{"id":9241,"date":"2024-04-09T11:21:52","date_gmt":"2024-04-09T03:21:52","guid":{"rendered":"\/?p=9241"},"modified":"2024-04-09T11:21:52","modified_gmt":"2024-04-09T03:21:52","slug":"%e5%88%a9%e7%94%a8rnn%e5%be%aa%e7%8e%af%e7%a5%9e%e7%bb%8f%e7%bd%91%e7%bb%9c%e5%af%b9mnist%e6%89%8b%e5%86%99%e6%95%b0%e5%ad%97%e8%af%86%e5%88%ab","status":"publish","type":"post","link":"\/?p=9241","title":{"rendered":"\u5229\u7528RNN\u5faa\u73af\u795e\u7ecf\u7f51\u7edc\u5bf9MNIST\u624b\u5199\u6570\u5b57\u8bc6\u522b"},"content":{"rendered":"

\u6982\u5ff5<\/h2>\n

\u4e3a\u4ec0\u4e48\u9700\u8981 RNN(\u5faa\u73af\u795e\u7ecf\u7f51\u7edc)<\/p>\n

DNN\u90fd\u53ea\u80fd\u5355\u72ec\u7684\u53d6\u5904\u7406\u4e00\u4e2a\u4e2a\u7684\u8f93\u5165 \u524d\u4e00\u4e2a\u8f93\u5165\u548c\u540e\u4e00\u4e2a\u8f93\u5165\u662f\u5b8c\u5168\u6ca1\u6709\u5173\u7cfb\u7684\u3002<\/p>\n

\u4f46\u662f\uff0c\u67d0\u4e9b\u4efb\u52a1\u9700\u8981\u80fd\u591f\u66f4\u597d\u7684\u5904\u7406\u5e8f\u5217\u7684\u4fe1\u606f\uff0c\u5373\u524d\u9762\u7684\u8f93\u5165\u548c\u540e\u9762\u7684\u8f93\u5165\u662f\u6709\u5173\u7cfb\u7684\u3002<\/p>\n

\u6bd4\u5982\uff0c\u5f53\u6211\u4eec\u5728\u7406\u89e3\u4e00\u53e5\u8bdd\u610f\u601d\u65f6,\u5b64\u7acb\u7684\u7406\u89e3\u8fd9\u53e5\u8bdd\u7684\u6bcf\u4e2a\u8bcd\u662f\u4e0d\u591f\u7684,\u6211\u4eec\u9700\u8981\u5904\u7406\u8fd9\u4e9b\u8bcd\u8fde\u63a5\u8d77\u6765\u7684\u6574\u4e2a\u5e8f\u5217; <\/p>\n

\u5f53\u6211\u4eec\u5904\u7406\u89c6\u9891\u7684\u65f6\u5019\uff0c\u6211\u4eec\u4e5f\u4e0d\u80fd\u53ea\u5355\u72ec\u7684\u53bb\u5206\u6790\u6bcf\u4e00\u5e27\uff0c\u800c\u8981\u5206\u6790\u8fd9\u4e9b\u5e27\u8fde\u63a5\u8d77\u6765\u7684\u6574\u4e2a\u5e8f\u5217\u3002<\/p>\n

\u4ee5 np \u7684\u4e00\u4e2a\u6700\u7b80\u5355\u8bcd\u6027\u6807\u6ce8\u4efb\u52a1\u6765\u8bf4,\u5c06 \u6211 \u5403 \u82f9\u679c \u4e09\u4e2a\u5355\u8bcd\u6807\u6ce8\u8bcd\u6027\u4e3a \u6211\/nn \u5403\/v \u82f9\u679c\/nn\u3002<\/p>\n

