update dpo

examples/libritts/cosyvoice/local/prepare_data.py

@@ -49,5 +49,7 @@ if __name__ == "__main__":
                         type=str)
     parser.add_argument('--des_dir',
                         type=str)
+    parser.add_argument('--ref_model',
+                        type=str)
     args = parser.parse_args()
     main()

examples/libritts/cosyvoice/local/prepare_reject_sample.py

@@ -0,0 +1,49 @@
+import argparse
+import logging
+import os
+from tqdm import tqdm
+import torch, torchaudio
+from cosyvoice.cli.cosyvoice import CosyVoice2
+from cosyvoice.utils.file_utils import load_wav
+
+
+logger = logging.getLogger()
+
+
+def main():
+    cosyvoice = CosyVoice2(args.ref_model)
+
+    utt2wav, utt2text = {}, {}
+    with open('{}/wav.scp'.format(args.src_dir)) as f:
+        for l in f:
+            l = l.split('\n')[0].split()
+            utt2wav[l[0]] = l[1]
+    with open('{}/text'.format(args.src_dir)) as f:
+        for l in f:
+            l = l.split('\n')[0].split()
+            utt2text[l[0]] = ' '.join(l[1:])
+
+    os.makedirs('{}/wav'.format(args.des_dir), exist_ok=True)
+    with open('{}/wav.scp'.format(args.des_dir), 'w') as f:
+        for utt, wav in tqdm(utt2wav.items()):
+            prompt_speech_16k = load_wav(wav, 16000)
+            if prompt_speech_16k.shape[1] >= 30 * 16000:
+                continue
+            speech_list = []
+            for i, j in enumerate(cosyvoice.inference_zero_shot(utt2text[utt], utt2text[utt], prompt_speech_16k, stream=False, text_frontend=False)):
+                speech_list.append(j['tts_speech'])
+            negative_wav = os.path.abspath('{}/wav/{}'.format(args.des_dir, os.path.basename(wav)))
+            torchaudio.save(negative_wav, torch.concat(speech_list, dim=1), cosyvoice.sample_rate, backend='soundfile')
+            f.write('{} {}\n'.format(utt, negative_wav))
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--src_dir',
+                        type=str)
+    parser.add_argument('--des_dir',
+                        type=str)
+    parser.add_argument('--ref_model',
+                        type=str)
+    args = parser.parse_args()
+    main()

examples/libritts/cosyvoice/run.sh

@@ -51,23 +51,6 @@ if [ ${stage} -le 3 ] && [ ${stop_stage} -ge 3 ]; then
   done
 fi
 
-# inference
-if [ ${stage} -le 4 ] && [ ${stop_stage} -ge 4 ]; then
-  echo "Run inference. Please make sure utt in tts_text is in prompt_data"
-  for mode in sft zero_shot; do
-    python cosyvoice/bin/inference.py --mode $mode \
-      --gpu 0 \
-      --config conf/cosyvoice.yaml \
-      --prompt_data data/test-clean/parquet/data.list \
-      --prompt_utt2data data/test-clean/parquet/utt2data.list \
-      --tts_text `pwd`/tts_text.json \
-      --llm_model $pretrained_model_dir/llm.pt \
-      --flow_model $pretrained_model_dir/flow.pt \
-      --hifigan_model $pretrained_model_dir/hift.pt \
-      --result_dir `pwd`/exp/cosyvoice/test-clean/$mode
-  done
-fi
-
 # train llm
 export CUDA_VISIBLE_DEVICES="0,1,2,3"
 num_gpus=$(echo $CUDA_VISIBLE_DEVICES | awk -F "," '{print NF}')

examples/libritts/cosyvoice2/run.sh

@@ -51,25 +51,6 @@ if [ ${stage} -le 3 ] && [ ${stop_stage} -ge 3 ]; then
   done
 fi
 
