{"id":1864,"date":"2025-03-07T11:51:52","date_gmt":"2025-03-07T03:51:52","guid":{"rendered":"https:\/\/www.forillusion.com\/?p=1864"},"modified":"2025-03-07T11:51:54","modified_gmt":"2025-03-07T03:51:54","slug":"6-5-rnn-pytorch","status":"publish","type":"post","link":"https:\/\/www.forillusion.com\/index.php\/6-5-rnn-pytorch\/","title":{"rendered":"6.5 \u5faa\u73af\u795e\u7ecf\u7f51\u7edc\u7684\u7b80\u6d01\u5b9e\u73b0"},"content":{"rendered":"\n<p><div class=\"has-toc have-toc\"><\/div><\/p>\n\n\n\n<p>\u5bfc\u5165\u5e93\u5e76\u52a0\u8f7d\u6570\u636e\u96c6<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>import time\nimport math\nimport numpy as np\nimport torch\nfrom torch import nn, optim\nimport torch.nn.functional as F\nimport zipfile\n\ndevice = torch.device('cuda' if torch.cuda.is_available() else 'cpu')\n\n(corpus_indices, char_to_idx, idx_to_char, vocab_size) = load_data_jay_lyrics()<\/code><\/pre>\n\n\n\n<h2 class=\"wp-block-heading\">\u5b9a\u4e49\u6a21\u578b<\/h2>\n\n\n\n<p>PyTorch\u4e2d\u7684<code>nn<\/code>\u6a21\u5757\u63d0\u4f9b\u4e86\u5faa\u73af\u795e\u7ecf\u7f51\u7edc\u7684\u5b9e\u73b0\u3002\u4e0b\u9762\u6784\u9020\u4e00\u4e2a\u542b\u5355\u9690\u85cf\u5c42\u3001\u9690\u85cf\u5355\u5143\u4e2a\u6570\u4e3a256\u7684\u5faa\u73af\u795e\u7ecf\u7f51\u7edc\u5c42<code>rnn_layer<\/code>\u3002<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>num_hiddens = 256\n# rnn_layer = nn.LSTM(input_size=vocab_size, hidden_size=num_hiddens) # \u5df2\u6d4b\u8bd5\nrnn_layer = nn.RNN(input_size=vocab_size, hidden_size=num_hiddens)<\/code><\/pre>\n\n\n\n<p>\u4e0e\u4e0a\u4e00\u8282\u4e2d\u5b9e\u73b0\u7684\u5faa\u73af\u795e\u7ecf\u7f51\u7edc\u4e0d\u540c\uff0c\u8fd9\u91cc<code>rnn_layer<\/code>\u7684\u8f93\u5165\u5f62\u72b6\u4e3a(\u65f6\u95f4\u6b65\u6570, \u6279\u91cf\u5927\u5c0f, \u8f93\u5165\u4e2a\u6570)\u3002\u5176\u4e2d\u8f93\u5165\u4e2a\u6570\u5373one-hot\u5411\u91cf\u957f\u5ea6\uff08\u8bcd\u5178\u5927\u5c0f\uff09\u3002\u6b64\u5916\uff0c<code>rnn_layer<\/code>\u4f5c\u4e3a<code>nn.RNN<\/code>\u5b9e\u4f8b\uff0c\u5728\u524d\u5411\u8ba1\u7b97\u540e\u4f1a\u5206\u522b\u8fd4\u56de\u8f93\u51fa\u548c\u9690\u85cf\u72b6\u6001h\uff0c\u5176\u4e2d\u8f93\u51fa\u6307\u7684\u662f\u9690\u85cf\u5c42\u5728<strong>\u5404\u4e2a\u65f6\u95f4\u6b65<\/strong>\u4e0a\u8ba1\u7b97\u5e76\u8f93\u51fa\u7684\u9690\u85cf\u72b6\u6001\uff0c\u5b83\u4eec\u901a\u5e38\u4f5c\u4e3a\u540e\u7eed\u8f93\u51fa\u5c42\u7684\u8f93\u5165\uff0c\u4e5f\u5c31\u662f\u8bf4\uff0c\u4e00\u6b21\u6027\u8ba1\u7b97\u51fa\u6240\u6709\u65f6\u95f4\u6b65\u7684\u9690\u85cf\u5c42\u7684\u72b6\u6001\u3002\u9700\u8981\u5f3a\u8c03\u7684\u662f\uff0c\u8be5\u201c\u8f93\u51fa\u201