app.py

import streamlit as st
import pandas as pd
import torch

from backend import inference
from backend.config import MODELS_ID, QA_MODELS_ID, SEARCH_MODELS_ID
from backend.utils import load_gender_data

st.title('Flax-Sentence-Tranformers')

st.sidebar.image("./hf-sbert.jpg", width=300)
st.sidebar.title('Navigation')
menu = st.sidebar.radio("", options=["Contributions & Evaluation", "Sentence Similarity", "Asymmetric QA", "Search / Cluster",
                                     "Gender Bias Evaluation"], index=0)

st.markdown('''

**Sentence Transformers** is a set of frameworks & models that are trained to generate Embeddings from input sentences. 
Generated Sentence Embeddings can be used for Sentence Similarity / Asymmetric QA / Semantic Search / Clustering
among other tasks.

We trained multiple general-purpose Sentence Transformers models based on different LMs including 
distilroberta, mpnet and MiniLM-l6. They were trained using Siamese network configuration with custom **Contrastive Loss** 
inspired by OpenAI CLIP. The models were trained on a dataset comprising of [1 Billion+ training corpus](https://huggingface.co/flax-sentence-embeddings/all_datasets_v4_MiniLM-L6#training-data) with the v3 setup.

We have trained [20 models](https://huggingface.co/flax-sentence-embeddings) focused on general-purpose, QuestionAnswering and Code search and **achieved SOTA on multiple benchmarks.**
We also uploaded [8 datasets](https://huggingface.co/flax-sentence-embeddings) specialized for Question Answering, Sentence-Similiarity and Gender Evaluation.
You can view our models and datasets [here](https://huggingface.co/flax-sentence-embeddings).

''')

if menu == "Contributions & Evaluation":
    st.markdown('''    
## Contributions
    
- **20 Sentence Embedding models** that can be utilized for Sentence Simliarity / Asymmetric QA / Search & Clustering.
- **8 Datasets** from Stackexchange and StackOverflow, PAWS, Gender Evaluation uploaded to HuggingFace Hub.
- **Achieve SOTA** on multiple general purpose Sentence Similarity evaluation tasks by utilizing large TPU memory to maximize
 customized Contrastive Loss. [Full Evaluation here](https://docs.google.com/spreadsheets/d/1vXJrIg38cEaKjOG5y4I4PQwAQFUmCkohbViJ9zj_Emg/edit#gid=1809754143).
- **Gender Bias demonstration** that explores inherent bias in general purpose datasets.
- **Search / Clustering demonstration** that showcases real world use-cases for Sentence Embeddings.

## Model Evaluations

| Model     | [FullEvaluation](https://docs.google.com/spreadsheets/d/1vXJrIg38cEaKjOG5y4I4PQwAQFUmCkohbViJ9zj_Emg/edit#gid=1809754143) Average     | 20Newsgroups Clustering | StackOverflow DupQuestions | Twitter SemEval2015    |
|-----------|---------------------------------------|-------|-------|-------|
| paraphrase-mpnet-base-v2 (previous SOTA)  | 67.97 | 47.79 | 49.03 | 72.36 |
| **all_datasets_v3_roberta-large (400k steps)** | **70.22** | **50.12** | **52.18** | **75.28** |
| **all_datasets_v3_mpnet-base (440k steps)** | **70.01**   | **50.22** | **52.24** | **76.27** |
''')
elif menu == "Sentence Similarity":
    st.header('Sentence Similarity')
    st.markdown('''
**Instructions**: You can compare the similarity of the main text with other texts of your choice. In the background, 
we'll create an embedding for each text, and then we'll use the cosine similarity function to calculate a similarity 
metric between our main sentence and the others.