-# inference
-if [ ${stage} -le 4 ] && [ ${stop_stage} -ge 4 ]; then
-  echo "Run inference. Please make sure utt in tts_text is in prompt_data"
-  # TODO consider remove bin/inference.py, or use similar initilization method as in readme
-  for mode in sft zero_shot; do
-    python cosyvoice/bin/inference.py --mode $mode \
-      --gpu 0 \
-      --config conf/cosyvoice2.yaml \
-      --prompt_data data/test-clean/parquet/data.list \
-      --prompt_utt2data data/test-clean/parquet/utt2data.list \
-      --tts_text `pwd`/tts_text.json \
-      --qwen_pretrain_path $pretrained_model_dir/CosyVoice-BlankEN \
-      --llm_model $pretrained_model_dir/llm.pt \
-      --flow_model $pretrained_model_dir/flow.pt \
-      --hifigan_model $pretrained_model_dir/hift.pt \
-      --result_dir `pwd`/exp/cosyvoice/test-clean/$mode
-  done
-fi
-
 # train llm
 export CUDA_VISIBLE_DEVICES="0,1,2,3"
 num_gpus=$(echo $CUDA_VISIBLE_DEVICES | awk -F "," '{print NF}')
@@ -86,7 +67,7 @@ if [ ${stage} -le 5 ] && [ ${stop_stage} -ge 5 ]; then
   cat data/{train-clean-100,train-clean-360,train-other-500}/parquet/data.list > data/train.data.list
   cat data/{dev-clean,dev-other}/parquet/data.list > data/dev.data.list
   # NOTE will update llm/hift training later
-  for model in llm flow; do
+  for model in llm flow hifigan; do
     torchrun --nnodes=1 --nproc_per_node=$num_gpus \
         --rdzv_id=$job_id --rdzv_backend="c10d" --rdzv_endpoint="localhost:1234" \
       cosyvoice/bin/train.py \

