Merge branch 'master' into lachesis

hyperion-ml · Jun 10, 2021 · fe2e1b8 · fe2e1b8
2 parents fbc0c83 + b8a0948
commit fe2e1b8
Show file tree

Hide file tree

Showing 62 changed files with 4,141 additions and 34 deletions.
diff --git a/egs/sre20-cts/v1/conf/linfbank64_stmn_8k.pyconf b/egs/sre20-cts/v1/conf/linfbank64_stmn_8k.pyconf
@@ -0,0 +1,20 @@
+--feats-audio-feat
+logfb
+--feats-sample-frequency
+8000
+--feats-frame-length
+25
+--feats-fb-type
+linear
+--feats-low-freq
+20 
+--feats-high-freq
+3700
+--feats-num-filters
+64
+--feats-snip-edges
+false
+--feats-use-energy
+false
+--mvn-context
+150
diff --git a/...fig_fbank64_stmn_effnetb4_v2_arcs30m0.3_trn_alllangs_nocv_nocnceleb_adam_lr0.01.amp.v1.sh b/...fig_fbank64_stmn_effnetb4_v2_arcs30m0.3_trn_alllangs_nocv_nocnceleb_adam_lr0.01.amp.v1.sh
@@ -40,18 +40,18 @@ nnet=$nnet_dir/model_ep0060.pth
 ft_batch_size_1gpu=4
 ft_eff_batch_size=128 # effective batch size
 ft_min_chunk=10
-ft_max_chunk=60
+ft_max_chunk=10
 ft_ipe=1
 ft_lr=0.05
-ft_nnet_num_epochs=21
+ft_nnet_num_epochs=30
 ft_margin=0.3
 ft_margin_warmup=3
 
 ft_opt_opt="--opt-optimizer sgd --opt-lr $ft_lr --opt-momentum 0.9 --opt-weight-decay 1e-5 --use-amp --var-batch-size"
 ft_lrs_opt="--lrsch-lrsch-type cos_lr --lrsch-t 2500 --lrsch-t-mul 2 --lrsch-warm-restarts --lrsch-gamma 0.75 --lrsch-min-lr 1e-4 --lrsch-warmup-steps 100 --lrsch-update-lr-on-opt-step"
 ft_nnet_name=${nnet_name}.ft_${ft_min_chunk}_${ft_max_chunk}_arcm${ft_margin}_sgdcos_lr${ft_lr}_b${ft_eff_batch_size}_amp.v2
 ft_nnet_dir=exp/xvector_nnets/$ft_nnet_name
-ft_nnet=$ft_nnet_dir/model_ep0020.pth
+ft_nnet=$ft_nnet_dir/model_ep0030.pth
 
 
 # xvector last-layer finetuning alllangs

diff --git a/...stmn_res2net50w26s4_eina_hln_arcs30m0.3_trn_alllangs_nocv_nocnceleb_adam_lr0.01.amp.v1.sh b/...stmn_res2net50w26s4_eina_hln_arcs30m0.3_trn_alllangs_nocv_nocnceleb_adam_lr0.01.amp.v1.sh
@@ -0,0 +1,116 @@
+# Res2Net50 x-vector with mixed precision training
+
+# acoustic features
+feat_config=conf/fbank64_mvn_8k.pyconf
+feat_type=fbank64_stmn
+
+
+# x-vector training 
+nnet_data=alllangs_nocv_nocnceleb
+nnet_num_augs=4
+aug_opt="--train-aug-cfg conf/reverb_noise_aug.yml --val-aug-cfg conf/reverb_noise_aug.yml"
+
+batch_size_1gpu=8
+eff_batch_size=512 # effective batch size
+ipe=$nnet_num_augs
+min_chunk=4
+max_chunk=4
+lr=0.01
+
+nnet_type=res2net50 
+dropout=0
+embed_dim=256
+width_factor=1.625
+scale=4
+ws_tag=w26s4
+
+s=30
+margin_warmup=20
+margin=0.3
+
