将BERT模型转换为TFLite

Question

我有这个代码的语义搜索引擎建立使用预先训练的bert模型。我想将此模型转换为tflite，以便将其部署到google mlkit。我想知道如何转换它。我想知道是否有可能将其转换为tflite。这可能是因为它在tensorflow官方网站上提到了:https://www.tensorflow.org/lite/convert。但我不知道从何说起

代码：

from sentence_transformers import SentenceTransformer

# Load the BERT model. Various models trained on Natural Language Inference (NLI) https://github.com/UKPLab/sentence-transformers/blob/master/docs/pretrained-models/nli-models.md and 
# Semantic Textual Similarity are available https://github.com/UKPLab/sentence-transformers/blob/master/docs/pretrained-models/sts-models.md

model = SentenceTransformer('bert-base-nli-mean-tokens')

# A corpus is a list with documents split by sentences.

sentences = ['Absence of sanity', 
             'Lack of saneness',
             'A man is eating food.',
             'A man is eating a piece of bread.',
             'The girl is carrying a baby.',
             'A man is riding a horse.',
             'A woman is playing violin.',
             'Two men pushed carts through the woods.',
             'A man is riding a white horse on an enclosed ground.',
             'A monkey is playing drums.',
             'A cheetah is running behind its prey.']

# Each sentence is encoded as a 1-D vector with 78 columns
sentence_embeddings = model.encode(sentences)

print('Sample BERT embedding vector - length', len(sentence_embeddings[0]))

print('Sample BERT embedding vector - note includes negative values', sentence_embeddings[0])

#@title Sematic Search Form

# code adapted from https://github.com/UKPLab/sentence-transformers/blob/master/examples/application_semantic_search.py

query = 'Nobody has sane thoughts' #@param {type: 'string'}

queries = [query]
query_embeddings = model.encode(queries)

# Find the closest 3 sentences of the corpus for each query sentence based on cosine similarity
number_top_matches = 3 #@param {type: "number"}

print("Semantic Search Results")

for query, query_embedding in zip(queries, query_embeddings):
    distances = scipy.spatial.distance.cdist([query_embedding], sentence_embeddings, "cosine")[0]

    results = zip(range(len(distances)), distances)
    results = sorted(results, key=lambda x: x[1])

    print("\n\n======================\n\n")
    print("Query:", query)
    print("\nTop 5 most similar sentences in corpus:")

    for idx, distance in results[0:number_top_matches]:
        print(sentences[idx].strip(), "(Cosine Score: %.4f)" % (1-distance))

Answer 1

首先，你需要用TensorFlow编写你的模型，你使用的包是用PyTorch写的。Huggingface的Transformers有TensorFlow模型，你可以从它开始。此外，他们还有安卓版的TFLite-ready models。

通常，您首先要有一个TensorFlow模型。将其保存为SavedModel格式：

tf.saved_model.save(pretrained_model, "/tmp/pretrained-bert/1/")

你可以在上面运行这个转换器。

Answer 2

你有没有试过运行转换工具(tflite_convert)，它有什么问题吗？

顺便说一句，你可能想看看使用Bert模型的TFLite团队的QA示例。https://github.com/tensorflow/examples/tree/master/lite/examples/bert_qa/android

Answer 3

我找不到任何关于使用BERT模型在移动设备上获取文档嵌入并计算k最近文档搜索的信息，如您的示例所示。这可能也不是一个好主意，因为BERT模型的执行成本很高，并且有大量的参数，因此模型文件大小(400mb+)也很大。

然而，you can now use BERT和MobileBERT用于文本分类和移动问答。也许你可以从他们的demo app开始，它与MobileBERT的tflite模型接口，正如迅开提到的那样。我相信在不久的将来，会有更好的支持你的用例。