Upload 3 files

app.py

@@ -1,182 +1,183 @@
-import os
-import torch
-import random
-import shutil
-import librosa
-import warnings
-import numpy as np
-import gradio as gr
-import librosa.display
-import matplotlib.pyplot as plt
-import torchvision.transforms as transforms
-from utils import get_modelist, find_wav_files
-from collections import Counter
-from model import EvalNet
-from PIL import Image
-
-
-CLASSES = ["m_bel", "f_bel", "m_folk", "f_folk"]
-
-
-def most_common_element(input_list):
-    # 使用 Counter 统计每个元素的出现次数
-    counter = Counter(input_list)
-    # 使用 most_common 方法获取出现次数最多的元素
-    most_common_element, _ = counter.most_common(1)[0]
-    return most_common_element
-
-
-def wav_to_mel(audio_path: str, width=1.6, topdb=40):
-    os.makedirs("./tmp", exist_ok=True)
-    try:
-        y, sr = librosa.load(audio_path, sr=48000)
-        non_silents = librosa.effects.split(y, top_db=topdb)
-        non_silent = np.concatenate([y[start:end] for start, end in non_silents])
-        mel_spec = librosa.feature.melspectrogram(y=non_silent, sr=sr)
-        log_mel_spec = librosa.power_to_db(mel_spec, ref=np.max)
-        dur = librosa.get_duration(y=non_silent, sr=sr)
-        total_frames = log_mel_spec.shape[1]
-        step = int(width * total_frames / dur)
-        count = int(total_frames / step)
-        begin = int(0.5 * (total_frames - count * step))
-        end = begin + step * count
-        for i in range(begin, end, step):
-            librosa.display.specshow(log_mel_spec[:, i : i + step])
-            plt.axis("off")
-            plt.savefig(
-                f"./tmp/mel_{round(dur, 2)}_{i}.jpg",
-                bbox_inches="tight",
-                pad_inches=0.0,
-            )
-            plt.close()
-
-    except Exception as e:
-        print(f"Error converting {audio_path} : {e}")
-
-
-def wav_to_cqt(audio_path: str, width=1.6, topdb=40):
-    os.makedirs("./tmp", exist_ok=True)
-    try:
-        y, sr = librosa.load(audio_path, sr=48000)
-        non_silents = librosa.effects.split(y, top_db=topdb)
-        non_silent = np.concatenate([y[start:end] for start, end in non_silents])
-        cqt_spec = librosa.cqt(y=non_silent, sr=sr)
-        log_cqt_spec = librosa.power_to_db(np.abs(cqt_spec) ** 2, ref=np.max)
-        dur = librosa.get_duration(y=non_silent, sr=sr)
-        total_frames = log_cqt_spec.shape[1]
-        step = int(width * total_frames / dur)
-        count = int(total_frames / step)
-        begin = int(0.5 * (total_frames - count * step))
-        end = begin + step * count
-        for i in range(begin, end, step):
-            librosa.display.specshow(log_cqt_spec[:, i : i + step])
-            plt.axis("off")
-            plt.savefig(
-                f"./tmp/cqt_{round(dur, 2)}_{i}.jpg",
-                bbox_inches="tight",
-                pad_inches=0.0,
-            )
-            plt.close()
-
-    except Exception as e:
-        print(f"Error converting {audio_path} : {e}")
-
-
-def wav_to_chroma(audio_path: str, width=1.6, topdb=40):
-    os.makedirs("./tmp", exist_ok=True)
-    try:
-        y, sr = librosa.load(audio_path, sr=48000)
-        non_silents = librosa.effects.split(y, top_db=topdb)
-        non_silent = np.concatenate([y[start:end] for start, end in non_silents])
-        chroma_spec = librosa.feature.chroma_stft(y=non_silent, sr=sr)
-        log_chroma_spec = librosa.power_to_db(np.abs(chroma_spec) ** 2, ref=np.max)
-        dur = librosa.get_duration(y=non_silent, sr=sr)
-        total_frames = log_chroma_spec.shape[1]
-        step = int(width * total_frames / dur)
-        count = int(total_frames / step)
-        begin = int(0.5 * (total_frames - count * step))
-        end = begin + step * count
-        for i in range(begin, end, step):
