Tags: Natooz/MidiTok
Tags
Better tests + minor improvements (#108) * parametrizing tests, improvements in preprocess_midi, fixes for miditoolkit 1.0.1 * fixing absolute path for data aug and io tests * fix test file tree, data aug report saved in out_path * fix in data augmentation saving paths * forced disabling original in out_dir when calling data aug from tok_dataset * using pytest tmp_path to write files, and TEST_LOG_DIR if required * lighter and more elegant MIDI assertions + covering check_midi_equals * better tokenization test sets, set_midi_max_tick method, renamed "nb" contractions to "num", handling empty tokens lists in methods * dealing with empty midi file (#110) * dealing with empty midi file * add a new test midi tokenizer file instead of changing the original one * delete test_midi_tokenizer * Adding check empty input for _ids_to_tokens as well --------- Co-authored-by: Nathan Fradet <[email protected]> * adding tests for empty MIDI and associated fixes * fixes from tests with empty midi + retry hf hub tests when http errors * fix convert_sequence_to_tokseq when list in last dim is empty * better tok test sets * testing with multiple time resolutions, adjusting notes ends * fix _quantize_time_signatures (delete_equal_successive_time_sig_changes) --------- Co-authored-by: feiyuehchen <[email protected]>
one_token_stream_for_programs config param and successive TimeShift t… …okens (#70) * one_token_stream_for_programs config param * Successive TimeShift tokens for MIDILike and TSD, quantize note now clips note durations to the maximum duration that the tokenizer can represent, * Rests are now represented the same as TimeShift / Duration and can now be successive, rest_range renamed beat_res_rest, outdated useless code cleaned * fixes in rest workflow and save/load config
PreviousNext