d\u672c\u8eab\u5e76\u4e0d\u6d89\u53ca\u8f93\u51fa\u5c42\u8ba1\u7b97\uff0c\u5f62\u72b6\u4e3a(\u65f6\u95f4\u6b65\u6570, \u6279\u91cf\u5927\u5c0f, \u9690\u85cf\u5355\u5143\u4e2a\u6570)\u3002\u800c<code>nn.RNN<\/code>\u5b9e\u4f8b\u5728\u524d\u5411\u8ba1\u7b97\u8fd4\u56de\u7684\u9690\u85cf\u72b6\u6001\u6307\u7684\u662f\u9690\u85cf\u5c42\u5728<strong>\u6700\u540e\u65f6\u95f4\u6b65<\/strong>\u7684\u9690\u85cf\u72b6\u6001\uff1a\u5f53\u9690\u85cf\u5c42\u6709\u591a\u5c42\u65f6\uff0c\u6bcf\u4e00\u5c42\u7684\u9690\u85cf\u72b6\u6001\u90fd\u4f1a\u8bb0\u5f55\u5728\u8be5\u53d8\u91cf\u4e2d\u3002\u5173\u4e8e\u5faa\u73af\u795e\u7ecf\u7f51\u7edc\uff08\u4ee5LSTM\u4e3a\u4f8b\uff09\u7684\u8f93\u51fa\uff0c\u53ef\u4ee5\u53c2\u8003\u4e0b\u56fe\u3002<\/p>\n\n\n\n<figure class=\"wp-block-image\"><img decoding=\"async\"   class=\"lazyload\" data-src=\"https:\/\/cos.forillusion.top\/wp-content\/uploads\/2025\/03\/6.5.png\" src=\"https:\/\/cdn.forillusion.com\/moezx\/img\/svg\/loader\/trans.ajax-spinner-preloader.svg\" onerror=\"imgError(this)\"  alt=\"\"\/><\/figure >\n<noscript><img decoding=\"async\" src=\"https:\/\/cos.forillusion.top\/wp-content\/uploads\/2025\/03\/6.5.png\" alt=\"\"\/><\/figure><\/noscript>\n\n\n\n<p>\u8f93\u51fa\u5f62\u72b6\u4e3a(\u65f6\u95f4\u6b65\u6570, \u6279\u91cf\u5927\u5c0f, \u9690\u85cf\u5355\u5143\u4e2a\u6570)\uff0c\u9690\u85cf\u72b6\u6001h\u7684\u5f62\u72b6\u4e3a(\u5c42\u6570, \u6279\u91cf\u5927\u5c0f, \u9690\u85cf\u5355\u5143\u4e2a\u6570)\u3002<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>num_steps = 35\nbatch_size = 2\nstate = None<\/code><\/pre>\n\n\n\n<p>\u63a5\u4e0b\u6765\u7ee7\u627f<code>Module<\/code>\u7c7b\u6765\u5b9a\u4e49\u4e00\u4e2a\u5b8c\u6574\u7684\u5faa\u73af\u795e\u7ecf\u7f51\u7edc\u3002\u5b83\u9996\u5148\u5c06\u8f93\u5165\u6570\u636e\u4f7f\u7528one-hot\u5411\u91cf\u8868\u793a\u540e\u8f93\u5165\u5230<code>rnn_layer<\/code>\u4e2d\uff0c\u7136\u540e\u4f7f\u7528\u5168\u8fde\u63a5\u8f93\u51fa\u5c42\u5f97\u5230\u8f93\u51fa\u3002\u8f93\u51fa\u4e2a\u6570\u7b49\u4e8e\u8bcd\u5178\u5927\u5c0f<code>vocab_size<\/code>\u3002<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>class RNNModel(nn.Module):\n    def __init__(self, rnn_layer, vocab_size):\n        super(RNNModel, self).__init__()\n        self.rnn = rnn_layer \n        self.hidden_size = rnn_layer.hidden_size * (2 if rnn_layer.bidirectional else 1) \n        self.vocab_size = vocab_size\n        self.dense = nn.Linear(self.hidden_size, vocab_size)\n        self.state = None\n\n    def forward(self, inputs, state): # inputs: (batch, seq_len)\n        # \u83b7\u53d6one-hot\u5411\u91cf\u8868\u793a\n        X = to_onehot(inputs, self.vocab_size) # X\u662f\u4e00\u4e2a\u957f\u5ea6\u4e3aseq_len\u7684\u5217\u8868\uff0c\u6bcf\u4e2a\u5143\u7d20\u5f62\u72b6\u4e3a(batch, vocab_size)\u7684\u77e9\u9635\n\n        Y, self.state = self.rnn(torch.stack(X), state) # stack\u51fd\u6570\u5c06X\u8fd9\u4e2a\u5217\u8868\u8f6c\u6362\u4e3a\u77e9\u9635\uff0c\u5373\u5f62\u72b6\u4e3a(seq_len, batch, vocab_size)\u7684\u77e9\u9635\n        # Y\u7684\u5f62\u72b6\u4e3a(seq_len, batch, hidden_size)\n\n        # \u5168\u8fde\u63a5\u5c42\u4f1a\u9996\u5148\u5c06Y\u7684\u5f62\u72b6\u53d8\u6210(num_steps * batch_size, num_hiddens)\uff0c\u5b83\u7684\u8f93\u51fa\n        # \u5f62\u72b6\u4e3a(num_steps * batch_size, vocab_size)\n        output = self.dense(Y.view(-1, Y.shape&#91;-1])) \n        return output, self.state<\/code><\/pre>\n\n\n\n<h2 class=\"wp-block-heading\">\u8bad\u7ec3\u6a21\u578b<\/h2>\n\n\n\n<p>\u5b9a\u4e49\u4e00\u4e2a\u9884\u6d4b\u51fd\u6570\u3002\u8fd9\u91cc\u7684\u5b9e\u73b0\u533a\u522b\u5728\u4e8e\u524d\u5411\u8ba1\u7b97\u548c\u521d\u59cb\u5316\u9690\u85cf\u72b6\u6001\u7684\u51fd\u6570\u63a5\u53e3\u3002<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>def predict_rnn_pytorch(prefix, num_chars, model, vocab_size, device, idx_to_char,\n                      char_to_idx):\n    state = None\n    output = &#91;char_to_idx&#91;prefix&#91;0]]] # output\u4f1a\u8bb0\u5f55prefix\u52a0\u4e0a\u8f93\u51fa\n    for t in range(num_chars + len(prefix) - 1):\n        X = torch.tensor(&#91;output&#91;-1]], device=device).view(1, 1)\n        if state is not None:\n            if isinstance(state, tuple): # LSTM, state:(h, c)  \n                state = (state&#91;0].to(device), state&#91;1].to(device))\n            else:   \n                state = state.to(device)\n\n        (Y, state) = model(X, state)\n        if t &lt; len(prefix) - 1:\n            output.append(char_to_idx&#91;prefix&#91;t + 1]])\n        else:\n            output.append(int(Y.argmax(dim=1).item()))\n    return ''.join(&#91;idx_to_char&#91;i] for i in output])<\/code><\/pre>\n\n\n\n<p>\u63a5\u4e0b\u6765\u5b9e\u73b0\u8bad\u7ec3\u51fd\u6570\u3002\u7b97\u6cd5\u540c\u4e0a\u4e00\u8282\u7684\u4e00\u6837\uff0c\u4f46\u8fd9\u91cc\u53ea\u4f7f\u7528\u4e86\u76f8\u90bb\u91c7\u6837\u6765\u8bfb\u53d6\u6570\u636e\u3002<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>def train_and_predict_rnn_pytorch(model, num_hiddens, vocab_size, device,\n                                corpus_indices, idx_to_char, char_to_idx,\n                                num_epochs, num_steps, lr, clipping_theta,\n                                batch_size, pred_period, pred_len, prefixes):\n    loss = nn.CrossEntropyLoss()\n    