-            librosa.display.specshow(log_chroma_spec[:, i : i + step])
-            plt.axis("off")
-            plt.savefig(
-                f"./tmp/chroma_{round(dur, 2)}_{i}.jpg",
-                bbox_inches="tight",
-                pad_inches=0.0,
-            )
-            plt.close()
-
-    except Exception as e:
-        print(f"Error converting {audio_path} : {e}")
-
-
-def embed_img(img_path, input_size=224):
-    transform = transforms.Compose(
-        [
-            transforms.Resize([input_size, input_size]),
-            transforms.ToTensor(),
-            transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5)),
-        ]
-    )
-    img = Image.open(img_path).convert("RGB")
-    return transform(img).unsqueeze(0)
-
-
-def inference(wav_path: str, log_name: str, folder_path="./tmp"):
-    if os.path.exists(folder_path):
-        shutil.rmtree(folder_path)
-
-    if not wav_path:
-        wav_path = "./examples/f_bel.wav"
-
-    model = EvalNet(log_name).model
-    spec = log_name.split("_")[-1]
-    eval("wav_to_%s" % spec)(wav_path)
-    outputs = []
-    all_files = os.listdir(folder_path)
-    for file_name in all_files:
-        if file_name.lower().endswith(".jpg"):
-            file_path = os.path.join(folder_path, file_name)
-            input = embed_img(file_path)
-            output = model(input)
-            pred_id = torch.max(output.data, 1)[1]
-            outputs.append(pred_id)
-
-    max_count_item = most_common_element(outputs)
-    shutil.rmtree(folder_path)
-    return os.path.basename(wav_path), translate[CLASSES[max_count_item]]
-
-
-if __name__ == "__main__":
-    warnings.filterwarnings("ignore")
-
-    models = get_modelist()
-    translate = {
-        "m_bel": "Male bel canto",
-        "m_folk": "Male folk singing",
-        "f_bel": "Female bel canto",
-        "f_folk": "Female folk singing",
-    }
-    examples = []
-    example_wavs = find_wav_files()
-    model_num = len(models)
-    for wav in example_wavs:
-        examples.append([wav, models[random.randint(0, model_num - 1)]])
-
-    with gr.Blocks() as demo:
-        gr.Interface(
-            fn=inference,
-            inputs=[
-                gr.Audio(label="Uploading a recording", type="filepath"),
-                gr.Dropdown(choices=models, label="Select a model", value=models[0]),
-            ],
-            outputs=[
-                gr.Textbox(label="Audio filename", show_copy_button=True),
-                gr.Textbox(label="Singing style recognition", show_copy_button=True),
-            ],
-            examples=examples,
-            allow_flagging="never",
-            title="It is recommended to keep the recording length around 5s, too long will affect the recognition efficiency.",
-        )
-
-    demo.launch()
+import os
+import torch
+import random
+import shutil
+import librosa
+import warnings
+import numpy as np
+import gradio as gr
+import librosa.display
+import matplotlib.pyplot as plt
+import torchvision.transforms as transforms
+from utils import get_modelist, find_wav_files
+from collections import Counter
+from model import EvalNet
+from PIL import Image
+
+
+TRANSLATE = {
+    "m_bel": "男声美声唱法",
+    "f_bel": "女声美声唱法",
+    "m_folk": "男声民族唱法",
+    "f_folk": "女声民族唱法",
+}
+
+CLASSES = list(TRANSLATE.keys())
+
+
+def most_common_element(input_list):
+    # 使用 Counter 统计每个元素的出现次数
+    counter = Counter(input_list)
+    # 使用 most_common 方法获取出现次数最多的元素
+    most_common_element, _ = counter.most_common(1)[0]
+    return most_common_element
+
+
+def wav_to_mel(audio_path: str, width=1.6, topdb=40):
+    os.makedirs("./tmp", exist_ok=True)
+    try:
+        y, sr = librosa.load(audio_path, sr=48000)
+        non_silents = librosa.effects.split(y, top_db=topdb)
