Tensorflow 2.4 搭建單層和多層 Bi-LSTM 模型

前言

本文使用 cpu 版本的 TensorFlow 2.4 ,分別搭建單層 Bi-LSTM 模型和多層 Bi-LSTM 模型完成文本分類任務。

確保使用 numpy == 1.19.0 左右的版本,否則在調用 TextVectorization 的時候可能會報 NotImplementedError 。

實現過程

1. 獲取數據

(1)我們本文用到的數據是電影的影評數據,每個樣本包含瞭一個對電影的評論文本和一個情感標簽,1 表示積極評論,0 表示負面評論,也就是說這是一份二分類的數據。

(2)我們通過 TensorFlow 內置的函數,可以從網絡上直接下載 imdb_reviews 數據到本地的磁盤,並取出訓練數據和測試數據。

(3)通過使用 tf.data.Dataset 相關的處理函數,我們將訓練數據和測試數據分別進行混洗,並且設置每個 batch 大小都是 64 ,每個樣本都是 (text, label) 的形式。如下我們取瞭任意一個 batch 中的前兩個影評文本和情感標簽。

import numpy as np
import tensorflow_datasets as tfds
import tensorflow as tf
import matplotlib.pyplot as plt
tfds.disable_progress_bar()
BUFFER_SIZE = 10000
BATCH_SIZE = 64
dataset, info = tfds.load('imdb_reviews', with_info=True, as_supervised=True)
train_dataset, test_dataset = dataset['train'], dataset['test']
train_dataset = train_dataset.shuffle(BUFFER_SIZE).batch(BATCH_SIZE).prefetch(tf.data.AUTOTUNE)
test_dataset = test_dataset.batch(BATCH_SIZE).prefetch(tf.data.AUTOTUNE)
for example, label in train_dataset.take(1):
    print('text: ', example.numpy()[:2])
    print('label: ', label.numpy()[:2])

部分樣本顯示:

 text: [ 
     b"First of all, I have to say I have worked for blockbuster and have seen quite a few movies to the point its tough for me to find something I haven't seen. Taking this into account, I want everyone to know that this movie was by far the worst film ever made, it made me pine for Gigli, My Boss's Daughter, and any other piece of junk you've ever seen. BeLyt must be out of his mind, I've only found one person who liked it and even they couldn't tell me what the movie was about. If you are able to decipher this movie and are able to tell me what it was about you have to either be the writer or a fortune teller because there's any other way a person could figure this crap out.<br /><br />FOR THE LOVE OF G-D STAY AWAY!"
     b"Just got out and cannot believe what a brilliant documentary this is. Rarely do you walk out of a movie theater in such awe and amazement. Lately movies have become so over hyped that the thrill of discovering something truly special and unique rarely happens. Amores Perros did this to me when it first came out and this movie is doing to me now. I didn't know a thing about this before going into it and what a surprise. If you hear the concept you might get the feeling that this is one of those touchy movies about an amazing triumph covered with over the top music and trying to have us fully convinced of what a great story it is telling but then not letting us in. Fortunetly this is not that movie. The people tell the story! This does such a good job of capturing every moment of their involvement while we enter their world and feel every second with them. There is so much beyond the climb that makes everything they go through so much more tense. Touching the Void was also a great doc about mountain climbing and showing the intensity in an engaging way but this film is much more of a human story. I just saw it today but I will go and say that this is one of the best documentaries I have ever seen."
 ]
label:  [0 1]

2. 處理數據

(1)想要在模型中訓練這些數據,必須將這些文本中的 token 都轉換成機器可以識別的整數,最簡單的方法就是使用 TextVectorization 來制作一個編碼器 encoder,這裡隻將出現次數最多的 1000 個 token 當做詞表,另外規定每個影評處理之後隻能保留最長 200 的長度,如果超過則會被截斷,如果不足則用填充字符對應的整數 0 補齊。

