hyperparams.yaml

# Generated 2024-02-06 from:
# /mnt3/speechbrain/recipes/LibriMix/separation/hparams/sepformer-libri3mix.yaml
# yamllint disable
# ################################
# Model: SepFormer for source separation
# https://arxiv.org/abs/2010.13154
# Dataset : Libri3Mix
# ################################
#
# Basic parameters
# Seed needs to be set at top of yaml, before objects with parameters are made
#
seed: 1234
__set_seed: !apply:torch.manual_seed [1234]

# Data params

# e.g. '/yourpath/Libri3Mix/train-clean-360/'
# the data folder is needed even if dynamic mixing is applied
data_folder: /mnt3/Libri3Mix_48k_own/Libri3Mix/

# This is needed only if dynamic mixing is applied
base_folder_dm: /yourpath/LibriSpeech/train-clean-360/

experiment_name: sepformer-libri3mix-48k
output_folder: results/sepformer-libri3mix-48k/1234
train_log: results/sepformer-libri3mix-48k/1234/train_log.txt
save_folder: results/sepformer-libri3mix-48k/1234/save
train_data: results/sepformer-libri3mix-48k/1234/save/libri3mix_train-360.csv
valid_data: results/sepformer-libri3mix-48k/1234/save/libri3mix_test.csv
test_data: results/sepformer-libri3mix-48k/1234/save/libri3mix_test.csv
skip_prep: false

ckpt_interval_minutes: 60

# Experiment params
precision: fp16 # bf16, fp16 or fp32 # Set it to True for mixed precision
num_spks: 3
noprogressbar: false
save_audio: false # Save estimated sources on disk
sample_rate: 48000

# Training parameters
N_epochs: 50
batch_size: 1
lr: 0.0005
clip_grad_norm: 5
loss_upper_lim: 999999  # this is the upper limit for an acceptable loss
# if True, the training sequences are cut to a specified length
limit_training_signal_len: true
# this is the length of sequences if we choose to limit
# the signal length of training sequences
training_signal_len: 100000

# Set it to True to dynamically create mixtures at training time
dynamic_mixing: false
use_wham_noise: true

# Parameters for data augmentation
use_wavedrop: false
use_speedperturb: true
use_rand_shift: false
min_shift: -8000
max_shift: 8000

# Speed perturbation
speed_changes: &id001 [95, 100, 105]

# Frequency drop: randomly drops a number of frequency bands to zero.
speed_perturb: !new:speechbrain.augment.time_domain.SpeedPerturb
  orig_freq: 48000
  speeds: *id001
drop_freq_low: 0  # Min frequency band dropout probability
drop_freq_high: 1  # Max frequency band dropout probability
drop_freq_count_low: 1  # Min number of frequency bands to drop
drop_freq_count_high: 3  # Max number of frequency bands to drop
drop_freq_width: 0.05  # Width of frequency bands to drop

drop_freq: !new:speechbrain.augment.time_domain.DropFreq
  drop_freq_low: 0
  drop_freq_high: 1
  drop_freq_count_low: 1
  drop_freq_count_high: 3
  drop_freq_width: 0.05

# Time drop: randomly drops a number of temporal chunks.
drop_chunk_count_low: 1  # Min number of audio chunks to drop
drop_chunk_count_high: 5  # Max number of audio chunks to drop
drop_chunk_length_low: 1000  # Min length of audio chunks to drop
drop_chunk_length_high: 2000  # Max length of audio chunks to drop

drop_chunk: !new:speechbrain.augment.time_domain.DropChunk
  drop_length_low: 1000
  drop_length_high: 2000
  drop_count_low: 1
  drop_count_high: 5

# loss thresholding -- this thresholds the training loss
threshold_byloss: true
threshold: -30

# Encoder parameters
N_encoder_out: 256
out_channels: 256
kernel_size: 16
kernel_stride: 8
d_ffn: 1024

# Dataloader options
dataloader_opts:
  batch_size: 1
  num_workers: 3


# Specifying the network
Encoder: &id004 !new:speechbrain.lobes.models.dual_path.Encoder
  kernel_size: 16
  out_channels: 256


SBtfintra: &id002 !new:speechbrain.lobes.models.dual_path.SBTransformerBlock
  num_layers: 8
  d_model: 256
  nhead: 8
  d_ffn: 1024
  dropout: 0
  use_positional_encoding: true
  norm_before: true

SBtfinter: &id003 !new:speechbrain.lobes.models.dual_path.SBTransformerBlock
  num_layers: 8
  d_model: 256
  nhead: 8
  d_ffn: 1024
  dropout: 0
  use_positional_encoding: true
  norm_before: true

MaskNet: &id006 !new:speechbrain.lobes.models.dual_path.Dual_Path_Model

  num_spks: 3
  in_channels: 256
  out_channels: 256
  num_layers: 2
  K: 250
  intra_model: *id002
  inter_model: *id003
  norm: ln
  linear_layer_after_inter_intra: false
  skip_around_intra: true

Decoder: &id005 !new:speechbrain.lobes.models.dual_path.Decoder
  in_channels: 256
  out_channels: 1
  kernel_size: 16
  stride: 8
  bias: false

optimizer: !name:torch.optim.Adam
  lr: 0.0005
  weight_decay: 0

loss: !name:speechbrain.nnet.losses.get_si_snr_with_pitwrapper

lr_scheduler: !new:speechbrain.nnet.schedulers.ReduceLROnPlateau
  factor: 0.5
  patience: 2
  dont_halve_until_epoch: 5

epoch_counter: &id007 !new:speechbrain.utils.epoch_loop.EpochCounter
        # lr_scheduler: !ref <lr_scheduler>

  limit: 50

modules:
  encoder: *id004
  decoder: *id005
  masknet: *id006
checkpointer: !new:speechbrain.utils.checkpoints.Checkpointer
  checkpoints_dir: results/sepformer-libri3mix-48k/1234/save
  recoverables:
    encoder: *id004
    decoder: *id005
    masknet: *id006
    counter: *id007
train_logger: !new:speechbrain.utils.train_logger.FileTrainLogger
  save_file: results/sepformer-libri3mix-48k/1234/train_log.txt


pretrainer: !new:speechbrain.utils.parameter_transfer.Pretrainer
  loadables:
     encoder: !ref <Encoder>
     masknet: !ref <MaskNet>
     decoder: !ref <Decoder>