add vits configs

bin/conformer_ppg_feature_extract_multi_jobs.sh

@@ -14,7 +14,7 @@ for split in $splits ; do
     speakers=$(cat data/$dataset/$split/speakers.txt)
     for spk in $speakers ; do 
         b=$script_dir/conformer_ppg_feature_extraction_${split}_${spk}.sh
-        l=logs/enc_dec_conformer_ppg_feature_extraction${split}_${spk}.log
+        l=logs/enc_dec_conformer_ppg_feature_extraction_${split}_${spk}.log
         cat <<EOF > $b
 #!/bin/bash
 source $conda/bin/activate $conda_env

bin/hubertsoft_feature_extraction_multi_jobs.sh

@@ -2,30 +2,34 @@
 
 conda=/share/mini1/sw/std/python/anaconda3-2019.07/v3.7
 conda_env=torch_1.9
-splits="train_nodev_clean dev_clean eval_clean"
+splits="train_nodev_all dev_all eval_all"
+dataset=vctk
 
-script_dir=scripts/libritts/hubert_soft
+
+script_dir=scripts/$dataset/hubert_soft
 [ ! -e $script_dir ]  && mkdir -p  $script_dir 
 [ ! -e logs ] && mkdir logs
+
+
 for split in $splits ; do
 
-    echo "[hubert_soft feature extraction]: $split for libritts"
-    speakers=$(cat data/libritts/$split/speakers.txt)
+    echo "[hubert_soft feature extraction]: $split for ${dataset}"
+    speakers=$(cat data/$dataset/$split/speakers.txt)
     for spk in $speakers ; do 
-        b=$script_dir/hubertsoft_feature_extraction_${spk}.sh
-        l=logs/enc_dec_hubertsoft_feature_extraction.${spk}.log
+        b=$script_dir/hubertsoft_feature_extraction_${split}_${spk}.sh
+        l=logs/hubertsoft_feature_extraction_${split}_${spk}.log
         cat <<EOF > $b
 #!/bin/bash
 source $conda/bin/activate $conda_env
 python3 ling_encoder/hubert_soft/extract_features.py \
-    --metadata data/libritts/$split/metadata.csv \
-    --dump_dir dump/libritts \
+    --metadata data/$dataset/$split/metadata.csv \
+    --dump_dir dump/$dataset \
     --split $split \
     --max_workers 20 \
     --speaker $spk
 EOF
     chmod +x $b
     submitjob -m 10000 $l $b
-    echo "submitjob for $spk"
+    echo "submitjob for [$dataset $split] [$spk], see log in $l"
     done
 done        

configs/vctk_conformerppg_uttdvec_ppgvcf0_vits_none.yaml

@@ -0,0 +1,100 @@
+# experiment
+dataset: vctk
+train_meta: data/vctk/train_nodev_all/metadata.csv
+dev_meta: data/vctk/dev_all/metadata.csv
+train_set: train_nodev_all
+dev_set: dev_all
+
+
+# encoder-decoder
+ling_enc: conformer_ppg
+spk_enc: utt_dvec
+pros_enc: ppgvc_f0
+decoder: VITS
+mel_type: vits_spec
+
+
+# training 
+fp16_run: !!bool False
+epochs: 200
+save_freq: 1 # save ckpt frequency
+show_freq: 100 # show training information frequency
+load_only_params: !!bool False
+seed: !!int 1234
+trainer: VITSTrainer
+ngpu: 2
+
+#dataloader
+dataset_class: VITSDataset
+sampling_rate: !!int 24000
+vits_hop_size: !!int 240
+spec_max_len: !!int 240
+sort: !!bool True
+dump_dir: dump
+num_workers: !!int 4
+batch_size: !!int 12
+drop_last: !!bool True
+rm_long_utt: !!bool False # remove too long utterances from metadata
+max_utt_duration: !!float 10.0 # max utterance duration (seconds)
+
+
+# decoder params
+decoder_params: 
+    spk_emb_dim: 256
+    prosodic_rep_type: continuous
+    prosodic_net:
+        hidden_dim: 192
+    input_dim: !!int 144
+    spec_channels: !!int 513
+    inter_channels: !!int 192
+    hidden_channels: !!int 192
+    filter_channels: !!int 768
+    n_heads: !!int 2
+    n_layers: !!int 6
+    kernel_size: !!int 3
+    p_dropout: !!float 0.1
+    resblock : 1
+    resblock_kernel_sizes: [3,7,11]
+    resblock_dilation_sizes: [[1,3,5], [1,3,5], [1,3,5]]
+    upsample_rates: [10,6,2,2]
+    upsample_initial_channel: !!int 512
+    upsample_kernel_sizes: [20, 12, 4, 4]
+    n_layers_q: !!int 3
+    use_spectral_norm: !!bool False
+    filter_length: !!int 1024
+    n_mels_channels: !!int 80
+    win_length: !!int 1024
+    hop_length: !!int 240
+    sampling_rate: !!int 24000
+    segment_size: !!int 9600
+
+
+
+
+#optimizer & scheduler
+optimizer:
+    generator:
+        lr: !!float 1e-4
+        betas: [0.8,0.99]
+        eps: !!float 1e-9
+    discriminator:
+        lr: !!float 1e-4
+        betas: [0.8,0.99]
+        eps: !!float 1e-9
+scheduler:    
+    generator:
+        lr_decay: !!float 0.999875
+    discriminator:
+        lr_decay: !!float 0.999875    
+
+# loss hyper-parameters
+losses:
+   mel: !!int 45
+   kl: !!int 1
+
+
+
+
+
+
+

