From ae83d8070d4c8144d6f6335c0c55f3ae115c453c Mon Sep 17 00:00:00 2001 From: yaozengwei Date: Wed, 28 Feb 2024 20:37:42 +0800 Subject: [PATCH] minor updates related to the tokenizer change --- egs/ljspeech/TTS/vits/export-onnx.py | 3 +-- egs/ljspeech/TTS/vits/infer.py | 9 +++++---- egs/ljspeech/TTS/vits/test_onnx.py | 4 +++- 3 files changed, 9 insertions(+), 7 deletions(-) diff --git a/egs/ljspeech/TTS/vits/export-onnx.py b/egs/ljspeech/TTS/vits/export-onnx.py index f82f9dbe9..c607f0114 100755 --- a/egs/ljspeech/TTS/vits/export-onnx.py +++ b/egs/ljspeech/TTS/vits/export-onnx.py @@ -218,8 +218,7 @@ def main(): params.update(vars(args)) tokenizer = Tokenizer(params.tokens) - params.blank_id = tokenizer.blank_id - params.oov_id = tokenizer.oov_id + params.blank_id = tokenizer.pad_id params.vocab_size = tokenizer.vocab_size logging.info(params) diff --git a/egs/ljspeech/TTS/vits/infer.py b/egs/ljspeech/TTS/vits/infer.py index cf0d20ae2..9e7c71c6d 100755 --- a/egs/ljspeech/TTS/vits/infer.py +++ b/egs/ljspeech/TTS/vits/infer.py @@ -130,14 +130,16 @@ def infer_dataset( batch_size = len(batch["tokens"]) tokens = batch["tokens"] - tokens = tokenizer.tokens_to_token_ids(tokens) + tokens = tokenizer.tokens_to_token_ids( + tokens, intersperse_blank=True, add_sos=True, add_eos=True + ) tokens = k2.RaggedTensor(tokens) row_splits = tokens.shape.row_splits(1) tokens_lens = row_splits[1:] - row_splits[:-1] tokens = tokens.to(device) tokens_lens = tokens_lens.to(device) # tensor of shape (B, T) - tokens = tokens.pad(mode="constant", padding_value=tokenizer.blank_id) + tokens = tokens.pad(mode="constant", padding_value=tokenizer.pad_id) audio = batch["audio"] audio_lens = batch["audio_lens"].tolist() @@ -201,8 +203,7 @@ def main(): device = torch.device("cuda", 0) tokenizer = Tokenizer(params.tokens) - params.blank_id = tokenizer.blank_id - params.oov_id = tokenizer.oov_id + params.blank_id = tokenizer.pad_id params.vocab_size = tokenizer.vocab_size logging.info(f"Device: {device}") diff --git a/egs/ljspeech/TTS/vits/test_onnx.py b/egs/ljspeech/TTS/vits/test_onnx.py index fcbc1d663..4f46e8e6c 100755 --- a/egs/ljspeech/TTS/vits/test_onnx.py +++ b/egs/ljspeech/TTS/vits/test_onnx.py @@ -108,7 +108,9 @@ def main(): model = OnnxModel(args.model_filename) text = "I went there to see the land, the people and how their system works, end quote." - tokens = tokenizer.texts_to_token_ids([text]) + tokens = tokenizer.texts_to_token_ids( + [text], intersperse_blank=True, add_sos=True, add_eos=True + ) tokens = torch.tensor(tokens) # (1, T) tokens_lens = torch.tensor([tokens.shape[1]], dtype=torch.int64) # (1, T) audio = model(tokens, tokens_lens) # (1, T')