如何使用另一个模型的词汇表初始化 gensim 模型？

Question

I'm training some embeddings on a large corpus. I gather from gensim 's documentation that it builds the vocabulary before beginning training. In my case, building the vocabulary takes many hours. I'd like to save time by re-using the vocabulary from the first model. How can I do this? the .build_vocab method can't take the vocabulary object from another model.

Here's a dummy example:

from gensim.models import FastText, Word2Vec
sentences = ["where are my goats", "yay i found my goats"]
m1 = Word2Vec(sentences, size  = 3)
m2 = Word2Vec(size = 4)
m2.build_vocab(m1.vocabulary) # doesn't work

Answer 1

基于这个错误（“重用用 scan_vocab 构建的词汇是不可能的”）我相信这在这个时候是不可能的。

Answer 2

build_vocab() says to survey a corpus of texts & configure the model's vocabulary from that corpus – so it doesn't take another model's internal state.

But you could either:

• save the model after vocabulary-discovery, for later reuse from that point on; or
• just directly modify any model however you'd like, to mimic another's state

For example, consider an initial session:

vocab_model = Word2Vec(size=3)
vocab_model.build_vocab(sentences)
vocab_model.save('vocab_initialized_but_untrained_model.w2v')

Now, you could immediately continue to train that model...

vocab_model.train(sentences, total_examples=vocab_model.corpus_count, epochs=10)

...and then perhaps do other work with, and .save() , that trained model, too.

But then also, later, you could simply re-load the vocabulary-initialized model and do other tinkering/training:

prior_model = Word2Vec.load('vocab_initialized_but_untrained_model.w2v')
// more operations on that model

And further, you can always directly modify the parts of a model however you'd like – though in some cases this may break the existing code's expectations about the model's state. For example:

source_model = Word2Vec.load('original_model.w2v')
new_model = Word2Vec(size=4)
new_model.vocabulary = source_model.vocabulary

(You probably need to copy over some other fields as well, to mimic all the effects of the 1st model's initialization, and perhaps re-trigger the final steps of the build_vocab() with your new size /modes. See the source code & especially the methods having to do with _weights( or prepare_ .)