Merge 5867122df2d08eacbdb6ffc64691403fa00e54bb into 0be59c8043a12112934d474ff4cc65658d848e8f

* 修复gpt的loss计算问题

* fallback tts config

GPT_SoVITS/AR/models/t2s_model.py

@@ -356,7 +356,7 @@ class Text2SemanticDecoder(nn.Module):
         x = self.ar_text_embedding(x)
         x = x + self.bert_proj(bert_feature.transpose(1, 2))
         x = self.ar_text_position(x)
-        x_mask = make_pad_mask(x_lens)
+        x_mask = make_pad_mask_left(x_lens)
 
         y_mask = make_pad_mask(y_lens)
         y_mask_int = y_mask.type(torch.int64)
@@ -420,7 +420,7 @@ class Text2SemanticDecoder(nn.Module):
             mask=xy_attn_mask,
         )
         x_len = x_lens.max()
-        logits = self.ar_predict_layer(xy_dec[:, x_len:])
+        logits = self.ar_predict_layer(xy_dec[:, x_len-1:])
 
         ###### DPO #############
         reject_xy_pos, reject_xy_attn_mask, reject_targets = self.make_input_data(
@@ -432,7 +432,7 @@ class Text2SemanticDecoder(nn.Module):
             mask=reject_xy_attn_mask,
         )
         x_len = x_lens.max()
-        reject_logits = self.ar_predict_layer(reject_xy_dec[:, x_len:])
+        reject_logits = self.ar_predict_layer(reject_xy_dec[:, x_len-1:])
 
         # loss
         # from feiteng: 每次 duration 越多, 梯度更新也应该更多, 所以用 sum
@@ -455,7 +455,7 @@ class Text2SemanticDecoder(nn.Module):
         x = self.ar_text_embedding(x)
         x = x + self.bert_proj(bert_feature.transpose(1, 2))
         x = self.ar_text_position(x)
-        x_mask = make_pad_mask(x_lens)
+        x_mask = make_pad_mask_left(x_lens)
 
         y_mask = make_pad_mask(y_lens)
         y_mask_int = y_mask.type(torch.int64)
@@ -502,7 +502,7 @@ class Text2SemanticDecoder(nn.Module):
             (xy_pos, None),
             mask=xy_attn_mask,
         )
-        logits = self.ar_predict_layer(xy_dec[:, x_len:]).permute(0, 2, 1)
+        logits = self.ar_predict_layer(xy_dec[:, x_len-1:]).permute(0, 2, 1)
         # loss
         # from feiteng: 每次 duration 越多, 梯度更新也应该更多, 所以用 sum
         loss = F.cross_entropy(logits, targets, reduction="sum")
@@ -578,7 +578,7 @@ class Text2SemanticDecoder(nn.Module):
     def pad_y_eos(self, y, y_mask_int, eos_id):
         targets = F.pad(y, (0, 1), value=0) + eos_id * F.pad(y_mask_int, (0, 1), value=1)
         # 错位
-        return targets[:, :-1], targets[:, 1:]
+        return targets[:, :-1], targets
 
     def infer_panel_batch_infer(
         self,

GPT_SoVITS/TTS_infer_pack/TTS.py

@@ -304,10 +304,10 @@ class TTS_Config:
             configs: dict = self._load_configs(self.configs_path)
 
         assert isinstance(configs, dict)
-        version = configs.get("version", "v2").lower()
-        assert version in ["v1", "v2", "v3", "v4", "v2pro", "v2proplus"]
-        self.default_configs[version] = configs.get(version, self.default_configs[version])
-        self.configs: dict = configs.get("custom", deepcopy(self.default_configs[version]))
+        configs_ = deepcopy(self.default_configs)
+        configs_.update(configs)
+        self.configs: dict = configs_.get("custom", configs_["v2"])
+        self.default_configs = deepcopy(configs_)
 
         self.device = self.configs.get("device", torch.device("cpu"))
         if "cuda" in str(self.device) and not torch.cuda.is_available():
@@ -315,11 +315,13 @@ class TTS_Config:
             self.device = torch.device("cpu")
 
         self.is_half = self.configs.get("is_half", False)
-        # if str(self.device) == "cpu" and self.is_half:
-        #     print(f"Warning: Half precision is not supported on CPU, set is_half to False.")
-        #     self.is_half = False
+        if str(self.device) == "cpu" and self.is_half:
+            print(f"Warning: Half precision is not supported on CPU, set is_half to False.")
+            self.is_half = False
 
+        version = self.configs.get("version", None)
         self.version = version
+        assert self.version in ["v1", "v2", "v3", "v4", "v2Pro", "v2ProPlus"], "Invalid version!"
         self.t2s_weights_path = self.configs.get("t2s_weights_path", None)
         self.vits_weights_path = self.configs.get("vits_weights_path", None)
         self.bert_base_path = self.configs.get("bert_base_path", None)
@@ -576,6 +578,10 @@ class TTS:
         if self.configs.is_half and str(self.configs.device) != "cpu":
             self.vits_model = self.vits_model.half()
 
+        self.configs.save_configs()
+
+
+
     def init_t2s_weights(self, weights_path: str):
         print(f"Loading Text2Semantic weights from {weights_path}")
         self.configs.t2s_weights_path = weights_path

GPT_SoVITS/configs/tts_infer.yaml

@@ -1,4 +1,3 @@
-version: v2ProPlus
 custom:
   bert_base_path: GPT_SoVITS/pretrained_models/chinese-roberta-wwm-ext-large
   cnhuhbert_base_path: GPT_SoVITS/pretrained_models/chinese-hubert-base

GPT_SoVITS/inference_webui_fast.py

@@ -125,7 +125,8 @@ is_exist_s2gv4 = os.path.exists(path_sovits_v4)
 tts_config = TTS_Config("GPT_SoVITS/configs/tts_infer.yaml")
 tts_config.device = device
 tts_config.is_half = is_half
-tts_config.version = version
+# tts_config.version = version
+tts_config.update_version(version)
 if gpt_path is not None:
     if "！" in gpt_path or "!" in gpt_path:
         gpt_path = name2gpt_path[gpt_path]

tools/assets.py

@@ -59,7 +59,7 @@ top_html = """
         <a href="https://www.yuque.com/baicaigongchang1145haoyuangong/ib3g1e" target="_blank">
             <img src="https://img.shields.io/badge/简体中文-阅读文档-blue?style=for-the-badge&logo=googledocs&logoColor=white" style="width: auto; height: 30px;">
         </a>
-        <a href="https://github.com/RVC-Boss/GPT-SoVITS" target="_blank">
+        <a href="https://lj1995-gpt-sovits-proplus.hf.space/" target="_blank">
             <img src="https://img.shields.io/badge/免费在线体验-free_online_demo-yellow.svg?style=for-the-badge&logo=huggingface" style="width: auto; height: 30px;">
         </a>
         <a href="https://www.yuque.com/baicaigongchang1145haoyuangong/ib3g1e" target="_blank">