{"id":185,"date":"2025-06-27T14:05:38","date_gmt":"2025-06-27T06:05:38","guid":{"rendered":"https:\/\/mianka.xyz\/?p=185"},"modified":"2025-06-27T14:10:16","modified_gmt":"2025-06-27T06:10:16","slug":"pytorch","status":"publish","type":"post","link":"https:\/\/www.mianka.xyz\/?p=185","title":{"rendered":"\u6c49\u5b57\uff0c\u56fe\u7247\uff0c\u97f3\u9891\uff0c\u89c6\u9891\u600e\u4e48\u8f6c\u5316\u6210pytorch\u4e2d\u7684\u5f20\u91cf"},"content":{"rendered":"

\u4e00\u3001\u6c49\u5b57\uff08\u6587\u672c\uff09\u8f6c\u6362\u4e3a\u5f20\u91cf<\/h3>\n

\u6587\u672c\u6570\u636e\u9700\u8981\u7ecf\u8fc7\u5206\u8bcd\u3001\u7f16\u7801\u548c\u586b\u5145\u7b49\u6b65\u9aa4\u624d\u80fd\u8f6c\u6362\u4e3a\u5f20\u91cf\u3002<\/p>\n

1. \u5206\u8bcd\u4e0e\u7f16\u7801<\/span><\/h4>\n

\u4f7f\u7528\u5206\u8bcd\u5de5\u5177\uff08\u5982jieba\u3001spaCy\uff09\u5c06\u53e5\u5b50\u62c6\u5206\u4e3a\u5355\u8bcd\u6216\u5b50\u8bcd\uff0c\u518d\u901a\u8fc7\u8bcd\u8868\u6620\u5c04\u4e3a\u6570\u5b57ID\u3002<\/p>\n