+        non_silent = np.concatenate([y[start:end] for start, end in non_silents])
+        mel_spec = librosa.feature.melspectrogram(y=non_silent, sr=sr)
+        log_mel_spec = librosa.power_to_db(mel_spec, ref=np.max)
+        dur = librosa.get_duration(y=non_silent, sr=sr)
+        total_frames = log_mel_spec.shape[1]
+        step = int(width * total_frames / dur)
+        count = int(total_frames / step)
+        begin = int(0.5 * (total_frames - count * step))
+        end = begin + step * count
+        for i in range(begin, end, step):
+            librosa.display.specshow(log_mel_spec[:, i : i + step])
+            plt.axis("off")
+            plt.savefig(
+                f"./tmp/mel_{round(dur, 2)}_{i}.jpg",
+                bbox_inches="tight",
+                pad_inches=0.0,
+            )
+            plt.close()
+
+    except Exception as e:
+        print(f"Error converting {audio_path} : {e}")
+
+
+def wav_to_cqt(audio_path: str, width=1.6, topdb=40):
+    os.makedirs("./tmp", exist_ok=True)
+    try:
+        y, sr = librosa.load(audio_path, sr=48000)
+        non_silents = librosa.effects.split(y, top_db=topdb)
+        non_silent = np.concatenate([y[start:end] for start, end in non_silents])
+        cqt_spec = librosa.cqt(y=non_silent, sr=sr)
+        log_cqt_spec = librosa.power_to_db(np.abs(cqt_spec) ** 2, ref=np.max)
+        dur = librosa.get_duration(y=non_silent, sr=sr)
+        total_frames = log_cqt_spec.shape[1]
+        step = int(width * total_frames / dur)
+        count = int(total_frames / step)
+        begin = int(0.5 * (total_frames - count * step))
+        end = begin + step * count
+        for i in range(begin, end, step):
+            librosa.display.specshow(log_cqt_spec[:, i : i + step])
+            plt.axis("off")
+            plt.savefig(
+                f"./tmp/cqt_{round(dur, 2)}_{i}.jpg",
+                bbox_inches="tight",
+                pad_inches=0.0,
+            )
+            plt.close()
+
+    except Exception as e:
+        print(f"Error converting {audio_path} : {e}")
+
+
+def wav_to_chroma(audio_path: str, width=1.6, topdb=40):
+    os.makedirs("./tmp", exist_ok=True)
+    try:
+        y, sr = librosa.load(audio_path, sr=48000)
+        non_silents = librosa.effects.split(y, top_db=topdb)
+        non_silent = np.concatenate([y[start:end] for start, end in non_silents])
+        chroma_spec = librosa.feature.chroma_stft(y=non_silent, sr=sr)
+        log_chroma_spec = librosa.power_to_db(np.abs(chroma_spec) ** 2, ref=np.max)
+        dur = librosa.get_duration(y=non_silent, sr=sr)
+        total_frames = log_chroma_spec.shape[1]
+        step = int(width * total_frames / dur)
+        count = int(total_frames / step)
+        begin = int(0.5 * (total_frames - count * step))
+        end = begin + step * count
+        for i in range(begin, end, step):
+            librosa.display.specshow(log_chroma_spec[:, i : i + step])
+            plt.axis("off")
+            plt.savefig(
+                f"./tmp/chroma_{round(dur, 2)}_{i}.jpg",
+                bbox_inches="tight",
+                pad_inches=0.0,
+            )
+            plt.close()
+
+    except Exception as e:
+        print(f"Error converting {audio_path} : {e}")
+
+
+def embed_img(img_path, input_size=224):
+    transform = transforms.Compose(
+        [
+            transforms.Resize([input_size, input_size]),
+            transforms.ToTensor(),
+            transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5)),
+        ]
+    )
+    img = Image.open(img_path).convert("RGB")
+    return transform(img).unsqueeze(0)
+
+
+def inference(wav_path: str, log_name: str, folder_path="./tmp"):
+    if os.path.exists(folder_path):
+        shutil.rmtree(folder_path)
