update evaluation

.gitignore

@@ -18,6 +18,8 @@ pretrained_models
 # parts that are not public yet
 evaluation/UTMOS-demo
 evaluation/eval_list*.txt
+evaluation/speechbrain_asr_model_weights
+results/
 
 
 

bin/feature_extraction_multi_jobs.sh

@@ -5,10 +5,10 @@ conda_env=torch_1.7
 
 # setup
 
-dataset=vctk
-config=configs/preprocess_bigvgan_mel.yaml
-feature_type=bigvgan_mel
-splits="train_nodev_all dev_all"
+dataset=libritts
+config=configs/preprocess_ppgvc_f0.yaml
+feature_type=ppgvc_f0
+splits="train_nodev_clean dev_clean"
 
 script_dir=scripts/$dataset/preprocess
 

bin/feature_extraction_sge_multi_tasks.sh

@@ -0,0 +1,38 @@
+#!/bin/bash
+
+conda=/share/mini1/sw/std/python/anaconda3-2019.07/v3.7
+conda_env=torch_1.9
+
+# setup
+
+dataset=libritts
+config=configs/preprocess_ppgvc_mel.yaml
+feature_type=ppgvc_mel
+splits="train_nodev_clean dev_clean"
+
+script_dir=scripts/$dataset/preprocess
+
+[ ! -e $script_dir ]  && mkdir -p  $script_dir 
+
+for split in $splits ; do
+
+    echo "[feature extraction]: $split $dataset $feature_type"
+    b=$script_dir/feature_extraction_${feature_type}_${split}.sh
+    l=logs/feature_extraction_${feature_type}_${split}.log
+    cat <<EOF > $b
+#!/bin/bash
+source $conda/bin/activate $conda_env
+python3 feature_extraction.py \
+    --metadata data/$dataset/$split/metadata.csv \
+    --dump_dir dump/$dataset \
+    --config_path  $config \
+    --split $split \
+    --max_workers 20 \
+    --feature_type $feature_type \
+    --sge_task_id \$SGE_TASK_ID \
+    --sge_n_tasks 5000
+EOF
+    chmod +x $b
+    submitjob -m 10000 -n 5000  $l $b
+    echo "submitjob for $dataset $split  $feature_type see log $l"
+done        

bin/generate_eval_list.sh

@@ -1,9 +1,9 @@
 #!/bin/bash
 
 task=vc
-dataset=vctk
-split=eval_all
-eval_list=eval_list_m2m_vc_small_oneshot.json
+dataset=libritts
+split=eval_clean
+eval_list=eval_list_a2a_vc_small_oneshot.json
 n_trg_spk_samples=1
 n_src_spk_samples=4
 n_eval_spks=10

configs/libritts_conformerppg_uttdvec_ppgvcf0_fs2_ppgvchifigan.yaml

@@ -0,0 +1,89 @@
+# experiment
+dataset: libritts
+train_meta: data/libritts/train_nodev_clean/metadata.csv
+dev_meta: data/libritts/dev_clean/metadata.csv
+train_set: train_nodev_clean
+dev_set: dev_clean
+
+
+# encoder-decoder
+ling_enc: conformer_ppg
+spk_enc: utt_dvec
+pros_enc: ppgvc_f0
+decoder: FastSpeech2
+mel_type: ppgvc_mel
+vocoder: ppgvc_hifigan
+
+# training 
+fp16_run: !!bool True
+epochs: 200
+save_freq: 2 # save ckpt frequency
+show_freq: 100 # show training information frequency
+load_only_params: !!bool False
+seed: !!int 1234
+trainer: FS2Trainer
+ngpu: 1
+
+#dataloader
+dataset_class: Dataset
+sort: !!bool False
+dump_dir: dump
+num_workers: !!int 8
+batch_size: 32
+drop_last: !!bool True
+rm_long_utt: !!bool True # remove too long utterances from metadata
+max_utt_duration: !!float 10.0 # max utterance duration (seconds)
+
+
+# decoder params
+decoder_params: 
+    out_dim: 80
+    max_len: 1000
+    max_seq_len: 1000
+    spk_emb_dim: 256
+    prosodic_rep_type: continuous
+    prosodic_net:
+        hidden_dim: 256
+    prenet:
+        conv_kernel_size: 3
+        input_dim: 144
+        dropout: 0.1
+    postnet:
+        idim: 80
+        odim: 80
+        n_layers: 0
+        n_filts: 5
+        n_chans: 256
+        dropout_rate: 0.5
+    transformer:
+        encoder_layer: 4
+        encoder_head: 2
+        encoder_hidden: 256
+        decoder_layer: 4
+        decoder_head: 2
+        decoder_hidden: 256
+        conv_filter_size: 1024
+        conv_kernel_size: [3, 1]
+        encoder_dropout: 0.1
+        decoder_dropout: 0.1
+
+#optimizer & scheduler
+optimizer:
+    init_lr: !!float 1e-2
+    betas: [0.9,0.99]
+    weight_decay: 0.0        
+scheduler:    
+    warm_up_step: 4000
+    anneal_steps: [800000, 900000, 1000000]
+    anneal_rate: 0.3
+
+# loss hyper-parameters
+losses:
+   alpha: 1. 
+
+
+
+
+
+
+

