Spaces:
Running
Running
File size: 2,349 Bytes
a00b67a |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 |
# For De-limit task, Conv-TasNet.
# si_sdr loss
#
# ozone_train_fixed is about 6.36 hours
# 300,000 segments is about 333.33 hours
# ratio should be about 0.019
wandb_params:
use_wandb: true
entity: null # your wandb id
project: delimit # your wandb project
rerun_id: null # use when you rerun wandb.
sweep: false
sys_params:
nb_workers: 4
seed: 777
n_nodes: 1
port: null
rank: 0
task_params:
target: all # choices=["all"]
train: true
dataset: delimit # choices=["musdb", "delimit"]
dir_params:
root: /path/to/musdb18hq
output_directory: /path/to/results
exp_name: convtasnet_6_s # you MUST specify this
resume: null # "path of checkpoint folder"
continual_train: false # when we want to use a pre-trained model but not want to use lr_scheduler history.
delimit_valid_root: null
delimit_valid_L_root: null
ozone_root: /path/to/musdb-XL-train # you have to specify data_params.use_fixed
hyperparams:
batch_size: 8 # with 1 gpus (we used 2080ti 11GB)
epochs: 200
optimizer: adamw
weight_decay: 0.01
lr: 0.00003
lr_decay_gamma: 0.5
lr_decay_patience: 15
patience: 50
lr_scheduler: step_lr
gradient_clip: 5.0
ema: false
data_params:
nfft: 4096
nhop: 1024
nb_channels: 2
sample_rate: 44100
seq_dur: 4.0
singleset_num_frames: null
samples_per_track: 128 # "Number of samples per track to use for training."
limitaug_method: ozone
limitaug_mode: null
limitaug_custom_target_lufs: null
limitaug_custom_target_lufs_std: null
target_loudnorm_lufs: -14.0
random_mix: true
target_limitaug_mode: null
target_limitaug_custom_target_lufs: null
target_limitaug_custom_target_lufs_std: null
custom_limiter_attack_range: null
custom_limiter_release_range: null
use_fixed: 0.019 # range 0.0 ~ 1.0 => 1.0 will use fixed Ozoned_mixture training examples only.
model_loss_params:
architecture: conv_tasnet_mask_on_output # Sample-wise Gain Inversion (SGI)
train_loss_func: [si_sdr]
train_loss_scales: [1.]
valid_loss_func: [si_sdr]
valid_loss_scales: [1.]
conv_tasnet_params:
encoder_activation: relu
n_filters: 512
kernel_size: 128 # about 3ms in 44100Hz
stride: 64
n_blocks: 5
n_repeats: 2
bn_chan: 128
hid_chan: 512
skip_chan: 128
# conv_kernel_size:
# norm_type:
mask_act: relu
# causal:
decoder_activation: sigmoid
|