+nnet_opt="--resnet-type $nnet_type --in-feats 64 --in-channels 1 --in-kernel-size 3 --in-stride 1 --no-maxpool --norm-layer instance-norm-affine --head-norm-layer layer-norm  --no-maxpool --res2net-width-factor $width_factor --res2net-scale $scale"
+
+opt_opt="--opt-optimizer adam --opt-lr $lr --opt-beta1 0.9 --opt-beta2 0.95 --opt-weight-decay 1e-5 --opt-amsgrad" # --use-amp"
+lrs_opt="--lrsch-lrsch-type exp_lr --lrsch-decay-rate 0.5 --lrsch-decay-steps 10000 --lrsch-hold-steps 40000 --lrsch-min-lr 1e-5 --lrsch-warmup-steps 1000 --lrsch-update-lr-on-opt-step"
+
+nnet_name=${feat_type}_${nnet_type}${ws_tag}_eina_hln_e${embed_dim}_arcs${s}m${margin}_do${dropout}_adam_lr${lr}_b${eff_batch_size}_amp.v1.$nnet_data
+nnet_num_epochs=50
+nnet_dir=exp/xvector_nnets/$nnet_name
+nnet=$nnet_dir/model_ep0050.pth
+
+
+# xvector full net finetuning with out-of-domain
+ft_batch_size_1gpu=4
+ft_eff_batch_size=128 # effective batch size
+ft_min_chunk=10
+ft_max_chunk=20
+ft_ipe=1
+ft_lr=0.05
+ft_nnet_num_epochs=21
+ft_nnet_num_epochs=45
+ft_margin=0.3
+ft_margin_warmup=3
+
+ft_opt_opt="--opt-optimizer sgd --opt-lr $ft_lr --opt-momentum 0.9 --opt-weight-decay 1e-5 --use-amp --var-batch-size"
+ft_lrs_opt="--lrsch-lrsch-type cos_lr --lrsch-t 2500 --lrsch-t-mul 2 --lrsch-warm-restarts --lrsch-gamma 0.75 --lrsch-min-lr 1e-4 --lrsch-warmup-steps 100 --lrsch-update-lr-on-opt-step"
+ft_nnet_name=${nnet_name}.ft_${ft_min_chunk}_${ft_max_chunk}_arcm${ft_margin}_sgdcos_lr${ft_lr}_b${ft_eff_batch_size}_amp.v2
+ft_nnet_dir=exp/xvector_nnets/$ft_nnet_name
+ft_nnet=$ft_nnet_dir/model_ep0014.pth
+
+
+# xvector last-layer finetuning realtel
+reg_layers_classif=0
+reg_layers_enc="0 1 2 3 4"
+nnet_adapt_data=realtel
+ft2_batch_size_1gpu=16
+ft2_eff_batch_size=128 # effective batch size
+ft2_ipe=1
+ft2_lr=0.01
+ft2_nnet_num_epochs=35
+ft2_margin_warmup=3
+ft2_reg_weight_embed=0.1
+ft2_min_chunk=10
+ft2_max_chunk=60
+
+ft2_opt_opt="--opt-optimizer sgd --opt-lr $ft2_lr --opt-momentum 0.9 --opt-weight-decay 1e-5 --use-amp --var-batch-size"
+ft2_lrs_opt="--lrsch-lrsch-type cos_lr --lrsch-t 2500 --lrsch-t-mul 2 --lrsch-warm-restarts --lrsch-gamma 0.75 --lrsch-min-lr 1e-4 --lrsch-warmup-steps 100 --lrsch-update-lr-on-opt-step"
+ft2_nnet_name=${ft_nnet_name}.ft_eaffine_rege_w${ft2_reg_weight_embed}_${ft2_min_chunk}_${ft2_max_chunk}_sgdcos_lr${ft2_lr}_b${ft2_eff_batch_size}_amp.v2.$nnet_adapt_data
+ft2_nnet_dir=exp/xvector_nnets/$ft2_nnet_name
+ft2_nnet=$ft2_nnet_dir/model_ep0015.pth
+
+
+# xvector full nnet finetuning
+ft3_batch_size_1gpu=2
+ft3_eff_batch_size=128 # effective batch size
+ft3_ipe=1
+ft3_lr=0.01
+ft3_nnet_num_epochs=10
+ft3_margin_warmup=20
