TensorFlow实现句子语义相似度报错RuntimeError求助

这个错误的核心原因是你的代码用的是TensorFlow 1.x的图模式API，但当前环境默认开启了TensorFlow 2.x的即时执行（Eager Execution）。hub.Module()是TF1时代的模块加载方式，依赖计算图，而Eager模式下没有全局计算图，所以会抛出"Exporting/importing meta graphs is not supported..."的错误。

下面给你两种可行的解决方案，推荐第二种更贴合TF2的主流用法：

方案一：禁用Eager Execution，沿用TF1风格代码 #

只需要在导入TensorFlow之后，添加一行代码关闭Eager模式，就能让原来的代码正常运行：

import tensorflow as tf
# 添加这行，禁用Eager Execution
tf.compat.v1.disable_eager_execution()
import tensorflow_hub as hub
import numpy as np
from sklearn.metrics.pairwise import cosine_similarity

# get cosine similairty matrix
def cos_sim(input_vectors):
    similarity = cosine_similarity(input_vectors)
    return similarity

# get topN similar sentences
def get_top_similar(sentence, sentence_list, similarity_matrix, topN):
    # find the index of sentence in list
    index = sentence_list.index(sentence)
    # get the corresponding row in similarity matrix
    similarity_row = np.array(similarity_matrix[index, :])
    # get the indices of top similar
    indices = similarity_row.argsort()[-topN:][::-1]
    return [sentence_list[i] for i in indices]

module_url = "https://tfhub.dev/google/universal-sentence-encoder/2"
# Import the Universal Sentence Encoder's TF Hub module
embed = hub.Module(module_url)

# Reduce logging output.
tf.logging.set_verbosity(tf.logging.ERROR)

sentences_list = [
    # phone related
    'My phone is slow',
    'My phone is not good',
    'I need to change my phone. It does not work well',
    'How is your phone?',
    # age related
    'What is your age?',
    'How old are you?',
    'I am 10 years old',
    # weather related
    'It is raining today',
    'Would it be sunny tomorrow?',
    'The summers are here.'
]

with tf.Session() as session:
    session.run([tf.global_variables_initializer(), tf.tables_initializer()])
    sentences_embeddings = session.run(embed(sentences_list))
    similarity_matrix = cos_sim(np.array(sentences_embeddings))

sentence = "It is raining today"
top_similar = get_top_similar(sentence, sentences_list, similarity_matrix, 3)

# printing the list using loop
for x in range(len(top_similar)):
    print(top_similar[x])

方案二：改用TF2.x兼容的API（推荐） #

TF2.x推荐使用更简洁的hub.load()或者Keras层的方式加载模型，不需要手动管理tf.Session()，代码更简洁易维护：

import tensorflow as tf
import tensorflow_hub as hub
import numpy as np
from sklearn.metrics.pairwise import cosine_similarity

# get cosine similairty matrix
def cos_sim(input_vectors):
    similarity = cosine_similarity(input_vectors)
    return similarity

# get topN similar sentences
def get_top_similar(sentence, sentence_list, similarity_matrix, topN):
    index = sentence_list.index(sentence)
    similarity_row = np.array(similarity_matrix[index, :])
    indices = similarity_row.argsort()[-topN:][::-1]
    return [sentence_list[i] for i in indices]

# 使用TF2兼容的方式加载模型
module_url = "https://tfhub.dev/google/universal-sentence-encoder/4"  # 推荐用v4版本，更适配TF2
embed_model = hub.load(module_url)

sentences_list = [
    # phone related
    'My phone is slow',
    'My phone is not good',
    'I need to change my phone. It does not work well',
    'How is your phone?',
    # age related
    'What is your age?',
    'How old are you?',
    'I am 10 years old',
    # weather related
    'It is raining today',
    'Would it be sunny tomorrow?',
    'The summers are here.'
]

# 直接生成句向量，不需要Session
sentences_embeddings = embed_model(sentences_list).numpy()
similarity_matrix = cos_sim(sentences_embeddings)

sentence = "It is raining today"
top_similar = get_top_similar(sentence, sentences_list, similarity_matrix, 3)

for item in top_similar:
    print(item)

改动说明： #

1. 使用hub.load()替代hub.Module()，加载TF2兼容的模型版本（这里用了v4，你也可以选其他TF2兼容的版本）
2. 不需要手动初始化变量和Session，直接调用模型生成向量，用.numpy()把Tensor转成Numpy数组
3. 代码更简洁，符合TF2的编程风格

内容的提问来源于stack exchange，提问作者user12907213