Merge branch 'PaddlePaddle:develop' into develop

paddlespeech/server/engine/tts/paddleinference/tts_engine.py

@@ -247,7 +247,9 @@ class TTSServerExecutor(TTSExecutor):
             else:
                 # multi speaker  do not have static model
                 if am_dataset in {"aishell3", "vctk"}:
-                    pass
+                    am_result = run_model(
+                        self.am_predictor,
+                        [part_phone_ids.numpy(), np.array([spk_id])])
                 else:
                     am_result = run_model(self.am_predictor,
                                           [part_phone_ids.numpy()])

paddlespeech/t2s/exps/vits/normalize.py

@@ -187,7 +187,7 @@ def main():
             record["spk_emb"] = str(item["spk_emb"])
 
         output_metadata.append(record)
-    output_metadata.sort(key=itemgetter('utt_id'))
+    output_metadata.sort(key=itemgetter('feats_lengths'))
     output_metadata_path = Path(args.dumpdir) / "metadata.jsonl"
     with jsonlines.open(output_metadata_path, 'w') as writer:
         for item in output_metadata:

paddlespeech/t2s/exps/vits/preprocess.py

@@ -166,7 +166,7 @@ def process_sentences(config,
                     if record:
                         results.append(record)
 
-    results.sort(key=itemgetter("utt_id"))
+    results.sort(key=itemgetter("feats_lengths"))
     with jsonlines.open(output_dir / "metadata.jsonl", 'w') as writer:
         for item in results:
             writer.write(item)

paddlespeech/t2s/exps/vits/train.py

@@ -24,13 +24,13 @@ import yaml
 from paddle import DataParallel
 from paddle import distributed as dist
 from paddle.io import DataLoader
-from paddle.io import DistributedBatchSampler
 from paddle.optimizer import Adam
 from yacs.config import CfgNode
 
 from paddlespeech.t2s.datasets.am_batch_fn import vits_multi_spk_batch_fn
 from paddlespeech.t2s.datasets.am_batch_fn import vits_single_spk_batch_fn
 from paddlespeech.t2s.datasets.data_table import DataTable
+from paddlespeech.t2s.datasets.sampler import ErnieSATSampler
 from paddlespeech.t2s.models.vits import VITS
 from paddlespeech.t2s.models.vits import VITSEvaluator
 from paddlespeech.t2s.models.vits import VITSUpdater
@@ -107,12 +107,12 @@ def train_sp(args, config):
         converters=converters, )
 
     # collate function and dataloader
-    train_sampler = DistributedBatchSampler(
+    train_sampler = ErnieSATSampler(
         train_dataset,
         batch_size=config.batch_size,
         shuffle=True,
         drop_last=True)
-    dev_sampler = DistributedBatchSampler(
+    dev_sampler = ErnieSATSampler(
         dev_dataset,
         batch_size=config.batch_size,
         shuffle=False,

paddlespeech/t2s/modules/transformer/attention.py

@@ -196,7 +196,7 @@ class RelPositionMultiHeadedAttention(MultiHeadedAttention):
 
         if self.zero_triu:
             ones = paddle.ones((t1, t2))
-            x = x * paddle.tril(ones, t2 - 1)[None, None, :, :]
+            x = x * paddle.tril(ones, t2 - t1)[None, None, :, :]
 
         return x
 
@@ -299,7 +299,7 @@ class LegacyRelPositionMultiHeadedAttention(MultiHeadedAttention):
 
         if self.zero_triu:
             ones = paddle.ones((t1, t2))
-            x = x * paddle.tril(ones, t2 - 1)[None, None, :, :]
+            x = x * paddle.tril(ones, t2 - t1)[None, None, :, :]
 
         return x
 

setup.py

@@ -44,8 +44,9 @@ base = [
     "loguru",
     "matplotlib",
     "nara_wpe",
-    "onnxruntime==1.11.0",
+    "onnxruntime>=1.11.0",
     "opencc",
+    "opencc-python-reimplemented",
     "pandas",
     "paddlenlp>=2.4.8",
     "paddlespeech_feat",