WHAMR!: sep_reverb_noisy 1dB lower than the reported performance about asteroid HOT 7 CLOSED

The log is similar with mine. I think I can reproduce the results with more training epochs.
Thanks for your patience !

from asteroid.

Can you add you're training config please? I'll look for the log today.

from asteroid.

Thanks ! This is my config:
data:
mode: min
nondefault_nsrc: null
sample_rate: 8000
task: sep_reverb_noisy
train dir: data/wav8k/min/tr
valid dir: data/wav8k/min/cv
filterbank:
kernel size: 40
n_filters: 512
stride: 20
main_args:
exp_dir: exp/train_tasnet_sep_reverb_noisy_8kmin_db50b168
help: null
masknet:
dropout: 0.3
n_layers: 4
n_src: 2
n_units: 600
optim:
lr: 0.001
optimizer: adam
weight_decay: 1.0e-05
training:
batch size: 32
early_stoping: true
epochs: 200
half_lr: true
num_workers: 16

My final loss on the validation set is about -3.98 (SI-SNR)

from asteroid.

Here are the log files : logs.tar.gz

Here is the conf

data:
  mode: min
  nondefault_nsrc: null
  sample_rate: 8000
  task: sep_reverb_noisy
  train_dir: data/wav8k/min/tr
  valid_dir: data/wav8k/min/cv
filterbank:
  kernel_size: 40
  n_filters: 512
  stride: 20
main_args:
  exp_dir: exp/train_tasnet_sep_reverb_noisy_plain_pl0.6/
  gpus: '-1'
  help: null
masknet:
  dropout: 0.3
  n_layers: 4
  n_src: 2
  n_units: 600
optim:
  lr: 0.001
  optimizer: adam
  weight_decay: 1.0e-05
positional arguments: {}
training:
  batch_size: 32
  early_stop: true
  epochs: 200
  half_lr: true
  num_workers: 4

In pytorch lightning 0.6.0, early stopping was broken so training didn't stop and the performance improved. Could you also try without early stopping to see if this is better please?
I need to make the patience higher IMO

from asteroid.

Could you share your log here please? I'm curious when the early stopping happens

from asteroid.

Sure. I will download the log from the lab tomorrow

from asteroid.

Sorry for late reply. This is my log.
train_tasnet_sep_reverb_noisy_8kmin_db50b168.tar.gz

from asteroid.