\"image-20240409111619833\" <\/p>\n

\"image-20240409111641444\"<\/p>\n

\"image-20240409111650823\"<\/p>\n

\"ebf5a7264034578251f169acf05ff17\"<\/p>\n

RNN \u624b\u5199\u6570\u5b57\u8bc6\u522b<\/h2>\n
import tensorflow as tf\nfrom tensorflow.examples.tutorials.mnist import input_data\nimport matplotlib.pyplot as plt\n\n# \u8bfb\u53d6mnist\u6570\u636e\u96c6\uff0cone_hot=True\u5c06y\u5217\u7f16\u7801\u4e3a\u7ef4\u5ea610\u5206\u7c7b\u76840\uff0c1\u7f16\u7801\nmnist = input_data.read_data_sets('MNIST_data_bak', one_hot=True)\n# \u6253\u5370\u8f93\u51fa\u8bad\u7ec3\u96c6\u7684\u5f62\u72b6(55000, 784)\nprint(mnist.train.images.shape)\n\n# \u8d85\u53c2\u6570\nlr = 0.001\ntraining_iters = 1000000\nbatch_size = 128\nn_inputs = 28\nn_steps = 28\nn_hidden_units = 128\nn_classes = 10\n\n# \u56fe\u8f93\u5165\nx = tf.placeholder(tf.float32, [None, n_steps, n_inputs])\ny = tf.placeholder(tf.float32, [None, n_classes])\n\n# \u5b9a\u4e49\u6743\u91cd\nweights = {\n    'out': tf.Variable(tf.random_normal([n_hidden_units, n_classes]))\n}\nbiases = {\n    'out': tf.Variable(tf.constant(0.1, shape=[n_classes, ]))\n}\n\ndef RNN(X, weights, biases):\n    # \u5355\u5143\n    # forget_bias = 1.0 \u662f\u521d\u59cb\u503c\n    # lstm_cell = tf.nn.rnn_cell.BasicLSTMCell(n_hidden_units, forget_bias=1.0, state_is_tuple=True)\n    rnn_cell = tf.nn.rnn_cell.BasicRNNCell(n_hidden_units)\n    # lstm cell \u5206\u4e3a\u4e24\u90e8\u5206 (c_state, m_state), RNN\u4f1a\u8ba1\u7b97\u6bcf\u4e00\u4e2acell\u91cc\u9762\u7684\u7ed3\u679c\n    # _init_state = lstm_cell.zero_state(batch_size, dtype=tf.float32)\n    _init_state = rnn_cell.zero_state(batch_size, dtype=tf.float32)\n    # states\u91cc\u9762(c_state, m_state)\uff0c\u5982\u679c\u53ea\u662f\u4e00\u822c\u7684RNN\u7684\u8bdd\uff0c\u5c31\u53ea\u662fm_state\n    # states\u662f\u6700\u540e\u4e00\u4e2astate\uff0coutputs\u662f\u4e2alist\uff0c\u6bcf\u4e00\u6b65\u7684output\u90fd\u5b58\u5728\u91cc\u9762\n    # \u6709rnn\u548cdynamic_rnn\uff0c\u533a\u522b\u662f\u6bcf\u6279\u6b21\u7684\u7ef4\u5ea6\u53ef\u4ee5\u4e0d\u4e00\u6837\uff0crnn\u5fc5\u987b\u4e00\u6837\n    # 28 steps\u5c31\u662f\u6211\u4eec\u7684time\u8f74\uff0ctime_major\u662f\u4e0d\u662f\u7b2c\u4e00\u4e2a\u7ef4\u5ea6\uff0c\u6211\u4eec\u7684\u662f\u5728\u7ef4\u5ea6\u4e3a2\u7684\u5730\u65b9\uff0c\u6240\u6709False\n    # outputs, last_states = tf.nn.dynamic_rnn(lstm_cell, X, initial_state=_init_state, time_major=False)\n    outputs, last_state = tf.nn.dynamic_rnn(rnn_cell, X, initial_state=_init_state, time_major=False)\n    # \u5982\u679c\u662fTrue\uff0coutputs\u7684\u7ef4\u5ea6\u662f[steps, batch_size, depth]\uff0c\u53cd\u4e4b\u5c31\u662f[batch_size, steps, depth]\u3002\u5c31\u662f\u548c\u8f93\u5165\u662f\u4e00\u6837\u7684\n    # last_state\u5c31\u662f\u6574\u4e2aLSTM\u8f93\u51fa\u7684\u6700\u7ec8\u7684\u72b6\u6001\uff0c\u5305\u542bc\u548ch\u3002c\u548ch\u7684\u7ef4\u5ea6\u90fd\u662f[batch_size\uff0c