+
+    if not wav_path:
+        wav_path = "./examples/f_bel.wav"
+
+    model = EvalNet(log_name).model
+    spec = log_name.split("_")[-1]
+    eval("wav_to_%s" % spec)(wav_path)
+    outputs = []
+    all_files = os.listdir(folder_path)
+    for file_name in all_files:
+        if file_name.lower().endswith(".jpg"):
+            file_path = os.path.join(folder_path, file_name)
+            input = embed_img(file_path)
+            output = model(input)
+            pred_id = torch.max(output.data, 1)[1]
+            outputs.append(pred_id)
+
+    max_count_item = most_common_element(outputs)
+    shutil.rmtree(folder_path)
+    return os.path.basename(wav_path), TRANSLATE[CLASSES[max_count_item]]
+
+
+if __name__ == "__main__":
+    warnings.filterwarnings("ignore")
+
+    models = get_modelist()
+    examples = []
+    example_wavs = find_wav_files()
+    model_num = len(models)
+    for wav in example_wavs:
+        examples.append([wav, models[random.randint(0, model_num - 1)]])
+
+    with gr.Blocks() as demo:
+        gr.Interface(
+            fn=inference,
+            inputs=[
+                gr.Audio(label="上传录音", type="filepath"),
+                gr.Dropdown(choices=models, label="选择模型", value=models[0]),
+            ],
+            outputs=[
+                gr.Textbox(label="音频文件名", show_copy_button=True),
+                gr.Textbox(label="唱法识别", show_copy_button=True),
+            ],
+            examples=examples,
+            allow_flagging="never",
+            title="建议录音时长保持在 5s 左右, 过长会影响识别效率",
+        )
+
+    demo.launch()

model.py

@@ -1,142 +1,144 @@
-import torch
-import torch.nn as nn
-import torchvision.models as models
-from modelscope.msdatasets import MsDataset
-from utils import MODEL_DIR
-
-
-def get_backbone(ver, backbone_list):
-    for bb in backbone_list:
-        if ver == bb["ver"]:
-            return bb
-
-    print("Backbone name not found, using default option - alexnet.")
-    return backbone_list[0]
-
-
-def model_info(m_ver):
-    backbone_list = MsDataset.load("monetjoe/cv_backbones", split="train")
-    backbone = get_backbone(m_ver, backbone_list)
-    m_type = str(backbone["type"])
-    input_size = int(backbone["input_size"])
-    return m_type, input_size
-
-
-def Classifier(cls_num: int, output_size: int, linear_output: bool):
-    q = (1.0 * output_size / cls_num) ** 0.25
-    l1 = int(q * cls_num)
-    l2 = int(q * l1)
-    l3 = int(q * l2)
-
-    if linear_output:
-        return torch.nn.Sequential(
-            nn.Dropout(),
-            nn.Linear(output_size, l3),
-            nn.ReLU(inplace=True),
-            nn.Dropout(),
-            nn.Linear(l3, l2),
-            nn.ReLU(inplace=True),
-            nn.Dropout(),
-            nn.Linear(l2, l1),
-            nn.ReLU(inplace=True),
-            nn.Linear(l1, cls_num),
-        )
-
-    else:
-        return torch.nn.Sequential(
-            nn.Dropout(),
-            nn.Conv2d(output_size, l3, kernel_size=(1, 1), stride=(1, 1)),
-            nn.ReLU(inplace=True),
-            nn.AdaptiveAvgPool2d(output_size=(1, 1)),
-            nn.Flatten(),
-            nn.Linear(l3, l2),
-            nn.ReLU(inplace=True),
-            nn.Dropout(),
-            nn.Linear(l2, l1),
-            nn.ReLU(inplace=True),
-            nn.Linear(l1, cls_num),
-        )
-
-
-class EvalNet:
-    model = None
-    m_type = "squeezenet"
-    input_size = 224
-    output_size = 512
-
-    def __init__(self, log_name: str, cls_num=4):
-        saved_model_path = f"{MODEL_DIR}/{log_name}/save.pt"
-        m_ver = "_".join(log_name.split("_")[:-1])
-        self.m_type, self.input_size = model_info(m_ver)
-
-        if not hasattr(models, m_ver):
-            print("Unsupported model.")
