_features.txt }, target_path: ./data/bd/sorted_bd.bed.gz, intervals_path: ./data/bd/bd_genome_interval.bed, seed: 127, # A positive example is an 1000bp sequence with at least 1 class/feature annotated to it. # A negative sample has no classes/features annotated to the sequence. sample_negative: True, sequence_length: 1000, center_bin_to_predict: 200, test_holdout: [chr05], validation_holdout: [chr04], # The feature must take up 50% of the bin (200bp) for it to be considered # a feature annotated to that sequence. feature_thresholds: 0.5, mode: train, save_datasets: [train, validate, test] } train_model: !obj:selene_sdk.TrainModel { batch_size: 256, max_steps: 20000, # update this value for longer training report_stats_every_n_steps: 1000, n_validation_samples: 15000, n_test_samples: 15000, cpu_n_threads: 16, use_cuda: True, # TODO: update this if CUDA is not on your machine data_parallel: True } random_seed: 147 output_dir: ./models/bd_model create_subdirectory: False load_test_set: False ...