comparevec2vecwithada.py

# -*- coding: utf-8 -*-
"""compareVec2VecWithAda.ipynb

Automatically generated by Colaboratory.

Original file is located at
    https://colab.research.google.com/drive/1jPaNXdO0_oW6VczlWfm5RPUVpMtVQD9c
"""

import pandas as pd
import numpy as np
import openai
from sklearn.metrics.pairwise import cosine_similarity
from tensorflow.keras.models import load_model
from transformers import AutoTokenizer, AutoModel
import torch
import torch.nn.functional as F

# Load model (available from Hugging Face)
tokenizer = AutoTokenizer.from_pretrained('all-mpnet-base-v2')
model = AutoModel.from_pretrained('all-mpnet-base-v2')

# Define cosine similarity loss
def cosine_similarity_loss(y_true, y_pred):
    y_true = tf.nn.l2_normalize(y_true, axis=-1)
    y_pred = tf.nn.l2_normalize(y_pred, axis=-1)
    return -tf.reduce_mean(y_true * y_pred, axis=-1)


#Mean Pooling - Take attention mask into account for correct averaging
def mean_pooling(model_output, attention_mask):
    token_embeddings = model_output[0] #First element of model_output contains all token embeddings
    input_mask_expanded = attention_mask.unsqueeze(-1).expand(token_embeddings.size()).float()
    return torch.sum(token_embeddings * input_mask_expanded, 1) / torch.clamp(input_mask_expanded.sum(1), min=1e-9)

loaded_model = load_model('mpnet2adaE75V4.h5', custom_objects={'cosine_similarity_loss': cosine_similarity_loss})

openai.api_key="insert API key here"

# load in csv of 10,000 embeddings in our test set paired with the original reviews
df2 = pd.read_csv('Actual_Embeddings.csv')

# Convert strings of lists to numpy arrays. this takes a while
df2['Actual_Embeddings'] = df2['Actual_Embeddings'].apply(eval).apply(np.array)


def get_top_5_texts(query):
    encoded_input = tokenizer(query, padding=True, truncation=True, return_tensors='pt')

    with torch.no_grad():
        model_output = model(**encoded_input)

    mpnetEmbeddings = mean_pooling(model_output, encoded_input['attention_mask'])

    mpnetEmbeddings = F.normalize(mpnetEmbeddings, p=2, dim=1)
    mpnetEmbeddings = mpnetEmbeddings.detach().cpu().numpy()
    mpnetEmbeddings = np.reshape(mpnetEmbeddings, (1,-1))
    query_embedding = loaded_model.predict(mpnetEmbeddings)

    similarities = [cosine_similarity(query_embedding.reshape(1, -1), emb.reshape(1, -1))[0][0] for emb in df2['Actual_Embeddings']]

    print("Converted MPNet Embedding Results:")
    top_5_idx2 = np.argsort(similarities)[-5:][::-1]
    for i, idx in enumerate(top_5_idx2, 1):
        print(f'Text {i}')
        print(df2['combined'].iloc[idx])
        print("\n")

    response = openai.Embedding.create(input=query, model="text-embedding-ada-002")
    query_embedding = np.array(response['data'][0]['embedding'])
    similarities2 = [cosine_similarity(query_embedding.reshape(1, -1), emb.reshape(1, -1))[0][0] for emb in df2['Actual_Embeddings']]

    print("OpenAI Embedding Results:")
    top_5_idx2 = np.argsort(similarities2)[-5:][::-1]
    for i, idx in enumerate(top_5_idx2, 1):
        print(f'Text {i}')
        print(df2['combined'].iloc[idx])
        print("\n")

while True:
    query = input("Enter your query: ")
    get_top_5_texts(query)