-            exit()
-
-        self.model = eval("models.%s()" % m_ver)
-        linear_output = self._set_outsize()
-        self._set_classifier(cls_num, linear_output)
-        checkpoint = torch.load(saved_model_path, map_location="cpu")
-        if torch.cuda.is_available():
-            checkpoint = torch.load(saved_model_path)
-
-        self.model.load_state_dict(checkpoint, False)
-        self.model.eval()
-
-    def _set_outsize(self, debug_mode=False):
-        for name, module in self.model.named_modules():
-            if (
-                str(name).__contains__("classifier")
-                or str(name).__eq__("fc")
-                or str(name).__contains__("head")
-            ):
-                if isinstance(module, torch.nn.Linear):
-                    self.output_size = module.in_features
-                    if debug_mode:
-                        print(
-                            f"{name}(Linear): {self.output_size} -> {module.out_features}"
-                        )
-                    return True
-
-                if isinstance(module, torch.nn.Conv2d):
-                    self.output_size = module.in_channels
-                    if debug_mode:
-                        print(
-                            f"{name}(Conv2d): {self.output_size} -> {module.out_channels}"
-                        )
-                    return False
-
-        return False
-
-    def _set_classifier(self, cls_num, linear_output):
-        if self.m_type == "convnext":
-            del self.model.classifier[2]
-            self.model.classifier = nn.Sequential(
-                *list(self.model.classifier)
-                + list(Classifier(cls_num, self.output_size, linear_output))
-            )
-            return
-
-        if hasattr(self.model, "classifier"):
-            self.model.classifier = Classifier(cls_num, self.output_size, linear_output)
-            return
-
-        elif hasattr(self.model, "fc"):
-            self.model.fc = Classifier(cls_num, self.output_size, linear_output)
-            return
-
-        elif hasattr(self.model, "head"):
-            self.model.head = Classifier(cls_num, self.output_size, linear_output)
-            return
-
-        self.model.heads.head = Classifier(cls_num, self.output_size, linear_output)
-
-    def forward(self, x):
-        if torch.cuda.is_available():
-            x = x.cuda()
-            self.model = self.model.cuda()
-
-        if self.m_type == "googlenet" and self.training:
-            return self.model(x)[0]
-        else:
-            return self.model(x)
+import torch
+import torch.nn as nn
+import torchvision.models as models
+from modelscope.msdatasets import MsDataset
+from utils import MODEL_DIR
+
+
+def get_backbone(ver, backbone_list):
+    for bb in backbone_list:
+        if ver == bb["ver"]:
+            return bb
+
+    print("Backbone name not found, using default option - alexnet.")