optimizer = torch.optim.Adam(model.parameters(), lr=lr)\n    model.to(device)\n    state = None\n    for epoch in range(num_epochs):\n        l_sum, n, start = 0.0, 0, time.time()\n        data_iter = data_iter_consecutive(corpus_indices, batch_size, num_steps, device) # \u76f8\u90bb\u91c7\u6837\n        for X, Y in data_iter:\n            if state is not None:\n                # \u4f7f\u7528detach\u51fd\u6570\u4ece\u8ba1\u7b97\u56fe\u5206\u79bb\u9690\u85cf\u72b6\u6001, \u8fd9\u662f\u4e3a\u4e86\n                # \u4f7f\u6a21\u578b\u53c2\u6570\u7684\u68af\u5ea6\u8ba1\u7b97\u53ea\u4f9d\u8d56\u4e00\u6b21\u8fed\u4ee3\u8bfb\u53d6\u7684\u5c0f\u6279\u91cf\u5e8f\u5217(\u9632\u6b62\u68af\u5ea6\u8ba1\u7b97\u5f00\u9500\u592a\u5927)\n                if isinstance (state, tuple): # LSTM, state:(h, c)  \n                    state = (state&#91;0].detach(), state&#91;1].detach())\n                else:   \n                    state = state.detach()\n\n            (output, state) = model(X, state) # output: \u5f62\u72b6\u4e3a(num_steps * batch_size, vocab_size)\n\n            # Y\u7684\u5f62\u72b6\u662f(batch_size, num_steps)\uff0c\u8f6c\u7f6e\u540e\u518d\u53d8\u6210\u957f\u5ea6\u4e3a\n            # batch * num_steps \u7684\u5411\u91cf\uff0c\u8fd9\u6837\u8ddf\u8f93\u51fa\u7684\u884c\u4e00\u4e00\u5bf9\u5e94\n            y = torch.transpose(Y, 0, 1).contiguous().view(-1)\n            l = loss(output, y.long())\n\n            optimizer.zero_grad()\n            l.backward()\n            # \u68af\u5ea6\u88c1\u526a\n            grad_clipping(model.parameters(), clipping_theta, device)\n            optimizer.step()\n            l_sum += l.item() * y.shape&#91;0]\n            n += y.shape&#91;0]\n\n        try:\n            perplexity = math.exp(l_sum \/ n)\n        except OverflowError:\n            perplexity = float('inf')\n        if (epoch + 1) % pred_period == 0:\n            print('epoch %d, perplexity %f, time %.2f sec' % (\n                epoch + 1, perplexity, time.time() - start))\n            for prefix in prefixes:\n                print(' -', predict_rnn_pytorch(\n                    prefix, pred_len, model, vocab_size, device, idx_to_char,\n                    char_to_idx))<\/code><\/pre>\n\n\n\n<p>\u4f7f\u7528\u548c\u4e0a\u4e00\u8282\u5b9e\u9a8c\u4e2d\u4e00\u6837\u7684\u8d85\u53c2\u6570\uff08\u9664\u4e86\u5b66\u4e60\u7387\uff09\u6765\u8bad\u7ec3\u6a21\u578b\u3002<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>num_epochs, batch_size, lr, clipping_theta = 250, 32, 1e-3, 1e-2 # \u6ce8\u610f\u8fd9\u91cc\u7684\u5b66\u4e60\u7387\u8bbe\u7f6e\npred_period, pred_len, prefixes = 50, 