rosvot.yaml

base_config:
  - ./base.yaml

# This dataset configuration below is for M4Singer. Feel free to change to other datasets.
raw_data_dir: ''
processed_data_dir: data/processed/m4
binary_data_dir: data/binary/m4
metafile_path: data/processed/m4/metadata.json
ds_names: m4    # ds_names during binarizing, not necessary during training
ds_names_in_training: ''  # default: '', means all in training   ';' separated
ds_names_in_testing: ''   # default: '', means all in testing
noise_data_dir: data/binary/musan_24k

binarizer_cls: data_gen.rosvot_binarizer.RosvotBinarizer
binarization_args:
  with_wav: true
  with_mel: false

test_prefixes: [
  "Alto-3",
  "Alto-6#白羊",
  "Alto-7#此生不换",
  "Bass-1#半夏",
  "Bass-2#DEAR JOHN",
  "Bass-3#匆匆那年",
  "Soprano-2#蒲公英的约定",
  "Soprano-3#爱的供养",
  "Tenor-3#K歌之王",
  "Tenor-4",
  "Tenor-6#阿楚姑娘"
]

audio_sample_rate: 24000
fft_size: 512  # Extra window size is filled with 0 paddings to match this parameter
win_size: 512
hop_size: 128
fmin: 30
fmax: 12000
f0_min: 50
f0_max: 900
dropout: 0.1
hidden_size: 256

task_cls: tasks.rosvot.task.MidiExtractorTask

###########
# model related params
###########
model: rosvot
conformer_kernel: 9
updown_rates: 2-2-2-2
channel_multiples: 1-1-1-1
bkb_net: conformer   # conv|conformer|transformer|wavenet
bkb_layers: 2
unet_skip_layer: false

###########
# note prediction related hparams
###########
# f0
f0_add_noise: gaussian:0.04  # gaussian:0.04
# pitch
pe: rmvpe  # rmvpe|parselmouth|pw
pe_ckpt: checkpoints/rmvpe/model.pt
f0_bin: 512
f0_filepath: ''
pitch_type: frame
use_pitch_embed: true
use_soft_note: false
note_start: 30
note_num: 85
note_pitch_label_smoothing: 0.005
note_pitch_start: 0
note_pitch_temperature: 0.01
pitch_attn_num_head: 4
# note bd
min_note_dur: 80  # ms
note_bd_threshold: 0.8
note_bd_min_gap: 90  # ms
note_bd_ref_min_gap: 40    # ms. note_bd_ref_min_gap <= note_bd_min_gap / 2
use_soft_note_bd: true
soft_note_bd_func: gaussian:80
note_bd_ratio: 2.42312  # the avg number of note bd in 1 sec
note_bd_add_noise: gaussian:0.002 # gaussian|none
note_bd_start: 0
note_bd_temperature: 0.2
note_bd_focal_loss: 5.0   # the gamma value. [gamma]|none
# lambdas
lambda_note_bd: 1.0
lambda_note_pitch: 1.0
lambda_note_bd_focal: 3.0
lambda_note_bd_slur_punish: 0.0   # punish the over fractured note, such as slurs (or reduce label_pos_weight_decay)

label_pos_weight_decay: 0.95

###########
# noise
###########
noise_snr: 6-20
noise_prob: 0.8
noise_in_test: false

use_mel: true
use_mel_bins: 40
use_wav: false
use_wbd: true
min_word_dur: 20  # ms
mel_add_noise: gaussian:0.05 # gaussian|none|musan
frames_multiple: 16
dataset_downsample_rate: 1.0

lr: 0.00001
scheduler: step_lr
scheduler_lr_step_size: 500
warmup_updates: 0

max_epochs: 1000
max_updates: 60000
max_tokens: 60000
max_sentences: 32
val_check_interval: 1000
save_best: true
num_valid_plots: 10
num_valid_stats: 100
accumulate_grad_batches: 1
find_unused_parameters: true
pin_memory: true
ds_workers: 8

# inference
input_process_name: none
save_plot: true
infer_meta_path: ''
infer_regulate_real_note_itv: true
infer_print_skipped: true