+    return backbone_list[0]
+
+
+def model_info(m_ver):
+    backbone_list = MsDataset.load(
+        "monetjoe/cv_backbones", split="train"
+    )
+    backbone = get_backbone(m_ver, backbone_list)
+    m_type = str(backbone["type"])
+    input_size = int(backbone["input_size"])
+    return m_type, input_size
+
+
+def Classifier(cls_num: int, output_size: int, linear_output: bool):
+    q = (1.0 * output_size / cls_num) ** 0.25
+    l1 = int(q * cls_num)
+    l2 = int(q * l1)
+    l3 = int(q * l2)
+
+    if linear_output:
+        return torch.nn.Sequential(
+            nn.Dropout(),
+            nn.Linear(output_size, l3),
+            nn.ReLU(inplace=True),
+            nn.Dropout(),
+            nn.Linear(l3, l2),
+            nn.ReLU(inplace=True),
+            nn.Dropout(),
+            nn.Linear(l2, l1),
+            nn.ReLU(inplace=True),
+            nn.Linear(l1, cls_num),
+        )
+
+    else:
+        return torch.nn.Sequential(
+            nn.Dropout(),
+            nn.Conv2d(output_size, l3, kernel_size=(1, 1), stride=(1, 1)),
+            nn.ReLU(inplace=True),
+            nn.AdaptiveAvgPool2d(output_size=(1, 1)),
+            nn.Flatten(),
+            nn.Linear(l3, l2),
+            nn.ReLU(inplace=True),
+            nn.Dropout(),
+            nn.Linear(l2, l1),
+            nn.ReLU(inplace=True),
+            nn.Linear(l1, cls_num),
+        )
+
+
+class EvalNet:
+    model = None
+    m_type = "squeezenet"
+    input_size = 224
+    output_size = 512
+
+    def __init__(self, log_name: str, cls_num=4):
+        saved_model_path = f"{MODEL_DIR}/{log_name}/save.pt"
+        m_ver = "_".join(log_name.split("_")[:-1])
+        self.m_type, self.input_size = model_info(m_ver)
+
+        if not hasattr(models, m_ver):
+            print("Unsupported model.")
+            exit()
+
+        self.model = eval("models.%s()" % m_ver)
+        linear_output = self._set_outsize()
+        self._set_classifier(cls_num, linear_output)
+        checkpoint = torch.load(saved_model_path, map_location="cpu")
+        if torch.cuda.is_available():
+            checkpoint = torch.load(saved_model_path)
+
+        self.model.load_state_dict(checkpoint, False)
+        self.model.eval()
+
+    def _set_outsize(self, debug_mode=False):
+        for name, module in self.model.named_modules():
+            if (
+                str(name).__contains__("classifier")
+                or str(name).__eq__("fc")
+                or str(name).__contains__("head")
+            ):
+                if isinstance(module, torch.nn.Linear):
+                    self.output_size = module.in_features
+                    if debug_mode:
+                        print(
+                            f"{name}(Linear): {self.output_size} -> {module.out_features}"
+                        )
+                    return True
+
+                if isinstance(module, torch.nn.Conv2d):
+                    self.output_size = module.in_channels
+                    if debug_mode:
+                        print(
+                            f"{name}(Conv2d): {self.output_size} -> {module.out_channels}"
+                        )
+                    return False
+
+        return False
+
+    def _set_classifier(self, cls_num, linear_output):
+        if self.m_type == "convnext":
+            del self.model.classifier[2]
+            self.model.classifier = nn.Sequential(
+                *list(self.model.classifier)
+                + list(Classifier(cls_num, self.output_size, linear_output))
+            )
+            return
+
+        if hasattr(self.model, "classifier"):
+            self.model.classifier = Classifier(cls_num, self.output_size, linear_output)
+            return
+
+        elif hasattr(self.model, "fc"):
+            self.model.fc = Classifier(cls_num, self.output_size, linear_output)
+            return
+
+        elif hasattr(self.model, "head"):
+            self.model.head = Classifier(cls_num, self.output_size, linear_output)
+            return
+
+        self.model.heads.head = Classifier(cls_num, self.output_size, linear_output)
+
+    def forward(self, x):
+        if torch.cuda.is_available():
+            x = x.cuda()
+            self.model = self.model.cuda()
+
+        if self.m_type == "googlenet" and self.training:
+            return self.model(x)[0]
+        else:
+            return self.model(x)