+ft3_reg_weight_embed=0.1
+ft3_reg_weight_enc=0.1
+ft3_min_chunk=10
+ft3_max_chunk=60
+
+ft3_opt_opt="--opt-optimizer sgd --opt-lr $ft3_lr --opt-momentum 0.9 --opt-weight-decay 1e-5 --use-amp --var-batch-size"
+ft3_lrs_opt="--lrsch-lrsch-type cos_lr --lrsch-t 2500 --lrsch-t-mul 2 --lrsch-warm-restarts --lrsch-gamma 0.75 --lrsch-min-lr 1e-4 --lrsch-warmup-steps 100 --lrsch-update-lr-on-opt-step"
+ft3_nnet_name=${ft2_nnet_name}.ft_reg_wenc${ft3_reg_weight_enc}_we${ft3_reg_weigth_embed}_${ft3_min_chunk}_${ft3_max_chunk}_sgdcos_lr${ft3_lr}_b${ft3_eff_batch_size}_amp.v2
+ft3_nnet_name=${ft2_nnet_name}.ft_${ft3_min_chunk}_${ft3_max_chunk}_sgdcos_lr${ft3_lr}_b${ft3_eff_batch_size}_amp.v2
+ft3_nnet_dir=exp/xvector_nnets/$ft3_nnet_name
+ft3_nnet=$ft3_nnet_dir/model_ep0010.pth
+
+
+# back-end
+plda_aug_config=conf/noise_aug.yml
+plda_num_augs=0
+# if [ $plda_num_augs -eq 0 ]; then
+#     plda_data=sre_tel
+#     plda_adapt_data=sre18_cmn2_adapt_lab
+# else
+#     plda_data=sre_tel_augx${plda_num_augs}
+#     plda_adapt_data=sre18_cmn2_adapt_lab_augx${plda_num_augs}
+# fi
+# plda_type=splda
+# lda_dim=200
+# plda_y_dim=150
+# plda_z_dim=200
+
diff --git a/...snet34_eina_hln_bmhah64d8192_arcs30m0.3_trn_alllangs_nocv_nocnceleb_adam_lr0.01.amp.v2.sh b/...snet34_eina_hln_bmhah64d8192_arcs30m0.3_trn_alllangs_nocv_nocnceleb_adam_lr0.01.amp.v2.sh
@@ -0,0 +1,112 @@
+# ResNet34 x-vector with mixed precision training
+
+# acoustic features
+feat_config=conf/fbank64_mvn_8k.pyconf
+feat_type=fbank64_stmn
+
+
+# x-vector training 
+nnet_data=alllangs_nocv_nocnceleb
+nnet_num_augs=4
+aug_opt="--train-aug-cfg conf/reverb_noise_aug.yml --val-aug-cfg conf/reverb_noise_aug.yml"
+
+batch_size_1gpu=32
+eff_batch_size=512 # effective batch size
+ipe=$nnet_num_augs
+min_chunk=4
+max_chunk=4
+lr=0.01
+
+nnet_type=resnet34 
+dropout=0
+embed_dim=256
+
+s=30
+margin_warmup=20
+margin=0.3
+
+nnet_opt="--resnet-type $nnet_type --in-feats 64 --in-channels 1 --in-kernel-size 3 --in-stride 1 --no-maxpool --norm-layer instance-norm-affine --head-norm-layer layer-norm --pool-type scaled-dot-prod-att-v1 --pool-num-heads 64 --pool-d-k 128 --pool-d-v 128 --pool-bin-attn"
+
+opt_opt="--opt-optimizer adam --opt-lr $lr --opt-beta1 0.9 --opt-beta2 0.95 --opt-weight-decay 1e-5 --opt-amsgrad" # --use-amp"
+lrs_opt="--lrsch-lrsch-type exp_lr --lrsch-decay-rate 0.5 --lrsch-decay-steps 10000 --lrsch-hold-steps 40000 --lrsch-min-lr 1e-5 --lrsch-warmup-steps 1000 --lrsch-update-lr-on-opt-step"
+
+nnet_name=${feat_type}_${nnet_type}_eina_hln_bmhah64d8192_e${embed_dim}_arcs${s}m${margin}_do${dropout}_adam_lr${lr}_b${eff_batch_size}_amp.v1.$nnet_data
