使用 tensorflow 2/keras 对 imdb 数据集进行训练给出了奇怪的结果

Question

我是 tensorflow2/keras 的新手。我在 tensorflow 网站上关注了这个tutorial。我没有将文本数据下载到目录，而是使用tensorflow_datasets 将 imdb 数据集直接加载到张量/numpy 数组中。下面是我的代码。

import os
import re
import string
import pandas as pd
import numpy as np
import tensorflow as tf
import tensorflow_hub as hub
import tensorflow_datasets as tfds

import matplotlib.pyplot as plt
from tensorflow.keras import layers
from tensorflow.keras.layers.experimental.preprocessing import TextVectorization

print("Version: ", tf.__version__)
print("Eager mode: ", tf.executing_eagerly())
print("Hub version: ", hub.__version__)
print("GPU is", "available" if tf.config.list_physical_devices('GPU') else "NOT AVAILABLE")

train_data, test_data = tfds.load(name="imdb_reviews", split=["train", "test"], 
                                  batch_size=-1, as_supervised=True)

X_train, y_train = tfds.as_numpy(train_data)
X_test, y_test = tfds.as_numpy(test_data)

# process text
def custom_standardization(input_data):
  lowercase = tf.strings.lower(input_data)
  stripped_html = tf.strings.regex_replace(lowercase, '<br />', ' ')
  return tf.strings.regex_replace(stripped_html,
                                  '[%s]' % re.escape(string.punctuation),
                                  '')
max_features = 1000
sequence_length = 50

vectorize_layer = TextVectorization(
    standardize=custom_standardization,
    max_tokens=max_features,
    output_mode='int',
    output_sequence_length=sequence_length)
# Make a text-only dataset (without labels), then call adapt
vectorize_layer.adapt(X_train)

def vectorize_text(text):
  text = tf.expand_dims(text, -1)
  return vectorize_layer(text)

#check data
first_review, first_label = X_train[0], y_train[0]
print("Review", first_review)
print("Vectorized review", vectorize_text(first_review))
print("11 ---> ",vectorize_layer.get_vocabulary()[11])
print(" 44 ---> ",vectorize_layer.get_vocabulary()[44])
print('Vocabulary size: {}'.format(len(vectorize_layer.get_vocabulary())))


# vectorize both train and test text data
X_train = vectorize_text(X_train)
X_test = vectorize_text(X_test)


embedding_dim = 16
#define and compile model
model = tf.keras.Sequential([
  layers.Embedding(max_features + 1, embedding_dim),
  layers.GlobalAveragePooling1D(),
  layers.Dropout(0.2),
  layers.Dense(256, activation='relu'),
  layers.Dropout(0.2),
  layers.Dense(1)])
model.summary()
model.compile(optimizer='adam', loss='categorical_crossentropy', metrics=['accuracy'])
# fit the model
history = model.fit(X_train, y_train, epochs=5, batch_size=32, verbose=2, validation_data=(X_test, y_test))

我得到的输出如下：

_________________________________________________________________
Epoch 1/5
782/782 - 4s - loss: 0.0000e+00 - accuracy: 0.5000 - val_loss: 0.0000e+00 - val_accuracy: 0.5000
Epoch 2/5
782/782 - 4s - loss: 0.0000e+00 - accuracy: 0.5000 - val_loss: 0.0000e+00 - val_accuracy: 0.5000
Epoch 3/5
782/782 - 4s - loss: 0.0000e+00 - accuracy: 0.5000 - val_loss: 0.0000e+00 - val_accuracy: 0.5000
Epoch 4/5
782/782 - 4s - loss: 0.0000e+00 - accuracy: 0.5000 - val_loss: 0.0000e+00 - val_accuracy: 0.5000
Epoch 5/5
782/782 - 4s - loss: 0.0000e+00 - accuracy: 0.5000 - val_loss: 0.0000e+00 - val_accuracy: 0.5000

准确率都是 50%！有些不对劲。我很困惑为什么会这样？我按照教程，从头开始训练嵌入层。花了几个小时试图找出原因。 有谁知道为什么出错了？谢谢！

Answer 1

您的密集层是linear，这意味着您没有将任何激活传递给它。有几种解决方案，就像您正在做的那样binary_classification。

如果你想按原样使用Dense(1)，那么你应该将损失函数更改为：

model.compile(optimizer='adam', loss=tf.keras.losses.BinaryCrossentropy(from_logits=True),
              metrics=['accuracy'])

或者你可以使用Dense(1, activation = tf.nn.sigmoid)，那么你的损失函数应该是：

model.compile(optimizer='adam', loss='binary_crossentropy',
              metrics=['accuracy'])

如果您的标签是热编码的，或者另一种解决方案，您可以设置Dense(2, activation = tf.nn.softmax) 然后损失函数可以是：

model.compile(optimizer='adam', loss='categorical_crossentropy',
              metrics=['accuracy'])

我通过更改复制了您的确切代码：

  layers.Dense(1, activation = tf.nn.sigmoid)])
  model.compile(optimizer='adam', loss='binary_crossentropy', metrics= 
                                   ['accuracy'])

得到：

782/782 [==============================] - 4s 5ms/step - loss: 0.4836 - accuracy: 0.7675 - val_loss: 0.5065 - val_accuracy: 0.7447

编辑 2：sigmoid 将您的输出压缩到 [0,1] 的范围内。例如，要获得预测，您需要所有 x_test。 X_test[0] 会打乱预测，因为它只包含第一句话。使用以下方法获取预测：

y_hat = model.predict(X_test)
y_hat

array([[0.8105568 ],
       [0.6332975 ],
       [0.20526059],
       ...,
       [0.03132877],
       [0.5318063 ],
       [0.8626927 ]], dtype=float32)

现在设置一个阈值并将它们转换为 0 和 1。这可以通过以下方式完成：

y_pred = [1 * (x[0]>=0.5) for x in y_hat]

y_pred 将包含零和一。注意这里的阈值为0.5，这意味着如果sigmoid的输出大于0.5，则属于第二类。