(2)這裡展現出來瞭某個樣本的經過整數映射止之後的結果,可以看到影評對應的整數數組長度為 200 。

MAX_SEQ_LENGTH = 200
VOCAB_SIZE = 1000
encoder = tf.keras.layers.experimental.preprocessing.TextVectorization(max_tokens=VOCAB_SIZE, output_sequence_length=MAX_SEQ_LENGTH)
encoder.adapt(train_dataset.map(lambda text, label: text))
vocab = np.array(encoder.get_vocabulary())
encoded_example = encoder(example)[:1].numpy()
print(encoded_example)
print(label[:1])

隨機選取一個樣本進行證書映射結果:

[[ 86   5  32  10  26   6 130  10  26 926  16   1   3  26 108 176   4 164
   93   6   2 215  30   1  16  70   6 160 140  10 731 108 647  11  78   1
   10 178 305   6 118  12  11  18  14  33 234   2 240  20 122  91   9  91
   70   1  16   1  56   1 580   3  99  81 408   5   1 825 122 108   1 217
   28  46   5  25 349 195  61 249  29 409  37 405   9   3  54  35 404 360
   70  49   2  18  14  43  45  23  24 491   6   1  11  18   3  24 491   6
  360  70  49   9  14  43  23  26   6 352  28   2 762  42   4   1   1  80
  213  99  81  97   4 409  96 811  11 638   1  13  16   2 116   5   1 766
  242   0   0   0   0   0   0   0   0   0   0   0   0   0   0   0   0   0
    0   0   0   0   0   0   0   0   0   0   0   0   0   0   0   0   0   0
    0   0   0   0   0   0   0   0   0   0   0   0   0   0   0   0   0   0
    0   0]]
tf.Tensor([0], shape=(1,), dtype=int64)

3. 單層 Bi-LSTM 模型

(1) 第一層是我們剛才定義好的 encoder ,將輸入的文本進行整數的映射。

(2)第二層是 Embedding 層,我們這裡設置瞭每個詞的詞嵌入維度為 32 維。

(3)第三層是 Bi-LSTM 層,這裡我們設置瞭每個 LSTM 單元的輸出維度為 16 維。

(4)第四層是一個輸出 8 維向量的全連接層,並且使用的 relu 激活函數。

(5)第五層是 Dropout ,設置神經元丟棄率為 0.5 ,主要是為瞭防止過擬合。

(6)第六層是一個輸出 1 維向量的全連接層,也就是輸出層,表示的是該樣本的 logit 。

model = tf.keras.Sequential([
    encoder,
    tf.keras.layers.Embedding( input_dim=len(encoder.get_vocabulary()), output_dim=32, mask_zero=True),
    tf.keras.layers.Bidirectional(tf.keras.layers.LSTM(16)),
    tf.keras.layers.Dense(8, activation='relu'),
    tf.keras.layers.Dropout(0.5),
    tf.keras.layers.Dense(1)
])

(7)在沒有經過訓練的模型上對文本進行預測,如果輸出小於 0 則為消極評論,如果大於 0 則為積極評論,我們可以看出這條評論本來應該是積極評論,但是卻輸出的 logit 卻是負數,即錯誤預測成瞭消極的。

sample_text = ('The movie was cool. The animation and the graphics were out of this world. I would recommend this movie.')
model.predict(np.array([sample_text]))

預測結果為:

array([[-0.01437075]], dtype=float32)

array([[-0.01437075]], dtype=float32)

(8)我們使用 BinaryCrossentropy 作為損失函數,需要註意的是如果模型輸出結果給到 BinaryCrossentropy 的是一個 logit 值(值域范圍 [-∞, +∞] ),則應該設置 from_logits=True 。如果模型輸出結果給到 BinaryCrossentropy 的是一個概率值 probability (值域范圍 [0, 1] ),則應該設置為 from_logits=False 。

(9)我們使用 Adam 作為優化器,並且設置學習率為 1e-3 。

(10)我們使用準確率 accuracy 作為評估指標。

(11)使用訓練數據訓練 10 個 epoch,同時每經過一個 epoch 使用驗證數據對模型進行評估。

model.compile(loss=tf.keras.losses.BinaryCrossentropy(from_logits=True),
              optimizer=tf.keras.optimizers.Adam(1e-3),
              metrics=['accuracy'])
history = model.fit(train_dataset, epochs=10,  validation_data=test_dataset, validation_steps=30)