+nnet_num_epochs=50
+nnet_dir=exp/xvector_nnets/$nnet_name
+nnet=$nnet_dir/model_ep0050.pth
+
+
+# xvector full net finetuning with out-of-domain
+ft_batch_size_1gpu=4
+ft_eff_batch_size=128 # effective batch size
+ft_min_chunk=10
+ft_max_chunk=60
+ft_ipe=1
+ft_lr=0.05
+ft_nnet_num_epochs=21
+ft_margin=0.3
+ft_margin_warmup=3
+
+ft_opt_opt="--opt-optimizer sgd --opt-lr $ft_lr --opt-momentum 0.9 --opt-weight-decay 1e-5 --use-amp --var-batch-size"
+ft_lrs_opt="--lrsch-lrsch-type cos_lr --lrsch-t 2500 --lrsch-t-mul 2 --lrsch-warm-restarts --lrsch-gamma 0.75 --lrsch-min-lr 1e-4 --lrsch-warmup-steps 100 --lrsch-update-lr-on-opt-step"
+ft_nnet_name=${nnet_name}.ft_${ft_min_chunk}_${ft_max_chunk}_arcm${ft_margin}_sgdcos_lr${ft_lr}_b${ft_eff_batch_size}_amp.v2
+ft_nnet_dir=exp/xvector_nnets/$ft_nnet_name
+ft_nnet=$ft_nnet_dir/model_ep0021.pth
+
+
+# xvector last-layer finetuning realtel
+reg_layers_classif=0
+reg_layers_enc="0 1 2 3 4"
+nnet_adapt_data=realtel
+ft2_batch_size_1gpu=16
+ft2_eff_batch_size=128 # effective batch size
+ft2_ipe=1
+ft2_lr=0.01
+ft2_nnet_num_epochs=35
+ft2_margin_warmup=3
+ft2_reg_weight_embed=0.1
+ft2_min_chunk=10
+ft2_max_chunk=60
+
+ft2_opt_opt="--opt-optimizer sgd --opt-lr $ft2_lr --opt-momentum 0.9 --opt-weight-decay 1e-5 --use-amp --var-batch-size"
+ft2_lrs_opt="--lrsch-lrsch-type cos_lr --lrsch-t 2500 --lrsch-t-mul 2 --lrsch-warm-restarts --lrsch-gamma 0.75 --lrsch-min-lr 1e-4 --lrsch-warmup-steps 100 --lrsch-update-lr-on-opt-step"
+ft2_nnet_name=${ft_nnet_name}.ft_eaffine_rege_w${ft2_reg_weight_embed}_${ft2_min_chunk}_${ft2_max_chunk}_sgdcos_lr${ft2_lr}_b${ft2_eff_batch_size}_amp.v2.$nnet_adapt_data
+ft2_nnet_dir=exp/xvector_nnets/$ft2_nnet_name
+ft2_nnet=$ft2_nnet_dir/model_ep0015.pth
+
+
+# xvector full nnet finetuning
+ft3_batch_size_1gpu=2
+ft3_eff_batch_size=128 # effective batch size
+ft3_ipe=1
+ft3_lr=0.01
+ft3_nnet_num_epochs=10
+ft3_margin_warmup=20
+ft3_reg_weight_embed=0.1
+ft3_reg_weight_enc=0.1
+ft3_min_chunk=10
+ft3_max_chunk=60
+
+ft3_opt_opt="--opt-optimizer sgd --opt-lr $ft3_lr --opt-momentum 0.9 --opt-weight-decay 1e-5 --use-amp --var-batch-size"
+ft3_lrs_opt="--lrsch-lrsch-type cos_lr --lrsch-t 2500 --lrsch-t-mul 2 --lrsch-warm-restarts --lrsch-gamma 0.75 --lrsch-min-lr 1e-4 --lrsch-warmup-steps 100 --lrsch-update-lr-on-opt-step"
+ft3_nnet_name=${ft2_nnet_name}.ft_reg_wenc${ft3_reg_weight_enc}_we${ft3_reg_weigth_embed}_${ft3_min_chunk}_${ft3_max_chunk}_sgdcos_lr${ft3_lr}_b${ft3_eff_batch_size}_amp.v2
