網絡結構：
代碼如下：
1 # encoding: utf-8 2 import tensorflow as tf 3 from tensorflow import keras 4 from tensorflow.keras import layers, losses, optimizers, Model 5 # from exam_rnn import MyRNN # (可以分離寫) 6 7 batchsz = 128 # 批量大小 8 total_words = 10000 # 詞匯表大小N_vocab 9 max_review_len = 80 # 句子最大長度s，大於的句子部分將截斷，小於的將填充 10 embedding_len = 100 # 詞向量特征長度n 11 # 加載IMDB 數據集，此處的數據采用數字編碼，一個數字代表一個單詞 12 (x_train, y_train), (x_test, y_test) = keras.datasets.imdb.load_data(num_words=total_words) 13 # 打印輸入的形狀，標簽的形狀 14 print(x_train.shape, len(x_train[0]), y_train.shape) # (25000,) 218 (25000,) 15 print(x_test.shape, len(x_test[0]), y_test.shape) # (25000,) 68 (25000,) 16 17 '''# 查看編碼 18 # 數字編碼表 19 word_index = keras.datasets.imdb.get_word_index() 20 # 打印出編碼表的單詞和對應的數字 21 # for k,v in word_index.items(): 22 # print(k,v) 23 24 # 前面4 個ID 是特殊位 25 word_index = {k: (v + 3) for k, v in word_index.items()} 26 word_index["<PAD>"] = 0 # 填充標志 27 word_index["<START>"] = 1 # 起始標志 28 word_index["<UNK>"] = 2 # 未知單詞的標志 29 word_index["<UNUSED>"] = 3 30 # 翻轉編碼表 31 reverse_word_index = dict([(value, key) for (key, value) in 32 word_index.items()]) 33 34 35 # 對於一個數字編碼的句子，通過如下函數轉換為字符串數據： 36 def decode_review(text): 37 return ' '.join([reverse_word_index.get(i, '?') for i in text]) 38 39 40 print(decode_review(x_train[0])) 41 ''' 42 43 # 截斷和填充句子，使得等長，此處長句子保留句子后面的部分，短句子在前面填充 44 x_train = keras.preprocessing.sequence.pad_sequences(x_train, maxlen=max_review_len) 45 x_test = keras.preprocessing.sequence.pad_sequences(x_test, maxlen=max_review_len) 46 47 # 構建數據集，打散，批量，並丟掉最后一個不夠batchsz 的batch 48 db_train = tf.data.Dataset.from_tensor_slices((x_train, y_train)) 49 db_train = db_train.shuffle(1000).batch(batchsz, drop_remainder=True) 50 db_test = tf.data.Dataset.from_tensor_slices((x_test, y_test)) 51 db_test = db_test.batch(batchsz, drop_remainder=True) 52 53 print('x_train shape:', x_train.shape, tf.reduce_max(y_train), tf.reduce_min(y_train)) 54 # x_train shape: (25000, 80) tf.Tensor(1, shape=(), dtype=int64) tf.Tensor(0, shape=(), dtype=int64) 55 print('x_test shape:', x_test.shape) 56 # x_test shape: (25000, 80) 57 58 59 class MyRNN(Model): 60 # Cell 方式構建多層網絡 61 def __init__(self, units): 62 super(MyRNN, self).__init__() 63 # [b, 64]，構建Cell 初始化狀態向量，重復使用 64 self.state0 = [tf.zeros([batchsz, units])] # 128,64 65 self.state1 = [tf.zeros([batchsz, units])] # 128,64 66 # 詞向量編碼 [b, 80] => [b, 80, 100] 67 self.embedding = layers.Embedding(total_words, embedding_len, input_length=max_review_len) 68 # 構建2 個Cell，使用dropout 技術防止過擬合 69 self.rnn_cell0 = layers.SimpleRNNCell(units)# , dropout=0.5) 70 self.rnn_cell1 = layers.SimpleRNNCell(units)#, dropout=0.5) 71 # 構建分類網絡，用於將CELL 的輸出特征進行分類，2 分類 72 # [b, 80, 100] => [b, 64] => [b, 1] 73 self.outlayer = layers.Dense(1) 74 75 def call(self, inputs, training=None): 76 x = inputs # [128, 80] 77 # 獲取詞向量: [128, 80] => [128, 80, 100] 78 x = self.embedding(x) 79 # 通過2 個RNN CELL,[128, 80, 100] => [128, 64] 80 state0 = self.state0 81 state1 = self.state1 82 for word in tf.unstack(x, axis=1): # word: [128, 100] 83 out0, state0 = self.rnn_cell0(word, state0, training) 84 out1, state1 = self.rnn_cell1(out0, state1, training) 85 86 # 末層最后一個輸出作為分類網絡的輸入: [128, 64] => [128, 1] 87 x = self.outlayer(out1) 88 # 通過激活函數，p(y is pos|x) 89 prob = tf.sigmoid(x) 90 return prob 91 92 93 def main(): 94 units = 64 # RNN 狀態向量長度n 95 epochs = 20 # 訓練epochs 96 model = MyRNN(units) # 創建模型 97 # 裝配 98 model.compile(optimizer=optimizers.Adam(0.001), 99 loss=losses.BinaryCrossentropy(), metrics=['accuracy'], 100 experimental_run_tf_function=False) 101 # 訓練和驗證 102 model.fit(db_train, epochs=epochs, validation_data=db_test) 103 # 測試 104 scores = model.evaluate(db_test) 105 print("Final test loss and accuracy :", scores) 106 107 108 if __name__ == '__main__': 109 main()
測試的誤差和准確率：
Final test loss and accuracy : [1.3201157276447002, 0.80188304]
下一次更新：LSTM情感分類問題