From d2b5db9f2c00f3328cb2a6bff48b9c79f9c6f98e Mon Sep 17 00:00:00 2001
From: csh <458761603@qq.com>
Date: Thu, 29 May 2025 02:42:22 +0800
Subject: [PATCH] =?UTF-8?q?=E6=94=B9=E5=90=8D=20export=5Ftorch=5Fscript=5F?=
 =?UTF-8?q?v3.py=20=E4=B8=BA=20export=5Ftorch=5Fscript=5Fv3v4.py?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 ...{export_torch_script_v3.py => export_torch_script_v3v4.py} | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)
 rename GPT_SoVITS/{export_torch_script_v3.py => export_torch_script_v3v4.py} (99%)

diff --git a/GPT_SoVITS/export_torch_script_v3.py b/GPT_SoVITS/export_torch_script_v3v4.py
similarity index 99%
rename from GPT_SoVITS/export_torch_script_v3.py
rename to GPT_SoVITS/export_torch_script_v3v4.py
index dd8464be..0a9c2a19 100644
--- a/GPT_SoVITS/export_torch_script_v3.py
+++ b/GPT_SoVITS/export_torch_script_v3v4.py
@@ -473,7 +473,7 @@ class GPTSoVITSV4(torch.nn.Module):
             # 因为导出的模型在不同shape时会重新编译还是怎么的，会卡顿10s这样，
             # 所以在这里补0让他shape维持不变
             # 但是这样会导致生成的音频长度不对，所以在最后截取一下。
-            # 经过 bigvgan 之后音频长度就是 fea_todo.shape[2] * 256
+            # 经过 hifigan 之后音频长度就是 fea_todo.shape[2] * 480
             complete_len = chunk_len - fea_todo_chunk.shape[-1]
             if complete_len != 0:
                 fea_todo_chunk = torch.cat(
@@ -1090,7 +1090,7 @@ def test_export1(
 
     bert1 = bert1.T.to(device)
     bert2 = bert2.T.to(device)
-    top_k = torch.LongTensor([15]).to(device)
+    top_k = torch.LongTensor([20]).to(device)
 
     current_time = datetime.now().strftime("%Y-%m-%d %H:%M:%S")
     logger.info("start inference %s", current_time)