+ft3_nnet_name=${ft2_nnet_name}.ft_${ft3_min_chunk}_${ft3_max_chunk}_sgdcos_lr${ft3_lr}_b${ft3_eff_batch_size}_amp.v2
+ft3_nnet_dir=exp/xvector_nnets/$ft3_nnet_name
+ft3_nnet=$ft3_nnet_dir/model_ep0010.pth
+
+
+# back-end
+plda_aug_config=conf/noise_aug.yml
+plda_num_augs=0
+# if [ $plda_num_augs -eq 0 ]; then
+#     plda_data=sre_tel
+#     plda_adapt_data=sre18_cmn2_adapt_lab
+# else
+#     plda_data=sre_tel_augx${plda_num_augs}
+#     plda_adapt_data=sre18_cmn2_adapt_lab_augx${plda_num_augs}
+# fi
+# plda_type=splda
+# lda_dim=200
+# plda_y_dim=150
+# plda_z_dim=200
+
diff --git a/...et34_eina_hln_chattstatsi128_arcs30m0.3_trn_alllangs_nocv_nocnceleb_adam_lr0.01.amp.v1.sh b/...et34_eina_hln_chattstatsi128_arcs30m0.3_trn_alllangs_nocv_nocnceleb_adam_lr0.01.amp.v1.sh
@@ -0,0 +1,114 @@
+# ResNet34 x-vector with mixed precision training
+
+# acoustic features
+feat_config=conf/fbank64_mvn_8k.pyconf
+feat_type=fbank64_stmn
+
+
+# x-vector training 
+nnet_data=alllangs_nocv_nocnceleb
+nnet_num_augs=4
+aug_opt="--train-aug-cfg conf/reverb_noise_aug.yml --val-aug-cfg conf/reverb_noise_aug.yml"
+
+batch_size_1gpu=32
+eff_batch_size=512 # effective batch size
+ipe=$nnet_num_augs
+min_chunk=4
+max_chunk=4
+lr=0.01
+
+nnet_type=resnet34 
+dropout=0
+embed_dim=256
+
+s=30
+margin_warmup=20
+margin=0.3
+
+attstats_inner=128
+
+nnet_opt="--resnet-type $nnet_type --in-feats 64 --in-channels 1 --in-kernel-size 3 --in-stride 1 --no-maxpool --norm-layer instance-norm-affine --head-norm-layer layer-norm --pool-type ch-wise-att-mean-stddev --pool-inner-feats $attstats_inner"
+
+opt_opt="--opt-optimizer adam --opt-lr $lr --opt-beta1 0.9 --opt-beta2 0.95 --opt-weight-decay 1e-5 --opt-amsgrad" # --use-amp"
+lrs_opt="--lrsch-lrsch-type exp_lr --lrsch-decay-rate 0.5 --lrsch-decay-steps 10000 --lrsch-hold-steps 40000 --lrsch-min-lr 1e-5 --lrsch-warmup-steps 1000 --lrsch-update-lr-on-opt-step"
+
+nnet_name=${feat_type}_${nnet_type}_eina_hln_chattstatsi128_e${embed_dim}_arcs${s}m${margin}_do${dropout}_adam_lr${lr}_b${eff_batch_size}_amp.v1.$nnet_data
+nnet_num_epochs=50
+nnet_dir=exp/xvector_nnets/$nnet_name
+nnet=$nnet_dir/model_ep0050.pth
+
+
+# xvector full net finetuning with out-of-domain
+ft_batch_size_1gpu=4
+ft_eff_batch_size=128 # effective batch size
+ft_min_chunk=10
+ft_max_chunk=60
+ft_ipe=1
+ft_lr=0.05
+ft_nnet_num_epochs=21
+ft_margin=0.3
+ft_margin_warmup=3
+
+ft_opt_opt="--opt-optimizer sgd --opt-lr $ft_lr --opt-momentum 0.9 --opt-weight-decay 1e-5 --use-amp --var-batch-size"
