Sphereface2 configuration

[GoldMan6]

Hi, Great work and thanks for sharing !
i tried to train in redimnetB2 in wespeaker platform with sphereface2 projection and 15ms hop in features after your recommendations but i get worse results than by training with arc_margin. any idea what might be the problem ?
thanks

[didi1233]

Hi, Great work and thanks for sharing ! i tried to train in redimnetB2 in wespeaker platform with sphereface2 projection and 15ms hop in features after your recommendations but i get worse results than by training with arc_margin. any idea what might be the problem ? thanks

same question

[vanIvan]

Hi @GoldMan6 @didi1233 , we know that issue, that using wespeaker pipeline as-is it is not possible to achieve same results, but we haven't had an opportunity to dig into this problem. Could you @GoldMan6 please share training config file, you are using

[GoldMan6]

Hi @vanIvan, this is the config i used:
num_avg: 10
enable_amp: False # whether enable automatic mixed precision training

seed: 42
num_epochs: 120
save_epoch_interval: 5 # save model every 5 epochs
log_batch_interval: 100 # log every 100 batchs

dataloader_args:
batch_size: 200
num_workers: 16
pin_memory: false
prefetch_factor: 4
drop_last: true

dataset_args:
sample_num_per_epoch: 0
shuffle: True
shuffle_args:
shuffle_size: 2500
filter: True
filter_args:
min_num_frames: 100
max_num_frames: 800
resample_rate: 16000
speed_perturb: True
num_frms: 200
aug_prob: 0.6 # prob to add reverb & noise aug per sample
fbank_args:
num_mel_bins: 72
frame_shift: 15
frame_length: 25
dither: 1.0
spec_aug: False
spec_aug_args:
num_t_mask: 1
num_f_mask: 1
max_t: 10
max_f: 8
prob: 0.6

model: ReDimNetB2
model_init: null
model_args:
feat_dim: 72
embed_dim: 192
pooling_func: "ASTP" # TSTP, ASTP, MQMHASTP
two_emb_layer: False

projection_args:
project_type: "sphereface2" # add_margin, arc_margin, sphere, sphereface2, softmax, arc_margin_intertopk_subcenter
scale: 32.0
easy_margin: False

margin_scheduler: MarginScheduler
margin_update:
initial_margin: 0.0
final_margin: 0.2
increase_start_epoch: 20
fix_start_epoch: 40
update_margin: True
increase_type: "exp" # exp, linear
update_margin: true

loss: CrossEntropyLoss
loss_args: {}

optimizer: SGD
optimizer_args:
momentum: 0.9
nesterov: True
weight_decay: 2.0e-05

scheduler: ExponentialDecrease
scheduler_args:
initial_lr: 0.1
final_lr: 0.00005
warm_up_epoch: 6
warm_from_zero: True

thanks!

[vanIvan]

@GoldMan6 thank you, looks correct. Based on this config, everything should be fine, except the fact, that our fork of wespeaker is different in some way from original, in feature extraction, in fact that we have added separate argument for head weight decay multiplier. We'll consider sharing our modified wespeaker pipeline in proximate future.

[GoldMan6]

Thanks, I will try to look at the fork, is the difference in the feature extraction also here ?
Another question, I saw that you've tested the new models you've released on VOICES dataset. can you please share the verification test file you used ? Thanks a lot !