examples/libritts/cosyvoice2/run_dpo.sh

@@ -0,0 +1,123 @@
+#!/bin/bash
+# Copyright 2024 Alibaba Inc. All Rights Reserved.
+. ./path.sh || exit 1;
+
+stage=-1
+stop_stage=3
+
+data_url=www.openslr.org/resources/60
+data_dir=/mnt/lyuxiang.lx/data/tts/openslr/libritts
+pretrained_model_dir=../../../pretrained_models/CosyVoice2-0.5B
+
+if [ ${stage} -le -1 ] && [ ${stop_stage} -ge -1 ]; then
+  echo "Data Download"
+  for part in dev-clean test-clean dev-other test-other train-clean-100 train-clean-360 train-other-500; do
+    local/download_and_untar.sh ${data_dir} ${data_url} ${part}
+  done
+fi
+
+if [ ${stage} -le 0 ] && [ ${stop_stage} -ge 0 ]; then
+  echo "Data preparation, prepare wav.scp/text/utt2spk/spk2utt"
+  for x in train-clean-100 train-clean-360 train-other-500 dev-clean dev-other test-clean test-other; do
+    mkdir -p data/$x
+    python local/prepare_data.py --src_dir $data_dir/LibriTTS/$x --des_dir data/$x
+  done
+fi
+
+if [ ${stage} -le 0 ] && [ ${stop_stage} -ge 0 ]; then
+  echo "Prepare negative samples using CosyVoice2-0.5B, this is also our reference model.
+    Here we use CosyVoice2-0.5B generated audio as reject sample for simplicity, you can use metric like wer/similarity."
+  for x in train-clean-100 train-clean-360 train-other-500; do
+    mkdir -p data/${x}_reject
+    python local/prepare_reject_sample.py --src_dir data/$x --des_dir data/${x}_reject --ref_model $pretrained_model_dir
+  done
+fi
+
+if [ ${stage} -le 1 ] && [ ${stop_stage} -ge 1 ]; then
+  echo "Extract campplus speaker embedding, you will get spk2embedding.pt and utt2embedding.pt in data/$x dir"
+  for x in train-clean-100 train-clean-360 train-other-500 dev-clean dev-other test-clean test-other; do
+    tools/extract_embedding.py --dir data/$x \
+      --onnx_path $pretrained_model_dir/campplus.onnx
+  done
+fi
+
+if [ ${stage} -le 2 ] && [ ${stop_stage} -ge 2 ]; then
+  echo "Extract discrete speech token, you will get utt2speech_token.pt in data/$x dir"
+  for x in train-clean-100 train-clean-360 train-other-500 train-clean-100_reject train-clean-360_reject dev-clean dev-other test-clean test-other; do
+    tools/extract_speech_token.py --dir data/$x \
+      --onnx_path $pretrained_model_dir/speech_tokenizer_v2.onnx
+  done
+fi
+
+if [ ${stage} -le 3 ] && [ ${stop_stage} -ge 3 ]; then
+  echo "Prepare required parquet format data, you should have prepared wav.scp/text/utt2spk/spk2utt/utt2embedding.pt/spk2embedding.pt/utt2speech_token.pt"
+  for x in train-clean-100 train-clean-360 train-other-500 dev-clean dev-other test-clean test-other; do
+    mkdir -p data/$x/parquet
+    tools/make_parquet_list.py --num_utts_per_parquet 1000 \
+      --num_processes 10 \
+      --dpo \
+      --src_dir data/$x \
+      --des_dir data/$x/parquet
+  done
+fi
+
+# train llm
+export CUDA_VISIBLE_DEVICES="0,1,2,3"
+num_gpus=$(echo $CUDA_VISIBLE_DEVICES | awk -F "," '{print NF}')
+job_id=1986
+dist_backend="nccl"
+num_workers=2
+prefetch=100
+train_engine=torch_ddp
+if [ ${stage} -le 5 ] && [ ${stop_stage} -ge 5 ]; then
+  echo "Run train. We only support llm traning for now. If your want to train from scratch, please use conf/cosyvoice.fromscratch.yaml"
+  if [ $train_engine == 'deepspeed' ]; then
+    echo "Notice deepspeed has its own optimizer config. Modify conf/ds_stage2.json if necessary"
+  fi
+  cat data/{train-clean-100,train-clean-360,train-other-500}/parquet/data.list > data/train.data.list
+  cat data/{dev-clean,dev-other}/parquet/data.list > data/dev.data.list
+  # NOTE only llm supports dpo
+  for model in llm; do
+    torchrun --nnodes=1 --nproc_per_node=$num_gpus \
+        --rdzv_id=$job_id --rdzv_backend="c10d" --rdzv_endpoint="localhost:1234" \
+      cosyvoice/bin/train.py \
+      --train_engine $train_engine \
+      --config conf/cosyvoice2.yaml \
+      --train_data data/train.data.list \
+      --cv_data data/dev.data.list \
+      --qwen_pretrain_path $pretrained_model_dir/CosyVoice-BlankEN \
+      --model $model \
+      --checkpoint $pretrained_model_dir/$model.pt \
+      --ref_model $pretrained_model_dir/llm.pt \
+      --model_dir `pwd`/exp/cosyvoice2/$model/$train_engine \
+      --tensorboard_dir `pwd`/tensorboard/cosyvoice2/$model/$train_engine \
+      --ddp.dist_backend $dist_backend \
+      --num_workers ${num_workers} \
+      --prefetch ${prefetch} \
+      --pin_memory \
+      --use_amp \
+      --dpo \
+      --deepspeed_config ./conf/ds_stage2.json \
+      --deepspeed.save_states model+optimizer
+  done
+fi
+
+# average model
+average_num=5
+if [ ${stage} -le 6 ] && [ ${stop_stage} -ge 6 ]; then
+  for model in llm flow hifigan; do
+    decode_checkpoint=`pwd`/exp/cosyvoice/$model/$train_engine/${model}.pt
+    echo "do model average and final checkpoint is $decode_checkpoint"
+    python cosyvoice/bin/average_model.py \
+      --dst_model $decode_checkpoint \
+      --src_path `pwd`/exp/cosyvoice/$model/$train_engine  \
+      --num ${average_num} \
+      --val_best
+  done
+fi
+
+if [ ${stage} -le 7 ] && [ ${stop_stage} -ge 7 ]; then
+  echo "Export your model for inference speedup. Remember copy your llm or flow model to model_dir"
+  python cosyvoice/bin/export_jit.py --model_dir $pretrained_model_dir
+  python cosyvoice/bin/export_onnx.py --model_dir $pretrained_model_dir
+fi

examples/magicdata-read/cosyvoice/run.sh

@@ -51,23 +51,6 @@ if [ ${stage} -le 3 ] && [ ${stop_stage} -ge 3 ]; then
   done
 fi
 
-# inference
-if [ ${stage} -le 4 ] && [ ${stop_stage} -ge 4 ]; then
-  echo "Run inference. Please make sure utt in tts_text is in prompt_data"
-  for mode in sft zero_shot; do
-    python cosyvoice/bin/inference.py --mode $mode \
-      --gpu 0 \
-      --config conf/cosyvoice.yaml \
-      --prompt_data data/test/parquet/data.list \
-      --prompt_utt2data data/test/parquet/utt2data.list \
-      --tts_text `pwd`/tts_text.json \
-      --llm_model $pretrained_model_dir/llm.pt \
-      --flow_model $pretrained_model_dir/flow.pt \
-      --hifigan_model $pretrained_model_dir/hift.pt \
-      --result_dir `pwd`/exp/cosyvoice/test/$mode
-  done
-fi
-
 # train llm
 export CUDA_VISIBLE_DEVICES="0,1,2,3"
 num_gpus=$(echo $CUDA_VISIBLE_DEVICES | awk -F "," '{print NF}')