Skip to content

Latest commit

 

History

History
58 lines (52 loc) · 3.11 KB

TODO.md

File metadata and controls

58 lines (52 loc) · 3.11 KB
Raw

TODOs

  • GUI/website
  • Save the stage 1 probabilities and have interactive visualizer as you change threshold.
  • Finish writing main README
  • Finish writing training README, for stage_two.
  • stage_two dataset uses state_space_functions, while general_beat_saber dataset uses level_processing_functions. Check if need to merge these two, as they have a fair of common functionality I think.
  • level_processing_functions is even more ugly now, because of new fix to speed up training by preprocessing levels into numpy tensors, and leaving the old functions just in case (feature creep :P). So should remove the old functions, as memory is cheaper than computing generally, so preprocessing is generally preferred if it speeds up computing!
  • is there a way go make get_reduced_tensors_from_level in level_processing_functions faster? -- now that we've fixed the bug, and made the sequence length longer, it's slowed down training :(. Yes there is, by preprocessing, but it introduced new TODO above!
  • make stage_two_dataset work with new beat saber data.
  • fix importing. E.g. base_options.py imports from base, that doesn't exist now....
  • Make DDC port
  • Train on new data
  • Testing code. Use perplexity, as in NLP literature
  • Make stage two that uses multi_mel features, or even the DDC embeddings as its inputs
  • Make non-reduced-state data work, wavenet.
  • Obstacles, etc.

Folder structure and refactoring comments

/base - !!! Lots of stuff in here, needs filtering and organising into training, generation and models !!! /data /extracted_data - Beatsaber Levels and Songs /metadata - [PLACEHOLDER] Beastsaber complete metadata file (guillermo) /statespace - Contains outputs of identify_state_space.py (tim/ralph) /models /scripts /data_retrieval /download_data.py - Perhaps integrate Beastsaber as download source /data_processing - lots of redundant code here, repetition in /base/models /difficulty_analysis.py [INCOMPLETE] /state_space_functions.py (level_processing_functions.py) produce_distinct_state_space_representations() compute_explicit_states_from_json() compute_explicit_states_from_bs_level() - Wrapper compute_shortest_inter_event_beat_gap() produce_transition_probability_matrix_from_distinct_state_spaces() compute_state_sequence_representation_from_json() - Sorts output of produce_distinct_state_space_representations for use in get_block_sequence_with_deltas() get_block_sequence_with_deltas() /feature_extraction - possible duplications in /base/models /example_decode_encode.py [LEGACY?] perhaps redundant (DEV testing code?) /feature_extraction.py /features_base.py - contains non-ML approach to level generation (should be extracted) /process_songs.py - needs to be integrated or deleted /training /generation /evaluation /rule_check_states.py - check/comment /graph_visualisation.py - check/comment /misc /io_functions.py - Check through /bash_scripts - Misc Bash scripts (Guillermo) /experiment_name - Unknown options file (Andrea/Guillermo) /web - [DEV] The beginnings of a web api for level generation