def get_similar_sentences(sen: str, no_of_variations: int, language_code: str):
check_input_language(language_code)
# get embedding vectors for sen
tok = LanguageTokenizer(language_code)
token_ids = tok.numericalize(sen)
embedding_vectors = get_embedding_vectors(sen, language_code)
# get learner
defaults.device = torch.device('cpu')
path = Path(__file__).parent
learn = load_learner(path / 'models' / f'{language_code}')
encoder = get_model(learn.model)[0]
embeddings = encoder.state_dict()['encoder.weight']
embeddings = np.array(embeddings)
# cos similarity of vectors
scores = []
for word_vec in embedding_vectors:
scores.append([cos_sim(word_vec, embdg) for embdg in embeddings])
word_ids = [np.argpartition(-np.array(score), no_of_variations)[:no_of_variations] for score in scores]
new_token_ids = []
# generating more variations than required so that we can then filter out the best ones
no_of_vars_per_token = ceil(no_of_variations/len(token_ids))*3
for i in range(len(token_ids)):
word_ids_list = word_ids[i].tolist()
word_ids_list.remove(token_ids[i])
for j in range(no_of_vars_per_token):
new_token_ids.append(token_ids[:i] + word_ids_list[j:j+1] + token_ids[i+1:])
new_sens = [tok.textify(tok_id) for tok_id in new_token_ids]
while sen in new_sens:
new_sens.remove(sen)
sen_with_sim_score = [(new_sen, get_sentence_similarity(sen, new_sen, language_code)) for new_sen in new_sens]
sen_with_sim_score.sort(key=lambda x: x[1], reverse=True)
new_sens = [sen for sen, _ in sen_with_sim_score]
return new_sens[:no_of_variations]
def get_sentence_encoding(input: str, language_code: str):
check_input_language(language_code)
tok = LanguageTokenizer(language_code)
token_ids = tok.numericalize(input)
# get learner
defaults.device = torch.device('cpu')
path = Path(__file__).parent
learn = load_learner(path / 'models' / f'{language_code}')
m = learn.model
kk0 = m[0](Tensor([token_ids]).to(torch.int64))
return np.array(kk0[0][-1][0][-1])
def get_similar_sentences(sen: str,
no_of_variations: int,
language_code: str,
degree_of_aug: float = 0.1):
check_input_language(language_code)
# get embedding vectors for sen
tok = LanguageTokenizer(language_code)
token_ids = tok.numericalize(sen)
embedding_vectors = get_embedding_vectors(sen, language_code)
# get learner
defaults.device = torch.device('cpu')
path = Path(__file__).parent
learn = load_learner(path / 'models' / f'{language_code}')
encoder = get_model(learn.model)[0]
encoder.reset()
embeddings = encoder.state_dict()['encoder.weight']
embeddings = np.array(embeddings)
# cos similarity of vectors
scores = cosine_similarity(embedding_vectors, embeddings)
word_ids = [
np.argpartition(-np.array(score),
no_of_variations + 1)[:no_of_variations + 1]
for score in scores
]
word_ids = [ids.tolist() for ids in word_ids]
for i, ids in enumerate(word_ids):
word_ids[i] = [wid for wid in word_ids[i] if wid != token_ids[i]]
# generating more variations than required so that we can then filter out the best ones
buffer_multiplicity = 2
new_sen_tokens = []
for i in range(no_of_variations):
for k in range(buffer_multiplicity):
new_token_ids = []
ids = sorted(
random.sample(range(len(token_ids)),
max(1, int(degree_of_aug * len(token_ids)))))
for j in range(len(token_ids)):
if j in ids:
new_token_ids.append(word_ids[j][(i + k) %
len(word_ids[j])])
else:
new_token_ids.append(token_ids[j])
new_token_ids = list(map(lambda x: int(x), new_token_ids))
new_sen_tokens.append(new_token_ids)
new_sens = [tok.textify(sen_tokens) for sen_tokens in new_sen_tokens]
while sen in new_sens:
new_sens.remove(sen)
sen_with_sim_score = [(new_sen,
get_sentence_similarity(sen, new_sen,
language_code))
for new_sen in new_sens]
sen_with_sim_score.sort(key=lambda x: x[1], reverse=True)
new_sens = [sen for sen, _ in sen_with_sim_score]
return new_sens[:no_of_variations]
def get_embedding_vectors(input: str, language_code: str):
check_input_language(language_code)
tok = LanguageTokenizer(language_code)
token_ids = tok.numericalize(input)
# get learner
defaults.device = torch.device('cpu')
path = Path(__file__).parent
learn = load_learner(path / 'models' / f'{language_code}')
encoder = get_model(learn.model)[0]
embeddings = encoder.state_dict()['encoder.weight']
embeddings = np.array(embeddings)
embedding_vectors = []
for token in token_ids:
embedding_vectors.append(embeddings[token])
return embedding_vectors