fruitpicker01's picture
Update utils.py
6a6025e verified
import re
import numpy as np
import pandas as pd
import pymorphy2
import torch
import torch.nn.functional as F
from transformers import AutoTokenizer, AutoModel
morph = pymorphy2.MorphAnalyzer()
tokenizer = AutoTokenizer.from_pretrained("ai-forever/ru-en-RoSBERTa")
model = AutoModel.from_pretrained("ai-forever/ru-en-RoSBERTa")
def cosine_similarity(embedding1, embedding2):
embedding1 = np.array(embedding1)
embedding2 = np.array(embedding2)
dot_product = np.dot(embedding1, embedding2)
norm_a = np.linalg.norm(embedding1)
norm_b = np.linalg.norm(embedding2)
return dot_product / (norm_a * norm_b)
def pool(hidden_state, mask, pooling_method="cls"):
if pooling_method == "mean":
s = torch.sum(hidden_state * mask.unsqueeze(-1).float(), dim=1)
d = mask.sum(axis=1, keepdim=True).float()
return s / d
elif pooling_method == "cls":
return hidden_state[:, 0]
def text_to_embedding(text, tokenizer, model):
# Токенизация текста
tokenized_inputs = tokenizer(text, max_length=512, padding=True, truncation=True, return_tensors="pt")
with torch.no_grad():
outputs = model(**tokenized_inputs)
embeddings = pool(
outputs.last_hidden_state,
tokenized_inputs["attention_mask"],
pooling_method="cls" # or try "mean"
)
embeddings = F.normalize(embeddings, p=2, dim=1).numpy()
return embeddings
def preprocess_text(text):
lemmas = [] # Для хранения лемм
for token in text.split():
parsed = morph.parse(token)[0] # Морфологический разбор токена
# Лемматизация
if parsed.normal_form and parsed.normal_form.strip():
lemmas.append(parsed.normal_form) # Добавляем лемму
return " ".join(lemmas) if lemmas else ""
def product_extraction(text):
lemmas = preprocess_text(text)
if 'кредитный бизнес-' in lemmas:
return 'кредитная бизнес-карта'
elif 'выпустить бизнес-карта' in lemmas:
return 'бизнес-карта'
elif ('расчётный счёт' in lemmas) or ('открыть счёт' in lemmas):
return 'расчетный счет'
elif 'бизнес-карта' in lemmas:
return 'бизнес-карта'
elif 'бизнес-кешбэк' in lemmas:
return 'cashback'
elif 'перевод' in lemmas:
return 'переводы'
elif 'кредит' in lemmas:
return 'кредит'
elif 'эквайринг' in lemmas:
return 'эквайринг'
elif 'зарплатный проект' in lemmas:
return 'зарплатный проект'
elif 'вклад' in lemmas:
return 'вклад'
elif 'депозит' in lemmas:
return 'депозит'
return 'прочее'
def best_text_choice(texts, core_df, tokenizer, model, coef=1):
'''
Функция для выбора лучшего текста, и оценки его успешности
'''
scoring_list = []
embeddings_df = core_df.copy()
texts_df = pd.DataFrame(texts, columns=['texts'])
texts_df['texts_lower'] = texts_df['texts'].apply(lambda x: x.lower())
texts_df['texts_'] = 'search_query: ' + texts_df['texts_lower']
texts_df['embeddings'] = texts_df['texts_'].apply(lambda x: text_to_embedding(x, tokenizer, model)[0])
texts_df['product'] = texts_df['texts'].apply(product_extraction)
best_text = ''
score = 0
for index, row in texts_df.iterrows():
product = row['product']
embeddings_df['similarity'] = embeddings_df['embedding'].apply(lambda x: cosine_similarity(x, row['embeddings']))
embeddings_df['score'] = embeddings_df['value'] * embeddings_df['similarity']
score_ = np.mean([(embeddings_df
.sort_values(by=['product_type', 'score'], ascending=[True, False])
.query('product_type == @product')['score'][:3].mean() * coef),
embeddings_df
.sort_values(by='similarity', ascending=False)
.query('product_type != @product')['score'][:3].mean()])
scoring_list.append([row['texts'], 100*score_ / embeddings_df.query('product_type == @product')['value'].max()])
if score_ > score:
score = score_
best_text = row['texts']
# ratio = score / embeddings_df.query('product_type == @product')['value'].max()
scoring_df = pd.DataFrame(scoring_list, columns=['text', 'score'])
scoring_df = scoring_df.sort_values(by='score', ascending=False).reset_index(drop=True)
scoring_df.index += 1
return scoring_df.reset_index().rename(columns={'index': 'Место'})[['Место', 'text']]