50, &#91;'\u5206\u5f00', '\u4e0d\u5206\u5f00']\ntrain_and_predict_rnn_pytorch(model, num_hiddens, vocab_size, device,\n                            corpus_indices, idx_to_char, char_to_idx,\n                            num_epochs, num_steps, lr, clipping_theta,\n                            batch_size, pred_period, pred_len, prefixes)<\/code><\/pre>\n\n\n\n<p>\u8f93\u51fa\uff1a<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>epoch 50, perplexity 10.340747, time 0.03 sec\n - \u5206\u5f00\u59cb\u7684\u7f8e \u8fd8\u662f\u6211 \u4e00\u4e2a\u4eba \u4e00\u4e2a \u6211\u60f3\u8981 \u4e00\u4e2a\u8d70 \u4e00\u679d\u6768\u67f3 \u4f60\u5728\u90a3\u91cc \u5728\u5c0f\u6751\u5916\u7684\u6eaa\u8fb9\u6cb3\u53e3\u9ed8\u9ed8\u9ed8\u9ed8\u9ed8\u9ed8\u9ed8\u9ed8 \n - \u4e0d\u5206\u5f00\u4e0d\u591a \u6709 \u4f60\u7684\u624b\u4e0d\u653e\u5f00 \u4e0d\u8981\u518d\u8fd9\u6837\u6253\u6211\u5988\u5988 \u60f3\u8981\u4f60\u7684\u5fae\u7b11 \u6211\u60f3\u4f60\u7684\u7231\u5199 \u6211\u60f3\u8981\u518d\u60f3\u4f60 \u6211 \u4f60\u7684\u624b\u4e0d\u653e\nepoch 100, perplexity 1.279130, time 0.03 sec\n - \u5206\u5f00 \u4e00\u76f4\u5230\u6211\u5988\u5988 \u8bf4\u4e0d\u4e86\u5f88\u4e45\u4e86\u542c\u600e\u4e48\u5c31\u662f\u6211\u60f3\u4f60\u548c\u6c49\u5821 \u6211\u60f3\u8981\u4f60\u7684\u5fae\u7b11\u6bcf\u5929\u90fd\u80fd\u770b\u5230  \u6211\u77e5\u9053\u8fd9\u91cc\u5f88\u7f8e\u4f46\u5bb6\n - \u4e0d\u5206\u5f00\u4e0d \u6211\u4e0d\u8981\u518d\u60f3\u4f60 \u4e0d \u6211\u4e0d\u591a\u592a\u591a \u6211\u60f3\u4f60\u8fd9\u6837\u7684\u542c\u4e0d\u77e5\u4e0d\u8fd9 \u6211\u60f3\u8981\u548c\u4f60\u600e\u4e48\u9762\u5bf9\u4f60 \u7231\u4f60\u5728\u6211\u5988\u4e0d\u77e5\u9053\u8fd9\nepoch 150, perplexity 1.059548, time 0.03 sec\n - \u5206\u5f00 \u4e86\u8fc7\u53bb\u79cd\u6162 \u4ece\u5c0f \u5927\u6709  \u5370\u5730\u5b89\u8001\u6591\u9e20 \u817f\u77ed\u6bdb\u4e0d\u591a \u51e0\u5929\u90fd\u6ca1\u6709\u559d\u6c34\u4e5f\u80fd\u6d3b \u8111\u888b\u74dc\u6709\u4e00\u70b9\u79c0\u9017 \u730e\u7269\u6b7b\n - \u4e0d\u5206\u5f00\u4e0d \u6211\u4e0d\u80fd\u591f\u8fdc\u8fdc\u770b\u8457 \u8fd9\u4e9b\u6211 \u505a\u5f97\u5230 \u4f46\u90a3\u4e2a\u4eba\u5df2\u7ecf\u4e0d\u662f\u6211 \u6ca1\u6709\u4f60\u5728 \u6211\u6709\u591a\u96be\u71ac  \u6ca1\u6709\u4f60\u5728\u6211\u6709\u591a\u96be\u71ac\nepoch 200, perplexity 1.031833, time 0.04 sec\n - \u5206\u5f00 \u4e86\u8fc7\u4e91\u79cd   \u4e00\u76f4\u5728\u8eab\u8fb9\u7684\u8ba9\u529b\u8fd8\u8bf4\u5206\u8fb9 \u624b\u8fc7\u5fc3\u75bc\u7684\u7384\u5bb6\u5ca9 \u522b\u5929\u7684\u624b\u5728\u4f1a\u8fd0\u505a\u5c06\u5230 \u6cea\u90a3\u5979\u517b\u7684\u9ed1\u732b\u7b11\u8d77 \n - \u4e0d\u5206\u5f00\u4e0d \u6211\u4e0d\u8981\u518d\u60f3\u4f60\u770b\u7231 \u4f60 \u4e00\u8d77\u770b\u5230\u73b0  \u4e00\u76f4\u5230\u8eab\u8fb9 \u6211\u7231\u80fd\u7684\u7247\u5473\u8bed   \u7231\u60c5\u5728\u7684\u5f62\u53e3\u8bed\u8457  \u98ce\u4ec0\u4e48\u6211\nepoch 250, perplexity 1.019352, time 0.03 sec\n - \u5206\u5f00 \u4e86\u8fc7\u4e91\u5c42 \u6211\u8bd5\u8457\u52aa\u529b\u5411\u4f60\u5954\u8dd1 \u7231\u624d\u9001\u5230 \u4f60\u5374\u5df2\u5728\u522b\u4eba\u6000\u62b1 \u5c31\u662f\u5f00\u4e0d\u4e86\u53e3\u8ba9\u5979\u77e5\u9053 \u6211\u4e00\u5b9a\u4f1a\u5475\u62a4\u8457\u4f60\n - \u4e0d\u5206\u5f00\u4e0d \u6211\u4e0d\u8981\u518d\u60f3\u4f60 \u7231 \u8d70\u7684\u592a\u5feb \u50cf\u9f99\u5377\u98ce \u4e0d\u80fd\u627f\u53d7\u6211\u5df2\u65e0\u5904\u53ef\u8eb2 \u6211\u4e0d\u8981\u518d\u60f3 \u6211\u4e0d\u8981\u518d\u60f3 \u6211\u4e0d \u6211\u4e0d<\/code><\/pre>\n","protected":false},"excerpt":{"rendered":"<p>\u5bfc\u5165\u5e93\u5e76\u52a0\u8f7d\u6570\u636e\u96c6 \u5b9a\u4e49\u6a21\u578b PyTorch\u4e2d\u7684nn\u6a21\u5757\u63d0\u4f9b\u4e86\u5faa\u73af\u795e\u7ecf\u7f51\u7edc\u7684\u5b9e\u73b0\u3002\u4e0b\u9762\u6784\u9020\u4e00\u4e2a\u542b\u5355\u9690\u85cf\u5c42\u3001\u9690\u85cf\u5355\u5143\u4e2a\u6570\u4e3a256\u7684\u5faa &#8230;<\/p>","protected":false},"author":1,"featured_media":1863,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[46,3],"tags":[45,44,12,22],"class_list":["post-1864","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-46","category-3","tag-45","tag-44","tag-12","tag-22"],"_links":{"self":[{"href":"https:\/\/www.forillusion.com\/index.php\/wp-json\/wp\/v2\/posts\/1864","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.forillusion.com\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.forillusion.com\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.forillusion.com\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.forillusion.com\/index.php\/wp-json\/wp\/v2\/comments?post=1864"}],"version-history":[{"count":1,"href":"https:\/\/www.forillusion.com\/index.php\/wp-json\/wp\/v2\/posts\/1864\/revisions"}],"predecessor-version":[{"id":1865,"href":"https:\/\/www.forillusion.com\/index.php\/wp-json\/wp\/v2\/posts\/1864\/revisions\/1865"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.forillusion.com\/index.php\/wp-json\/wp\/v2\/media\/1863"}],"wp:attachment":[{"href":"https:\/\/www.forillusion.com\/index.php\/wp-json\/wp\/v2\/media?parent=1864"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.forillusion.com\/index.php\/wp-json\/wp\/v2\/categories?post=1864"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.forillusion.com\/index.php\/wp-json\/wp\/v2\/tags?post=1864"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}