cant change vocab vector

[DanielBeck93]

When I try to change the word embeddings nothing happens

`
import spacy_universal_sentence_encoder
nlp_use = spacy_universal_sentence_encoder.load_model('en_use_lg')

vector = nlp_use('her').vector
nlp_use.vocab.set_vector("Ertha", vector)

print(vector)
print(nlp_use("Ertha"))
`
The two print statements should be equal.

[MartinoMensio]

Hi @DanielBeck93,
At the moment this wrapper library did not allow to set the vectors, it provided only a read-only copy of the values from TensorFlow Hub.
Given your example, I understood that the feature is quite easy to be provided.

With the new v0.4.2, you can achieve setting any vectors to tokens, spans and even entire docs.
The comparison is done on the string that you pass to the vocab.set_vector function. Only exact matches will count.

import spacy_universal_sentence_encoder
nlp_use = spacy_universal_sentence_encoder.load_model('en_use_lg')

vector = nlp_use('her').vector
nlp_use.vocab.set_vector("Ertha", vector)

# compare the vectors, now they are equal
print((vector == nlp_use("Ertha").vector).all())

Please note that changing the vector to a single word does not change the vector of the spans/sentences/documents containing it.
If you try to compare the sentences "good for her" and "good for Ertha", they will be different, even if you have set the value of "Ertha" to be equal to "her". This happens because the USE model isn't a simple average of the word embeddings.
If you want to substitute the words in the document, my suggestion would be to substitute the string "Ertha" with "her" in the text before creating the document with nlp_use(text).

Let me know if this works for you.
Best,
Martino

[MartinoMensio]

The version 0.4.2 has been yanked due to a problem found in #14