+ft_lrs_opt="--lrsch-lrsch-type cos_lr --lrsch-t 2500 --lrsch-t-mul 2 --lrsch-warm-restarts --lrsch-gamma 0.75 --lrsch-min-lr 1e-4 --lrsch-warmup-steps 100 --lrsch-update-lr-on-opt-step"
+ft_nnet_name=${nnet_name}.ft_${ft_min_chunk}_${ft_max_chunk}_arcm${ft_margin}_sgdcos_lr${ft_lr}_b${ft_eff_batch_size}_amp.v2
+ft_nnet_dir=exp/xvector_nnets/$ft_nnet_name
+ft_nnet=$ft_nnet_dir/model_ep0021.pth
+
+
+# xvector last-layer finetuning realtel
+reg_layers_classif=0
+reg_layers_enc="0 1 2 3 4"
+nnet_adapt_data=realtel
+ft2_batch_size_1gpu=16
+ft2_eff_batch_size=128 # effective batch size
+ft2_ipe=1
+ft2_lr=0.01
+ft2_nnet_num_epochs=35
+ft2_margin_warmup=3
+ft2_reg_weight_embed=0.1
+ft2_min_chunk=10
+ft2_max_chunk=60
+
+ft2_opt_opt="--opt-optimizer sgd --opt-lr $ft2_lr --opt-momentum 0.9 --opt-weight-decay 1e-5 --use-amp --var-batch-size"
+ft2_lrs_opt="--lrsch-lrsch-type cos_lr --lrsch-t 2500 --lrsch-t-mul 2 --lrsch-warm-restarts --lrsch-gamma 0.75 --lrsch-min-lr 1e-4 --lrsch-warmup-steps 100 --lrsch-update-lr-on-opt-step"
+ft2_nnet_name=${ft_nnet_name}.ft_eaffine_rege_w${ft2_reg_weight_embed}_${ft2_min_chunk}_${ft2_max_chunk}_sgdcos_lr${ft2_lr}_b${ft2_eff_batch_size}_amp.v2.$nnet_adapt_data
+ft2_nnet_dir=exp/xvector_nnets/$ft2_nnet_name
+ft2_nnet=$ft2_nnet_dir/model_ep0015.pth
+
+
+# xvector full nnet finetuning
+ft3_batch_size_1gpu=2
+ft3_eff_batch_size=128 # effective batch size
+ft3_ipe=1
+ft3_lr=0.01
+ft3_nnet_num_epochs=10
+ft3_margin_warmup=20
+ft3_reg_weight_embed=0.1
+ft3_reg_weight_enc=0.1
+ft3_min_chunk=10
+ft3_max_chunk=60
+
+ft3_opt_opt="--opt-optimizer sgd --opt-lr $ft3_lr --opt-momentum 0.9 --opt-weight-decay 1e-5 --use-amp --var-batch-size"
+ft3_lrs_opt="--lrsch-lrsch-type cos_lr --lrsch-t 2500 --lrsch-t-mul 2 --lrsch-warm-restarts --lrsch-gamma 0.75 --lrsch-min-lr 1e-4 --lrsch-warmup-steps 100 --lrsch-update-lr-on-opt-step"
+ft3_nnet_name=${ft2_nnet_name}.ft_reg_wenc${ft3_reg_weight_enc}_we${ft3_reg_weigth_embed}_${ft3_min_chunk}_${ft3_max_chunk}_sgdcos_lr${ft3_lr}_b${ft3_eff_batch_size}_amp.v2
+ft3_nnet_name=${ft2_nnet_name}.ft_${ft3_min_chunk}_${ft3_max_chunk}_sgdcos_lr${ft3_lr}_b${ft3_eff_batch_size}_amp.v2
+ft3_nnet_dir=exp/xvector_nnets/$ft3_nnet_name
+ft3_nnet=$ft3_nnet_dir/model_ep0010.pth
+
+
+# back-end
+plda_aug_config=conf/noise_aug.yml
+plda_num_augs=0
+# if [ $plda_num_augs -eq 0 ]; then
+#     plda_data=sre_tel
+#     plda_adapt_data=sre18_cmn2_adapt_lab
+# else
+#     plda_data=sre_tel_augx${plda_num_augs}
+#     plda_adapt_data=sre18_cmn2_adapt_lab_augx${plda_num_augs}
+# fi
+# plda_type=splda
+# lda_dim=200
+# plda_y_dim=150
+# plda_z_dim=200
+