configs/libritts_conformerppg_uttdvec_ppgvcf0_tacoar_ppgvchifigan.yaml

@@ -0,0 +1,78 @@
+# experiment
+dataset: libritts
+train_meta: data/libritts/train_nodev_clean/metadata.csv
+dev_meta: data/libritts/dev_clean/metadata.csv
+train_set: train_nodev_clean
+dev_set: dev_clean
+
+
+# encoder-decoder
+ling_enc: conformer_ppg
+spk_enc: utt_dvec
+pros_enc: ppgvc_f0
+decoder: TacoAR
+mel_type: ppgvc_mel
+vocoder: ppgvc_hifigan
+
+
+# training 
+fp16_run: !!bool True
+epochs: 200
+save_freq: 2 # save ckpt frequency
+show_freq: 10
+load_only_params: !!bool False
+seed: !!int 1234
+trainer: TacoARTrainer
+ngpu: 2
+
+#dataloader
+dataset_class: Dataset
+sort: !!bool True
+dump_dir: dump
+num_workers: !!int 8
+batch_size: 64
+drop_last: !!bool True
+rm_long_utt: !!bool True # remove too long utterances from metadata
+max_utt_duration: !!float 10.0 # max utterance duration (seconds)
+
+
+# decoder params
+decoder_params: 
+    prosodic_rep_type: continuous
+    prosodic_net:
+        hidden_dim: 1024
+    input_dim: 144
+    output_dim: 80
+    resample_ratio: 1
+    spk_emb_integration_type: concat # add or concat
+    spk_emb_dim: 256
+    ar: True
+    encoder_type: "taco2"
+    hidden_dim: 1024
+    prenet_layers: 2  # if set 0, no prenet is used
+    prenet_dim: 256
+    prenet_dropout_rate: 0.5
+    lstmp_layers: 2
+    lstmp_dropout_rate: 0.2
+    lstmp_proj_dim: 256
+    lstmp_layernorm: False
+
+#optimizer & scheduler
+optimizer:
+    weight_decay: 0.0
+    betas: [0.9,0.99]  
+    lr: !!float 1e-4
+scheduler:    
+    num_training_steps: 500000
+    num_warmup_steps: 4000
+
+# loss hyper-parameters
+losses:
+   alpha: 1. 
+
+
+
+
+
+
+

configs/libritts_conformerppg_uttdvec_ppgvcf0_tacomol_ppgvchifigan.yaml

@@ -0,0 +1,67 @@
+# experiment
+dataset: libritts
+train_meta: data/libritts/train_nodev_clean/metadata.csv
+dev_meta: data/libritts/dev_clean/metadata.csv
+train_set: train_nodev_clean
+dev_set: dev_clean
+
+
+# encoder-decoder
+ling_enc: conformer_ppg
+spk_enc: utt_dvec
+pros_enc: ppgvc_f0
+decoder: TacoMOL
+mel_type: ppgvc_mel
+vocoder: ppgvc_hifigan
+
+
+# training 
+fp16_run: !!bool True
+epochs: 200
+save_freq: 2 # save ckpt frequency
+show_freq: 10
+load_only_params: !!bool False
+seed: !!int 1234
+trainer: TacoMOLTrainer
+ngpu: 2
+
+#dataloader
+dataset_class: Dataset
+sort: !!bool True
+dump_dir: dump
+num_workers: !!int 8
+batch_size: 64
+drop_last: !!bool True
+rm_long_utt: !!bool True # remove too long utterances from metadata
+max_utt_duration: !!float 10.0 # max utterance duration (seconds)
+frames_per_step: !!int 4
+
+
+# decoder params
+decoder_params: 
+    out_dim: 80
+    prosodic_rep_type: continuous
+    prosodic_net:
+        hidden_dim: 256
+    spk_embed_dim: 256
+    bottle_neck_feature_dim: 144
+
+#optimizer & scheduler
+optimizer:
+    weight_decay: !!float 1e-6
+    betas: [0.9,0.99]  
+    lr: !!float 1e-4
+scheduler:    
+    num_training_steps: 500000
+    num_warmup_steps: 4000
+
+# loss hyper-parameters
+losses:
+   alpha: 1. 
+
+
+
+
+
+
+