import jieba\nfrom torchtext.vocab import build_vocab_from_iterator\nimport torch\n\n# \u793a\u4f8b\u6587\u672c\ntexts = ["\u8fd9\u662f\u4e00\u4e2a\u793a\u4f8b\u53e5\u5b50", "\u8fd9\u662f\u53e6\u4e00\u4e2a\u4f8b\u5b50"]\n\n# \u5206\u8bcd\ntokenized_texts = [list(jieba.cut(text)) for text in texts]\n# \u7ed3\u679c\uff1a[['\u8fd9\u662f', '\u4e00\u4e2a', '\u793a\u4f8b', '\u53e5\u5b50'], ['\u8fd9\u662f', '\u53e6\u4e00\u4e2a', '\u4f8b\u5b50']]\n\n# \u6784\u5efa\u8bcd\u8868\nvocab = build_vocab_from_iterator(tokenized_texts, specials=["<unk>"])\nvocab.set_default_index(vocab["<unk>"])  # \u672a\u77e5\u8bcd\u5904\u7406\n\n# \u6587\u672c\u8f6cID\u5e8f\u5217\ntext_ids = [vocab(tokens) for tokens in tokenized_texts]\n# \u7ed3\u679c\uff1a[[1, 2, 3, 4], [1, 5, 6]]<\/pre>\n
2. \u586b\u5145\u5e8f\u5217<\/span><\/h4>\n
\u7edf\u4e00\u5e8f\u5217\u957f\u5ea6\u4ee5\u4fbf\u6279\u91cf\u5904\u7406\u3002<\/p>\n
from torch.nn.utils.rnn import pad_sequence\n\n# \u8f6c\u6362\u4e3a\u5f20\u91cf\u5217\u8868\ntensor_list = [torch.tensor(ids) for ids in text_ids]\n\n# \u586b\u5145\u5e8f\u5217\uff08\u957f\u5ea6\u8bbe\u4e3a\u6700\u957f\u5e8f\u5217\uff09\npadded_tensors = pad_sequence(tensor_list, batch_first=True, padding_value=0)\n# \u7ed3\u679c\uff1atensor([[1, 2, 3, 4], [1, 5, 6, 0]])<\/pre>\n
\n
\u4e8c\u3001\u56fe\u7247\u8f6c\u6362\u4e3a\u5f20\u91cf<\/h3>\n
\u4f7f\u7528PyTorch\u5185\u7f6e\u7684torchvision.transforms<\/code>\u5904\u7406\u56fe\u50cf\u6570\u636e\u3002<\/p>\n
    \n
  1. \n
     \u52a0\u8f7d\u5e76\u8f6c\u6362\u56fe\u50cf<\/span><\/h4>\n<\/li>\n<\/ol>\n
    from PIL import Image\nimport torchvision.transforms as transforms\n\n# \u52a0\u8f7d\u56fe\u50cf\nimage_path = "example.jpg"\nimage = Image.open(image_path)\n\n# \u5b9a\u4e49\u8f6c\u6362\u7ba1\u9053\ntransform = transforms.Compose([\n    transforms.Resize((256, 256)),  # \u8c03\u6574\u5927\u5c0f\n    transforms.ToTensor(),          # \u8f6c\u6362\u4e3a\u5f20\u91cf\uff08\u8303\u56f4[0,1]\uff09\n    transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])  # \u6807\u51c6\u5316\n])\n\n# \u5e94\u7528\u8f6c\u6362\nimage_tensor = transform(image)  # \u5f62\u72b6\uff1a[C, H, W]<\/pre>\n
    2. \u6279\u91cf\u5904\u7406<\/span><\/h4>\n
    \u4f7f\u7528DataLoader<\/code>\u52a0\u8f7d\u56fe\u50cf\u6570\u636e\u96c6\uff1a<\/p>\n
    from torchvision.datasets import ImageFolder\nfrom torch.utils.data import DataLoader\n\ndataset = ImageFolder(root="path\/to\/images", transform=transform)\ndataloader = DataLoader(dataset, batch_size=32, shuffle=True)\n\nfor batch in dataloader:\n    images, labels = batch  # images\u5f62\u72b6\uff1a[B, C, H, W]<\/pre>\n
    <\/span><\/p>\n
    \n
    \u4e09\u3001\u97f3\u9891\u8f6c\u6362\u4e3a\u5f20\u91cf<\/h3>\n
    \u97f3\u9891\u9700\u5148\u8f6c\u6362\u4e3a\u9891\u8c31\u56fe\uff08\u5982MFCC\u3001Mel\u9891\u8c31\uff09\u518d\u8f6c\u4e3a\u5f20\u91cf\u3002<\/p>\n
      \n
    1. \n
       \u4f7f\u7528torchaudio\u5904\u7406<\/span><\/h4>\n<\/li>\n<\/ol>\n
      import torchaudio\nimport torchaudio.transforms as T\n\n# \u52a0\u8f7d\u97f3\u9891\u6587\u4ef6\nwaveform, sample_rate = torchaudio.load("example.wav")  # \u5f62\u72b6\uff1a[C, T]\n\n# \u8f6c\u6362\u4e3aMel\u9891\u8c31\nmel_spectrogram = T.MelSpectrogram(sample_rate=sample_rate)(waveform)\n# \u7ed3\u679c\u5f62\u72b6\uff1a[F, T]\uff08\u9891\u7387\u00d7\u65f6\u95f4\uff09\n\n# \u6807\u51c6\u5316\u5e76\u6dfb\u52a0\u6279\u6b21\u7ef4\u5ea6\naudio_tensor = mel_spectrogram.log2()  # \u5bf9\u6570\u7f29\u653e\naudio_tensor = audio_tensor.unsqueeze(0)  # \u5f62\u72b6\uff1a[1, F, T]<\/pre>\n
      2. \u6279\u91cf\u5904\u7406<\/span><\/h4>\n
      \u81ea\u5b9a\u4e49Dataset<\/code>\u7c7b\uff1a<\/p>\n
      from torch.utils.data import Dataset\n\nclass AudioDataset(Dataset):\n    def __init__(self, file_list):\n        self.file_list = file_list\n\n    def __getitem__(self, idx):\n        waveform, _ = torchaudio.load(self.file_list[idx])\n        mel = T.MelSpectrogram()(waveform).log2()\n        return mel.squeeze(0)  # \u5f62\u72b6\uff1a[F, T]\n\n    def __len__(self):\n        return len(self.file_list)<\/pre>\n
      \n
      \u56db\u3001\u89c6\u9891\u8f6c\u6362\u4e3a\u5f20\u91cf<\/h3>\n
      \u89c6\u9891\u5904\u7406\u9700\u9010\u5e27\u63d0\u53d6\u5e76\u8f6c\u6362\u4e3a\u5f20\u91cf\u5e8f\u5217\u3002<\/p>\n
      1. \u9010\u5e27\u5904\u7406<\/span><\/h4>\n
      \u4f7f\u7528OpenCV\u63d0\u53d6\u89c6\u9891\u5e27\uff1a<\/p>\n
      import cv2\nimport numpy as np\n\ndef video_to_tensor(video_path, frame_interval=1):\n    cap = cv2.VideoCapture(video_path)\n    frames = []\n    \n    while cap.isOpened():\n        ret, frame = cap.read()\n        if not ret:\n            break\n        if len(frames) % frame_interval == 0:\n            # \u8f6c\u6362\u4e3aRGB\u5e76\u5f52\u4e00\u5316\n            frame = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)\n            frame = frame.astype(np.float32) \/ 255.0\n            frames.append(torch.from_numpy(frame).permute(2, 0, 1))  # \u5f62\u72b6\uff1a[C, H, W]\n    \n    cap.release()\n    return torch.stack(frames)  # \u5f62\u72b6\uff1a[T, C, H, W]\n\nvideo_tensor = video_to_tensor("example.mp4")<\/pre>\n
      2. \u4f7f\u7528\u9884\u8bad\u7ec3\u6a21\u578b\u5904\u7406<\/span><\/h4>\n
      \u4f7f\u7528torchvision.models.video<\/code>\u5904\u7406\u89c6\u9891\uff1a<\/p>\n
      from torchvision.models.video import r2plus1d_18\n\nmodel = r2plus1d_18(pretrained=True)\n# \u8f93\u5165\u5f62\u72b6\uff1a[B, C, T, H, W]\n# \u9700\u5148\u5bf9\u89c6\u9891\u8fdb\u884c\u9884\u5904\u7406\uff08\u8c03\u6574\u5927\u5c0f\u3001\u5f52\u4e00\u5316\u7b49\uff09<\/pre>\n
      \n
      \u4e94\u3001\u5173\u952e\u6ce8\u610f\u4e8b\u9879<\/h3>\n
        \n
      1. \n
        \u6570\u636e\u6807\u51c6\u5316<\/span>\uff1a\u56fe\u50cf\u548c\u97f3\u9891\u901a\u5e38\u9700\u8981\u6807\u51c6\u5316\uff08\u5982\u51cf\u5747\u503c\u3001\u9664\u65b9\u5dee\uff09\u3002<\/p>\n<\/li>\n
      2. \n
        \u6279\u6b21\u5904\u7406<\/span>\uff1a\u4f7f\u7528DataLoader<\/code>\u5b9e\u73b0\u6279\u91cf\u52a0\u8f7d\u548c\u5e76\u884c\u5904\u7406\u3002<\/p>\n<\/li>\n
      3. \n
        \u5185\u5b58\u7ba1\u7406<\/span>\uff1a\u89c6\u9891\u548c\u97f3\u9891\u6570\u636e\u53ef\u80fd\u5360\u7528\u5927\u91cf\u5185\u5b58\uff0c\u9700\u5408\u7406\u8bbe\u7f6e\u6279\u6b21\u5927\u5c0f\u6216\u4f7f\u7528\u6d41\u5f0f\u5904\u7406\u3002<\/p>\n<\/li>\n
      4. \n
        \u591a\u6a21\u6001\u878d\u5408<\/span>\uff1a\u82e5\u9700\u878d\u5408\u4e0d\u540c\u6570\u636e\u7c7b\u578b\uff08\u5982\u6587\u672c+\u56fe\u50cf\uff09\uff0c\u9700\u8bbe\u8ba1\u5bf9\u9f50\u673a\u5236\uff08\u5982\u65f6\u95f4\u6233\u5bf9\u9f50\u6216\u6ce8\u610f\u529b\u673a\u5236\uff09\u3002<\/p>\n<\/li>\n<\/ol>\n
        <\/p>\n","protected":false},"excerpt":{"rendered":"
        <\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[3],"tags":[],"class_list":["post-185","post","type-post","status-publish","format-standard","hentry","category-pythonbiji"],"_links":{"self":[{"href":"https:\/\/www.mianka.xyz\/index.php?rest_route=\/wp\/v2\/posts\/185","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.mianka.xyz\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.mianka.xyz\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.mianka.xyz\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.mianka.xyz\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=185"}],"version-history":[{"count":0,"href":"https:\/\/www.mianka.xyz\/index.php?rest_route=\/wp\/v2\/posts\/185\/revisions"}],"wp:attachment":[{"href":"https:\/\/www.mianka.xyz\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=185"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.mianka.xyz\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=185"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.mianka.xyz\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=185"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}