qqq

Docker/Dockerfile

@@ -1,12 +0,0 @@
-FROM python:3.9-bullseye
-VOLUME ["/app"]
-WORKDIR /app
-# Set apt to Chinese mirror
-RUN sed -i 's/deb.debian.org/mirrors.ustc.edu.cn/g' /etc/apt/sources.list
-RUN apt-get update && apt-get -y install cmake git
-RUN git clone https://huggingface.co/spaces/ikechan8370/vits-uma-genshin-honkai
-WORKDIR /app/vits-uma-genshin-honkai
-RUN sed -i "s/\.launch()/\.launch(server_name=\"0.0.0.0\")/" /app/vits-uma-genshin-honkai/app.py
-ADD vits.sh /app/vits.sh
-EXPOSE 7860
-ENTRYPOINT [ "/app/vits.sh" ]

Docker/vits.sh

@@ -1,20 +0,0 @@
-#!/bin/bash
-run() {
-    echo -e "\033[32m已完成初始化，启动服务...\033[0m"
-    python3 /app/vits-uma-genshin-honkai/app.py
-}
-install() {
-    echo -e "\033[33m正在初始化：安装依赖....\033[0m"
-    pip install -r /app/vits-uma-genshin-honkai/requirements.txt -i https://mirrors.ustc.edu.cn/pypi/web/simple
-    echo -e "\033[33m正在下载模型....\033[0m"
-    rm -f /app/vits-uma-genshin-honkai/model/G_953000.pth
-    wget -O /app/vits-uma-genshin-honkai/model/G_953000.pth https://huggingface.co/spaces/ikechan8370/vits-uma-genshin-honkai/resolve/main/model/G_953000.pth
-    echo -e "\033[32m初始化完成！\033[0m"
-    run
-}
-
-if [ ! -f "/app/vits-uma-genshin-honkai/model/G_953000.pth" ] || [ "$(stat -c%s "/app/vits-uma-genshin-honkai/model/G_953000.pth")" -lt 10000 ]; then
-    install
-else
-    run
-fi