n_hidden]\n\n    # \u9690\u85cf\u5c42\u5230\u8f93\u5165\u7ed3\u679c\n    # results = tf.matmul(last_states[1], weights['out']) + biases['out']\n    results = tf.matmul(last_state, weights['out']) + biases['out']\n\n    return results\n\npred = RNN(x, weights, biases)\ncost = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=pred, labels=y))\ntrain_op = tf.train.AdamOptimizer(lr).minimize(cost)\n\ncorrect_pred = tf.equal(tf.argmax(pred, 1), tf.argmax(y, 1))\naccuracy = tf.reduce_mean(tf.cast(correct_pred, tf.float32))\n\ninit = tf.global_variables_initializer()\nwith tf.Session() as sess:\n    sess.run(init)\n    step = 0\n    while step * batch_size < training_iters:\n        batch_xs, batch_ys = mnist.train.next_batch(batch_size)\n        batch_xs = batch_xs.reshape([batch_size, n_steps, n_inputs])\n        test_batch_xs, test_batch_ys = mnist.test.next_batch(batch_size)\n        test_batch_xs = test_batch_xs.reshape([batch_size, n_steps, n_inputs])\n        _, = sess.run([train_op], feed_dict={\n            x: batch_xs,\n            y: batch_ys\n        })\n        if step % 20 == 0:\n            print("Train set accuracy %s" % sess.run(accuracy, feed_dict={\n                x: test_batch_xs,\n                y: test_batch_ys\n            }))\n        step += 1\n\n    test_xs, test_ys = mnist.test.next_batch(128)\n    test_xs = test_xs.reshape([-1, n_steps, n_inputs])\n    print("Test set accuracy %s" % sess.run(accuracy, feed_dict={\n                x: test_xs,\n                y: test_ys\n            }))\n<\/code><\/pre>\n","protected":false},"excerpt":{"rendered":"
\u6982\u5ff5 \u4e3a\u4ec0\u4e48\u9700\u8981 RNN(\u5faa\u73af\u795e\u7ecf\u7f51\u7edc) DNN\u90fd\u53ea\u80fd\u5355\u72ec\u7684\u53d6\u5904\u7406\u4e00\u4e2a\u4e2a\u7684\u8f93\u5165 \u524d\u4e00\u4e2a\u8f93\u5165\u548c\u540e\u4e00\u4e2a\u8f93\u5165\u662f\u5b8c\u5168\u6ca1\u6709\u5173\u7cfb\u7684\u3002 \u4f46\u662f\uff0c\u67d0\u4e9b\u4efb\u52a1   \u2026 …<\/p>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"ngg_post_thumbnail":0},"categories":[262],"tags":[263,257],"_links":{"self":[{"href":"\/index.php?rest_route=\/wp\/v2\/posts\/9241"}],"collection":[{"href":"\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=9241"}],"version-history":[{"count":1,"href":"\/index.php?rest_route=\/wp\/v2\/posts\/9241\/revisions"}],"predecessor-version":[{"id":9242,"href":"\/index.php?rest_route=\/wp\/v2\/posts\/9241\/revisions\/9242"}],"wp:attachment":[{"href":"\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=9241"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=9241"},{"taxonomy":"post_tag","embeddable":true,"href":"\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=9241"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}