configs/vctk_hubertsoft_uttdvec_ppgvcf0_vits_none.yaml

@@ -0,0 +1,100 @@
+# experiment
+dataset: vctk
+train_meta: data/vctk/train_nodev_all/metadata.csv
+dev_meta: data/vctk/dev_all/metadata.csv
+train_set: train_nodev_all
+dev_set: dev_all
+
+
+# encoder-decoder
+ling_enc: hubert_soft
+spk_enc: utt_dvec
+pros_enc: ppgvc_f0
+decoder: VITS
+mel_type: vits_spec
+
+
+# training 
+fp16_run: !!bool False
+epochs: 200
+save_freq: 1 # save ckpt frequency
+show_freq: 100 # show training information frequency
+load_only_params: !!bool False
+seed: !!int 1234
+trainer: VITSTrainer
+ngpu: 2
+
+#dataloader
+dataset_class: VITSDataset
+sampling_rate: !!int 24000
+vits_hop_size: !!int 240
+spec_max_len: !!int 240
+sort: !!bool True
+dump_dir: dump
+num_workers: !!int 4
+batch_size: !!int 12
+drop_last: !!bool True
+rm_long_utt: !!bool False # remove too long utterances from metadata
+max_utt_duration: !!float 10.0 # max utterance duration (seconds)
+
+
+# decoder params
+decoder_params: 
+    spk_emb_dim: 256
+    prosodic_rep_type: continuous
+    prosodic_net:
+        hidden_dim: 192
+    input_dim: !!int 256
+    spec_channels: !!int 513
+    inter_channels: !!int 192
+    hidden_channels: !!int 192
+    filter_channels: !!int 768
+    n_heads: !!int 2
+    n_layers: !!int 6
+    kernel_size: !!int 3
+    p_dropout: !!float 0.1
+    resblock : 1
+    resblock_kernel_sizes: [3,7,11]
+    resblock_dilation_sizes: [[1,3,5], [1,3,5], [1,3,5]]
+    upsample_rates: [10,6,2,2]
+    upsample_initial_channel: !!int 512
+    upsample_kernel_sizes: [20, 12, 4, 4]
+    n_layers_q: !!int 3
+    use_spectral_norm: !!bool False
+    filter_length: !!int 1024
+    n_mels_channels: !!int 80
+    win_length: !!int 1024
+    hop_length: !!int 240
+    sampling_rate: !!int 24000
+    segment_size: !!int 9600
+
+
+
+
+#optimizer & scheduler
+optimizer:
+    generator:
+        lr: !!float 1e-4
+        betas: [0.8,0.99]
+        eps: !!float 1e-9
+    discriminator:
+        lr: !!float 1e-4
+        betas: [0.8,0.99]
+        eps: !!float 1e-9
+scheduler:    
+    generator:
+        lr_decay: !!float 0.999875
+    discriminator:
+        lr_decay: !!float 0.999875    
+
+# loss hyper-parameters
+losses:
+   mel: !!int 45
+   kl: !!int 1
+
+
+
+
+
+
+

submit_train.sh

@@ -6,9 +6,9 @@ conda_env=torch_1.9
 
 #choose config
 dataset=vctk
-ling=vqwav2vec
+#ling=vqwav2vec
 #ling=conformerppg
-#ling=hubertsoft
+ling=hubertsoft
 spk=uttdvec
 pros=ppgvcf0
 #dec=fs2
@@ -19,7 +19,7 @@ dec=vits
 #vocoder=ppgvchifigan
 vocoder=none
 
-exp_name=vctk_no16fp_split
+exp_name=vctk_first_train
 config=configs/${dataset}_${ling}_${spk}_${pros}_${dec}_${vocoder}.yaml
 if [ ! -e $config ] ; then
     echo "can't find config file $config"