app.py

@@ -1,124 +1,123 @@
-import time
-import gradio as gr
-import utils
-import commons
-from models import SynthesizerTrn
-from text import text_to_sequence
-from torch import no_grad, LongTensor
-import torch
-
-hps_ms = utils.get_hparams_from_file(r'./model/config.json')
-device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
-net_g_ms = SynthesizerTrn(
-    len(hps_ms.symbols),
-    hps_ms.data.filter_length // 2 + 1,
-    hps_ms.train.segment_size // hps_ms.data.hop_length,
-    n_speakers=hps_ms.data.n_speakers,
-    **hps_ms.model).to(device)
-_ = net_g_ms.eval()
-speakers = hps_ms.speakers
-model, optimizer, learning_rate, epochs = utils.load_checkpoint(r'./model/G_953000.pth', net_g_ms, None)
-
-def get_text(text, hps):
-    text_norm, clean_text = text_to_sequence(text, hps.symbols, hps.data.text_cleaners)
-    if hps.data.add_blank:
-        text_norm = commons.intersperse(text_norm, 0)
-    text_norm = LongTensor(text_norm)
-    return text_norm, clean_text
-
-def vits(text, language, speaker_id, noise_scale, noise_scale_w, length_scale):
-    start = time.perf_counter()
-    if not len(text):
-        return "输入文本不能为空！", None, None
-    text = text.replace('\n', ' ').replace('\r', '').replace(" ", "")
-    if len(text) > 500:
-        return f"输入文字过长！{len(text)}>100", None, None
-    if language == 0:
-        text = f"[ZH]{text}[ZH]"
-    elif language == 1:
-        text = f"[JA]{text}[JA]"
-    else:
-        text = f"{text}"
-    stn_tst, clean_text = get_text(text, hps_ms)
-    with no_grad():
-        x_tst = stn_tst.unsqueeze(0)
-        x_tst_lengths = LongTensor([stn_tst.size(0)])
-        speaker_id = LongTensor([speaker_id])
-        audio = net_g_ms.infer(x_tst, x_tst_lengths, sid=speaker_id, noise_scale=noise_scale, noise_scale_w=noise_scale_w,
-                               length_scale=length_scale)[0][0, 0].data.cpu().float().numpy()
-
-    return "生成成功!", (22050, audio), f"生成耗时 {round(time.perf_counter()-start, 2)} s"
-
-def search_speaker(search_value):
-    for s in speakers:
-        if search_value == s:
-            return s
-    for s in speakers:
-        if search_value in s:
-            return s
-
-def change_lang(language):
-    if language == 0:
-        return 0.6, 0.668, 1.2
-    else:
-        return 0.6, 0.668, 1.1
-
-download_audio_js = """
-() =>{{
-    let root = document.querySelector("body > gradio-app");
-    if (root.shadowRoot != null)
-        root = root.shadowRoot;
-    let audio = root.querySelector("#tts-audio").querySelector("audio");
-    let text = root.querySelector("#input-text").querySelector("textarea");
-    if (audio == undefined)
-        return;
-    text = text.value;
-    if (text == undefined)
-        text = Math.floor(Math.random()*100000000);
-    audio = audio.src;
-    let oA = document.createElement("a");
-    oA.download = text.substr(0, 20)+'.wav';
-    oA.href = audio;
-    document.body.appendChild(oA);
-    oA.click();
-    oA.remove();
-}}
-"""
-
-if __name__ == '__main__':
-    with gr.Blocks() as app:
-        gr.Markdown(
-            "# <center> VITS语音在线合成demo\n"
-            "<div align='center'>主要有赛马娘，原神中文，原神日语，崩坏3的音色</div>"
-            '<div align="center"><a><font color="#dd0000">结果有随机性，语调可能很奇怪，可多次生成取最佳效果</font></a></div>'
-            '<div align="center"><a><font color="#dd0000">标点符号会影响生成的结果</font></a></div>'
-        )
-
-        with gr.Tabs():
-            with gr.TabItem("vits"):
-                with gr.Row():
-                    with gr.Column():
-                        input_text = gr.Textbox(label="Text (100 words limitation)", lines=5, value="今天晚上吃啥好呢。", elem_id=f"input-text")
-                        lang = gr.Dropdown(label="Language", choices=["中文", "日语", "中日混合（中文用[ZH][ZH]包裹起来，日文用[JA][JA]包裹起来）"],
-                                    type="index", value="中文")
-                        btn = gr.Button(value="Submit")
-                        with gr.Row():
-                            search = gr.Textbox(label="Search Speaker", lines=1)
-                            btn2 = gr.Button(value="Search")
-                        sid = gr.Dropdown(label="Speaker", choices=speakers, type="index", value=speakers[228])
-                        with gr.Row():
-                            ns = gr.Slider(label="noise_scale(控制感情变化程度)", minimum=0.1, maximum=1.0, step=0.1, value=0.6, interactive=True)
-                            nsw = gr.Slider(label="noise_scale_w(控制音素发音长度)", minimum=0.1, maximum=1.0, step=0.1, value=0.668, interactive=True)
-                            ls = gr.Slider(label="length_scale(控制整体语速)", minimum=0.1, maximum=2.0, step=0.1, value=1.2, interactive=True)
-                    with gr.Column():
-                        o1 = gr.Textbox(label="Output Message")
-                        o2 = gr.Audio(label="Output Audio", elem_id=f"tts-audio")
-                        o3 = gr.Textbox(label="Extra Info")
-                        download = gr.Button("Download Audio")
-                    btn.click(vits, inputs=[input_text, lang, sid, ns, nsw, ls], outputs=[o1, o2, o3], api_name="generate")
-                    download.click(None, [], [], _js=download_audio_js.format())
-                    btn2.click(search_speaker, inputs=[search], outputs=[sid])
-                    lang.change(change_lang, inputs=[lang], outputs=[ns, nsw, ls])
-            with gr.TabItem("可用人物一览"):
-                gr.Radio(label="Speaker", choices=speakers, interactive=False, type="index")
-    app.queue(concurrency_count=1).launch()
+# coding=utf-8
+import time
+import gradio as gr
+import utils
+import commons
+from models import SynthesizerTrn
+from text import text_to_sequence
+from torch import no_grad, LongTensor
+
+hps_ms = utils.get_hparams_from_file(r'./model/config.json')
+net_g_ms = SynthesizerTrn(
+    len(hps_ms.symbols),
+    hps_ms.data.filter_length // 2 + 1,
+    hps_ms.train.segment_size // hps_ms.data.hop_length,
+    n_speakers=hps_ms.data.n_speakers,
+    **hps_ms.model)
+_ = net_g_ms.eval()
+speakers = hps_ms.speakers
+model, optimizer, learning_rate, epochs = utils.load_checkpoint(r'./model/G_953000.pth', net_g_ms, None)
+
+def get_text(text, hps):
+    text_norm, clean_text = text_to_sequence(text, hps.symbols, hps.data.text_cleaners)
+    if hps.data.add_blank:
+        text_norm = commons.intersperse(text_norm, 0)
+    text_norm = LongTensor(text_norm)
+    return text_norm, clean_text
+
+def vits(text, language, speaker_id, noise_scale, noise_scale_w, length_scale):
+    start = time.perf_counter()
+    if not len(text):
+        return "输入文本不能为空！", None, None
+    text = text.replace('\n', ' ').replace('\r', '').replace(" ", "")
+    if len(text) > 300:
+        return f"输入文字过长！{len(text)}>100", None, None
+    if language == 0:
+        text = f"[ZH]{text}[ZH]"
+    elif language == 1:
+        text = f"[JA]{text}[JA]"
+    else:
+        text = f"{text}"
+    stn_tst, clean_text = get_text(text, hps_ms)
+    with no_grad():
+        x_tst = stn_tst.unsqueeze(0)
+        x_tst_lengths = LongTensor([stn_tst.size(0)])
+        speaker_id = LongTensor([speaker_id])
+        audio = net_g_ms.infer(x_tst, x_tst_lengths, sid=speaker_id, noise_scale=noise_scale, noise_scale_w=noise_scale_w,
+                               length_scale=length_scale)[0][0, 0].data.float().numpy()
+
+    return "生成成功!", (22050, audio), f"生成耗时 {round(time.perf_counter()-start, 2)} s"
+
+def search_speaker(search_value):
+    for s in speakers:
+        if search_value == s:
+            return s
+    for s in speakers:
+        if search_value in s:
+            return s
+
+def change_lang(language):
+    if language == 0:
+        return 0.6, 0.668, 1.2
+    else:
+        return 0.6, 0.668, 1.1
+
+download_audio_js = """
+() =>{{
+    let root = document.querySelector("body > gradio-app");
+    if (root.shadowRoot != null)
+        root = root.shadowRoot;
+    let audio = root.querySelector("#tts-audio").querySelector("audio");
+    let text = root.querySelector("#input-text").querySelector("textarea");
+    if (audio == undefined)
+        return;
+    text = text.value;
+    if (text == undefined)
+        text = Math.floor(Math.random()*100000000);
+    audio = audio.src;
+    let oA = document.createElement("a");
+    oA.download = text.substr(0, 20)+'.wav';
+    oA.href = audio;
+    document.body.appendChild(oA);
+    oA.click();
+    oA.remove();
+}}
+"""
+
+if __name__ == '__main__':
+    with gr.Blocks() as app:
+        gr.Markdown(
+            "# <center> VITS语音在线合成demo\n"
+            "<div align='center'>主要有赛马娘，原神中文，原神日语，崩坏3的音色</div>"
+            '<div align="center"><a><font color="#dd0000">结果有随机性，语调可能很奇怪，可多次生成取最佳效果</font></a></div>'
+            '<div align="center"><a><font color="#dd0000">标点符号会影响生成的结果</font></a></div>'
+        )
+
+        with gr.Tabs():
+            with gr.TabItem("vits"):
+                with gr.Row():
+                    with gr.Column():
+                        input_text = gr.Textbox(label="Text (100 words limitation)", lines=5, value="今天晚上吃啥好呢。", elem_id=f"input-text")
+                        lang = gr.Dropdown(label="Language", choices=["中文", "日语", "中日混合（中文用[ZH][ZH]包裹起来，日文用[JA][JA]包裹起来）"],
+                                    type="index", value="中文")
+                        btn = gr.Button(value="Submit")
+                        with gr.Row():
+                            search = gr.Textbox(label="Search Speaker", lines=1)
+                            btn2 = gr.Button(value="Search")
+                        sid = gr.Dropdown(label="Speaker", choices=speakers, type="index", value=speakers[228])
+                        with gr.Row():
+                            ns = gr.Slider(label="noise_scale(控制感情变化程度)", minimum=0.1, maximum=1.0, step=0.1, value=0.6, interactive=True)
+                            nsw = gr.Slider(label="noise_scale_w(控制音素发音长度)", minimum=0.1, maximum=1.0, step=0.1, value=0.668, interactive=True)
+                            ls = gr.Slider(label="length_scale(控制整体语速)", minimum=0.1, maximum=2.0, step=0.1, value=1.2, interactive=True)
+                    with gr.Column():
+                        o1 = gr.Textbox(label="Output Message")
+                        o2 = gr.Audio(label="Output Audio", elem_id=f"tts-audio")
+                        o3 = gr.Textbox(label="Extra Info")
+                        download = gr.Button("Download Audio")
+                    btn.click(vits, inputs=[input_text, lang, sid, ns, nsw, ls], outputs=[o1, o2, o3], api_name="generate")
+                    download.click(None, [], [], _js=download_audio_js.format())
+                    btn2.click(search_speaker, inputs=[search], outputs=[sid])
+                    lang.change(change_lang, inputs=[lang], outputs=[ns, nsw, ls])
+            with gr.TabItem("可用人物一览"):
+                gr.Radio(label="Speaker", choices=speakers, interactive=False, type="index")
+    app.queue(concurrency_count=1).launch()

models.py

@@ -496,10 +496,9 @@ class SynthesizerTrn(nn.Module):
     return o, l_length, attn, ids_slice, x_mask, y_mask, (z, z_p, m_p, logs_p, m_q, logs_q)
 
   def infer(self, x, x_lengths, sid=None, noise_scale=1, length_scale=1, noise_scale_w=1., max_len=None):
-    device = next(self.parameters()).device  # 获取模型所在的设备
-    x, m_p, logs_p, x_mask = self.enc_p(x.to(device), x_lengths.to(device))
+    x, m_p, logs_p, x_mask = self.enc_p(x, x_lengths)
     if self.n_speakers > 0:
-      g = self.emb_g(sid.to(device)).unsqueeze(-1) # [b, h, 1]
+      g = self.emb_g(sid).unsqueeze(-1) # [b, h, 1]
     else:
       g = None