src/music/config.py

import numpy as np
import os

REPO_PATH = '/'.join(os.path.abspath(__file__).split('/')[:-3]) + '/'
AUDIO_PATH = REPO_PATH + 'data/music/audio/'
MIDI_PATH = REPO_PATH + 'data/music/midi/'
MUSIC_PATH = REPO_PATH + 'data/music/'
PROCESSED_PATH = REPO_PATH + 'data/music/processed/'
ENCODED_PATH = REPO_PATH + 'data/music/encoded/'
HANDCODED_REP_PATH = MUSIC_PATH + 'handcoded_reps/'
DATASET_PATH = REPO_PATH + 'data/music/encoded_new_structured/diverse_piano/'
SYNTH_RECORDED_AUDIO_PATH = AUDIO_PATH + 'synth_audio_recorded/'
SYNTH_RECORDED_MIDI_PATH = MIDI_PATH + 'synth_midi_recorded/'
CHECKPOINTS_PATH = REPO_PATH + 'checkpoints/'
EXPERIMENT_PATH = REPO_PATH + 'experiments/'
SEED = 0

# params for data download
ALL_URL_PATH = REPO_PATH + 'data/music/audio/all_urls.pickle'
ALL_FAILED_URL_PATH = REPO_PATH + 'data/music/audio/all_failed_urls.pickle'
RATE_AUDIO_SAVE = 16000
FROM_URL_PATH = AUDIO_PATH + 'from_url/'

# params transcription
CHKPT_PATH_TRANSCRIPTION = REPO_PATH + 'checkpoints/piano_transcription/note_F1=0.9677_pedal_F1=0.9186.pth'  # transcriptor chkpt path
FPS = 16000
RANDOM_CROP = True  # whether to use random crops in case of cropped audio
CROP_LEN = 26 * 60

# params midi scrubbing and processing
MAX_DEPTH = 5  # max depth when searching in folders for audio files
MAX_GAP_IN_SONG = 10 # in secs
MIN_LEN = 20  # actual min len could go down to MIN_LEN - 2 * (REMOVE_FIRST_AND_LAST / 5)
MAX_LEN = 25 * 60  # maximum audio len for playlist downloads, and maximum audio length for transcription (in sec)
MIN_NB_NOTES = 80  # min nb of notes per minute of recording
REMOVE_FIRST_AND_LAST = 10  # will be divided by 5 if cutting this makes the song fall below min len

# parameters encoding
NOISE_INJECTED = True
AUGMENTATION = True
NB_AUG = 4 if AUGMENTATION else 0
RANGE_NOTE_ON = 128
RANGE_NOTE_OFF = 128
RANGE_VEL = 32
RANGE_TIME_SHIFT = 100
MAX_EMBEDDING = RANGE_VEL + RANGE_NOTE_OFF + RANGE_TIME_SHIFT + RANGE_NOTE_ON
MAX_TEST_SIZE = 1000
CHECKSUM_PATH = REPO_PATH + 'data/music/midi/checksum.pickle'
CHUNK_SIZE = 512

ALL_AUGMENTATIONS = []
for p in [-3, -2, -1, 1, 2, 3]:
    ALL_AUGMENTATIONS.append((p))
ALL_AUGMENTATIONS = np.array(ALL_AUGMENTATIONS)

ALL_NOISE = []
for s in [-5, -2.5, 0, 2.5, 5]:
    for p in np.arange(-6, 7):
        if not ((s == 0) and (p==0)):
            ALL_NOISE.append((s, p))
ALL_NOISE = np.array(ALL_NOISE)

# music transformer params
REP_MODEL_NAME = REPO_PATH + "checkpoints/music_representation/sentence_embedding/smallbert_b256_r128_1/best_model"
MUSIC_REP_PATH = REPO_PATH + "checkpoints/b256_r128_represented/"
MUSIC_NN_PATH = REPO_PATH + "checkpoints/music_representation/b256_r128_represented/nn_model.pickle"

TRANSLATION_VAE_CHKP_PATH = REPO_PATH + "checkpoints/music2cocktails/music2flavor/b256_r128_classif001_ld40_meanstd_regground2.5_egg_bubbles/"

# piano solo evaluation
# META_DATA_PIANO_EVAL_PATH = REPO_PATH + 'data/music/audio/is_piano.csv'
# CHKPT_PATH_PIANO_EVAL = REPO_PATH + 'data/checkpoints/piano_detection/piano_solo_model_32k.pth'