訓練過程如下:

Epoch 1/10
391/391 [==============================] - 30s 65ms/step - loss: 0.6461 - accuracy: 0.5090 - val_loss: 0.4443 - val_accuracy: 0.8245
Epoch 2/10
391/391 [==============================] - 23s 58ms/step - loss: 0.4594 - accuracy: 0.6596 - val_loss: 0.3843 - val_accuracy: 0.8396
...
Epoch 10/10
391/391 [==============================] - 22s 57ms/step - loss: 0.3450 - accuracy: 0.8681 - val_loss: 0.3920 - val_accuracy: 0.8417

(12)訓練結束之後使用測試數據對模型進行測試,準確率可以達到 0.8319 。如果經過超參數的調整和足夠的訓練時間,效果會更好。

model.evaluate(test_dataset)

結果為:

391/391 [==============================] – 6s 15ms/step – loss: 0.3964 – accuracy: 0.8319

(13)使用訓練好的模型對影評進行分類預測,可以看出可以正確得識別文本的情感取向。因為負數表示的就是影評為負面情緒的。

sample_text = ('The movie was not cool. The animation and the graphics were bad. I would not recommend this movie.')
model.predict(np.array([sample_text]))

結果為:

array([[-1.6402857]], dtype=float32)

4. 多層 Bi-LSTM 模型

(1)我們上面隻是搭建瞭一層的 Bi-LSTM ,這裡我們搭建瞭兩層的 Bi-LSTM 模型,也就是在第二層 Bidirectional 之後又加瞭一層 Bidirectional ,這樣可以使我們的模型更加有效。我們使用的損失函數、優化器、評估指標都和上面一樣。

model = tf.keras.Sequential([
    encoder,
    tf.keras.layers.Embedding(len(encoder.get_vocabulary()), 32, mask_zero=True),
    tf.keras.layers.Bidirectional(tf.keras.layers.LSTM(32,  return_sequences=True)),
    tf.keras.layers.Bidirectional(tf.keras.layers.LSTM(16)),
    tf.keras.layers.Dense(8, activation='relu'),
    tf.keras.layers.Dropout(0.5),
    tf.keras.layers.Dense(1)
])
model.compile(loss=tf.keras.losses.BinaryCrossentropy(from_logits=True), optimizer=tf.keras.optimizers.Adam(1e-3), metrics=['accuracy'])
history = model.fit(train_dataset, epochs=10, validation_data=test_dataset,  validation_steps=30)

訓練過程如下:

Epoch 1/10
391/391 [==============================] - 59s 124ms/step - loss: 0.6170 - accuracy: 0.5770 - val_loss: 0.3931 - val_accuracy: 0.8135
Epoch 2/10
391/391 [==============================] - 45s 114ms/step - loss: 0.4264 - accuracy: 0.7544 - val_loss: 0.3737 - val_accuracy: 0.8380
...
Epoch 10/10
391/391 [==============================] - 45s 114ms/step - loss: 0.3138 - accuracy: 0.8849 - val_loss: 0.4069 - val_accuracy: 0.8323	

(2)訓練結束之後使用測試數據對模型進行測試,準確率可以達到 0.8217 。如果經過超參數的調整和足夠的訓練時間,效果會更好。

model.evaluate(test_dataset)

結果為:

391/391 [==============================] – 14s 35ms/step – loss: 0.4021 – accuracy: 0.8217

(3)使用訓練好的模型對影評進行分類預測,可以看出可以正確得識別文本的情感取向。因為正數表示的就是影評為積極情緒的。

sample_text = ('The movie was good. The animation and the graphics were very good. you should love movie.')
model.predict(np.array([sample_text]))

結果為:

array([[3.571126]], dtype=float32)

以上就是Tensorflow 2.4 搭建單層和多層 Bi-LSTM 模型的詳細內容,更多關於Tensorflow搭建Bi-LSTM模型的資料請關註WalkonNet其它相關文章!

推薦閱讀: