Update app.py

app.py

@@ -1,41 +1,57 @@
-import os, numpy as np, torch, gradio as gr, librosa
+import os
+import numpy as np
+import torch
+import gradio as gr
+import librosa
 from huggingface_hub import hf_hub_download
-from model import DCCRN  # 确保你上传了 model.py 和 utils 依赖
+from model import DCCRN  # 确保已有 model.py 与 utils/ 依赖
 
+# ===== 基本配置 =====
 DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
-SR = 16000  # 你的模型训练采样率
+SR = int(os.getenv("SAMPLE_RATE", "16000"))
 
-# 从环境变量里读取模型仓库名和权重文件名
-REPO_ID  = os.getenv("MODEL_REPO_ID", "Ada312/DCCRN")   # 你的模型仓库
-FILENAME = os.getenv("MODEL_FILENAME", "dccrn.ckpt")    # 权重文件
-TOKEN    = os.getenv("HF_TOKEN")  # 如果模型仓库是私有，就需要这个
+# 从环境变量读取模型仓库与权重文件
+REPO_ID  = os.getenv("MODEL_REPO_ID", "Ada312/DCCRN")
+FILENAME = os.getenv("MODEL_FILENAME", "dccrn.ckpt")
+TOKEN    = os.getenv("HF_TOKEN")  # 私有模型仓库才需要
 
-# 下载权重到本地缓存
+# ===== 下载并加载权重 =====
 ckpt_path = hf_hub_download(repo_id=REPO_ID, filename=FILENAME, token=TOKEN)
 
-# 初始化模型并加载权重
-net = DCCRN()
+net = DCCRN()  # 如果训练时用了自定义参数，请按实际填入
 ckpt = torch.load(ckpt_path, map_location=DEVICE)
 state = ckpt.get("state_dict", ckpt)
-state = {k.replace("model.","").replace("module.",""): v for k,v in state.items()}
+state = {k.replace("model.", "").replace("module.", ""): v for k, v in state.items()}
 net.load_state_dict(state, strict=False)
 net.to(DEVICE).eval()
 
-# 推理函数：输入 noisy audio → 输出 enhanced audio
+# ===== 推理函数 =====
 def enhance(audio_path: str):
     wav, _ = librosa.load(audio_path, sr=SR, mono=True)
-    x = torch.from_numpy(wav).float().to(DEVICE)[None, None, :]
+    x = torch.from_numpy(wav).float().to(DEVICE)
+    if x.ndim == 1:
+        x = x.unsqueeze(0)  # [1, T]
+
     with torch.no_grad():
-        y = net(x).squeeze().cpu().numpy()
+        # 许多 DCCRN 期望 [B, 1, T]，先尝试该形状；不行再退回 [B, T]
+        try:
+            y = net(x.unsqueeze(1))  # [1, 1, T]
+        except Exception:
+            y = net(x)               # [1, T]
+
+    y = y.squeeze().detach().cpu().numpy()
     return (SR, y)
 
-# Gradio 界面
+# ===== Gradio 界面 =====
 with gr.Blocks() as demo:
     gr.Markdown("## 🎧 DCCRN Speech Enhancement\n上传或录音，点击“去噪”。")
     with gr.Row():
-        inp = gr.Audio(sources=["upload","microphone"], type="filepath", label="Noisy speech")
+        inp = gr.Audio(sources=["upload", "microphone"], type="filepath", label="Noisy speech")
         out = gr.Audio(label="Enhanced speech")
-    gr.Button("去噪").click(enhance, inputs=inp, outputs=out)
+    btn = gr.Button("去噪")
+    # 新写法：把并发限制写在事件监听器上
+    btn.click(enhance, inputs=inp, outputs=out, concurrency_limit=1)
 
-demo.queue(concurrency_count=1, max_size=8)
+# 队列：保留排队上限即可（不再使用已废弃的 concurrency_count）
+demo.queue(max_size=8)
 demo.launch()