From aa07216bba05f624c9aa50c04a9c6a66bf04c8be Mon Sep 17 00:00:00 2001
From: RVC-Boss <129054828+RVC-Boss@users.noreply.github.com>
Date: Sun, 23 Feb 2025 15:09:22 +0800
Subject: [PATCH] =?UTF-8?q?=E4=BF=AE=E5=A4=8Dresume=20epoch=E6=95=B0?=
 =?UTF-8?q?=E8=AF=86=E5=88=AB=E9=94=99=EF=BC=8C=E6=AF=8F=E6=AC=A1resume?=
 =?UTF-8?q?=E9=83=BD=E8=A6=81=E9=83=BD=E8=AE=AD=E4=B8=80=E8=BD=AE=E7=9A=84?=
 =?UTF-8?q?=E9=97=AE=E9=A2=98?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

修复resume epoch数识别错，每次resume都要都训一轮的问题
---
 GPT_SoVITS/s2_train.py         | 7 +++++--
 GPT_SoVITS/s2_train_v3.py      | 5 ++++-
 GPT_SoVITS/s2_train_v3_lora.py | 5 ++++-
 3 files changed, 13 insertions(+), 4 deletions(-)

diff --git a/GPT_SoVITS/s2_train.py b/GPT_SoVITS/s2_train.py
index 5be43c9..2c7a426 100644
--- a/GPT_SoVITS/s2_train.py
+++ b/GPT_SoVITS/s2_train.py
@@ -205,6 +205,7 @@ def run(rank, n_gpus, hps):
             net_g,
             optim_g,
         )
+        epoch_str+=1
         global_step = (epoch_str - 1) * len(train_loader)
         # epoch_str = 1
         # global_step = 0
@@ -215,7 +216,7 @@ def run(rank, n_gpus, hps):
         if hps.train.pretrained_s2G != ""and hps.train.pretrained_s2G != None and os.path.exists(hps.train.pretrained_s2G):
             if rank == 0:
                 logger.info("loaded pretrained %s" % hps.train.pretrained_s2G)
-            print(
+            print("loaded pretrained %s" % hps.train.pretrained_s2G,
                 net_g.module.load_state_dict(
                     torch.load(hps.train.pretrained_s2G, map_location="cpu")["weight"],
                     strict=False,
@@ -227,7 +228,7 @@ def run(rank, n_gpus, hps):
         if hps.train.pretrained_s2D != ""and hps.train.pretrained_s2D != None and os.path.exists(hps.train.pretrained_s2D):
             if rank == 0:
                 logger.info("loaded pretrained %s" % hps.train.pretrained_s2D)
-            print(
+            print("loaded pretrained %s" % hps.train.pretrained_s2D,
                 net_d.module.load_state_dict(
                     torch.load(hps.train.pretrained_s2D, map_location="cpu")["weight"]
                 ) if torch.cuda.is_available() else net_d.load_state_dict(
@@ -251,6 +252,7 @@ def run(rank, n_gpus, hps):
     scaler = GradScaler(enabled=hps.train.fp16_run)
 
     for epoch in range(epoch_str, hps.train.epochs + 1):
+        print("start training from epoch %s"%epoch)
         if rank == 0:
             train_and_evaluate(
                 rank,
@@ -280,6 +282,7 @@ def run(rank, n_gpus, hps):
             )
         scheduler_g.step()
         scheduler_d.step()
+        print("training done")
 
 
 def train_and_evaluate(
diff --git a/GPT_SoVITS/s2_train_v3.py b/GPT_SoVITS/s2_train_v3.py
index a5f7da7..1d8ff30 100644
--- a/GPT_SoVITS/s2_train_v3.py
+++ b/GPT_SoVITS/s2_train_v3.py
@@ -178,6 +178,7 @@ def run(rank, n_gpus, hps):
             net_g,
             optim_g,
         )
+        epoch_str+=1
         global_step = (epoch_str - 1) * len(train_loader)
         # epoch_str = 1
         # global_step = 0
@@ -188,7 +189,7 @@ def run(rank, n_gpus, hps):
         if hps.train.pretrained_s2G != ""and hps.train.pretrained_s2G != None and os.path.exists(hps.train.pretrained_s2G):
             if rank == 0:
                 logger.info("loaded pretrained %s" % hps.train.pretrained_s2G)
-            print(
+            print("loaded pretrained %s" % hps.train.pretrained_s2G,
                 net_g.module.load_state_dict(
                     torch.load(hps.train.pretrained_s2G, map_location="cpu")["weight"],
                     strict=False,
@@ -225,6 +226,7 @@ def run(rank, n_gpus, hps):
 
     net_d=optim_d=scheduler_d=None
     for epoch in range(epoch_str, hps.train.epochs + 1):
+        print("start training from epoch %s"%epoch)
         if rank == 0:
             train_and_evaluate(
                 rank,
@@ -254,6 +256,7 @@ def run(rank, n_gpus, hps):
             )
         scheduler_g.step()
         # scheduler_d.step()
+        print("training done")
 
 
 def train_and_evaluate(
diff --git a/GPT_SoVITS/s2_train_v3_lora.py b/GPT_SoVITS/s2_train_v3_lora.py
index f10bde1..f2d5f83 100644
--- a/GPT_SoVITS/s2_train_v3_lora.py
+++ b/GPT_SoVITS/s2_train_v3_lora.py
@@ -161,6 +161,7 @@ def run(rank, n_gpus, hps):
             net_g,
             optim_g,
         )
+        epoch_str+=1
         global_step = (epoch_str - 1) * len(train_loader)
     except:  # 如果首次不能加载，加载pretrain
         # traceback.print_exc()
@@ -170,7 +171,7 @@ def run(rank, n_gpus, hps):
         if hps.train.pretrained_s2G != ""and hps.train.pretrained_s2G != None and os.path.exists(hps.train.pretrained_s2G):
             if rank == 0:
                 logger.info("loaded pretrained %s" % hps.train.pretrained_s2G)
-            print(
+            print("loaded pretrained %s" % hps.train.pretrained_s2G,
                 net_g.load_state_dict(
                     torch.load(hps.train.pretrained_s2G, map_location="cpu")["weight"],
                     strict=False,
@@ -198,6 +199,7 @@ def run(rank, n_gpus, hps):
 
     net_d=optim_d=scheduler_d=None
     for epoch in range(epoch_str, hps.train.epochs + 1):
+        print("start training from epoch %s"%epoch)
         if rank == 0:
             train_and_evaluate(
                 rank,
@@ -226,6 +228,7 @@ def run(rank, n_gpus, hps):
                 None,
             )
         scheduler_g.step()
+        print("training done")
 
 def train_and_evaluate(
     rank, epoch, hps, nets, optims, schedulers, scaler, loaders, logger, writers