# Alternative data configuration that demonstrates the use # of TF Dataset files instead of tabular data # The files can be created by executing the following: # python $ZERO2NEURO_PATH/zero2neuro.py @data.txt @experiment.txt @network.txt --data_save_folds ds/amino --data_representation tf-dataset --nogo # Must explicitly define tf-dataset as the incoming data format and # the internal representation --data_format=tf-dataset --data_representation=tf-dataset # We are still doing holistic cross-validation --data_set_type=holistic-cross-validation # Specify each of the folds separately. A single TF-Dataset cannot be # partitioned further to create other folds --data_files ds/amino_fold_00.ds,0 ds/amino_fold_01.ds,0 ds/amino_fold_02.ds,1 ds/amino_fold_03.ds,1 ds/amino_fold_04.ds,2 ds/amino_fold_05.ds,2 ds/amino_fold_06.ds,3 ds/amino_fold_07.ds,3 ds/amino_fold_08.ds,4 ds/amino_fold_09.ds,4 # Combine pairs of input files to form 5 folds --data_fold_split=group-by-file # TF training/evaluation arguments --data_tf_shuffle=32 --data_tf_repeat --data_steps_per_epoch=2 --data_tf_prefetch=4 # Preprocessing # We are using a tokenizer to translate the strings into lists of # integer tokens --tokenizer --tokenizer_max_tokens=24 --tokenizer_split=character --tokenizer_output_sequence_length=30 # Transform the integer tokens into 15-dimensional vectors # (this is a learned transformation) --embedding_dimensions=15