For more cool information on sentence embeddings, see the [sBert project](https://www.sbert.net/examples/applications/computing-embeddings/README.html).
''')
    select_models = st.multiselect("Choose models", options=list(MODELS_ID), default=list(MODELS_ID))

    anchor = st.text_input(
        'Please enter here the main text you want to compare:',
        value="That is a happy person"
    )

    n_texts = st.number_input(
        f'''How many texts you want to compare with: '{anchor}'?''',
        value=3,
        min_value=2)

    inputs = []
    
    defaults = ["That is a happy dog", "That is a very happy person", "Today is a sunny day"]
    for i in range(int(n_texts)):
        input = st.text_input(f'Text {i + 1}:', value=defaults[i] if i < len(defaults) else "")

        inputs.append(input)

    if st.button('Tell me the similarity.'):
        results = {model: inference.text_similarity(anchor, inputs, model, MODELS_ID) for model in select_models}
        df_results = {model: results[model] for model in results}

        index = [f"{idx + 1}:{input[:min(15, len(input))]}..." for idx, input in enumerate(inputs)]
        df_total = pd.DataFrame(index=index)
        for key, value in df_results.items():
            df_total[key] = [ts.item() for ts in torch.nn.functional.softmax(torch.from_numpy(value['score'].values))]

        st.write('Here are the results for selected models:')
        st.write(df_total)
        st.write('Visualize the results of each model:')
        st.line_chart(df_total)
elif menu == "Asymmetric QA":
    st.header('Asymmetric QA')
    st.markdown('''
**Instructions**: You can compare the Answer likeliness of a given Query with answer candidates of your choice. In the 
background, we'll create an embedding for each answer, and then we'll use the cosine similarity function to calculate a 
similarity metric between our query sentence and the others.
`mpnet_asymmetric_qa` model works best for hard-negative answers or distinguishing answer candidates that are actually questions 
due to separate models applied for encoding questions and answers.

For more cool information on sentence embeddings, see the [sBert project](https://www.sbert.net/examples/applications/computing-embeddings/README.html).
''')

    select_models = st.multiselect("Choose models", options=list(QA_MODELS_ID), default=list(QA_MODELS_ID)[0])

    anchor = st.text_input(
        'Please enter here the query you want to compare with given answers:',
        value="What is the weather in Paris?"
    )

    n_texts = st.number_input(
        f'''How many answers you want to compare with: '{anchor}'?''',
        value=3,
        min_value=2)

    inputs = []

    defaults = ["It is raining in Paris right now with 70 F temperature.", "What is the weather in Berlin?", "I have 3 brothers."]
    for i in range(int(n_texts)):
        input = st.text_input(f'Answer {i + 1}:', value=defaults[i] if i < len(defaults) else "")

        inputs.append(input)

    if st.button('Tell me Answer likeliness.'):
        results = {model: inference.text_similarity(anchor, inputs, model, QA_MODELS_ID) for model in select_models}
        df_results = {model: results[model] for model in results}

        index = [f"{idx + 1}:{input[:min(15, len(input))]}..." for idx, input in enumerate(inputs)]
        df_total = pd.DataFrame(index=index)
        for key, value in df_results.items():
            df_total[key] = [ts.item() for ts in torch.nn.functional.softmax(torch.from_numpy(value['score'].values))]

        st.write('Here are the results for selected models:')
        st.write(df_total)
        st.write('Visualize the results of each model:')
        st.line_chart(df_total)

elif menu == "Search / Cluster":
    st.header('Search / Cluster')
    st.markdown('''
**Instructions**: Make a query for anything related to "Python" and the model will return you nearby answers via dot-product.

For more cool information on sentence embeddings, see the [sBert project](https://www.sbert.net/examples/applications/computing-embeddings/README.html).
''')

    select_models = st.multiselect("Choose models", options=list(SEARCH_MODELS_ID), default=list(SEARCH_MODELS_ID)[0])

    anchor = st.text_input(
        'Please enter here your query about "Python", we will look for similar ones:',
        value="How do I sort a dataframe by column"
    )

    n_texts = st.number_input(
        f'''How many similar queries you want?''',
        value=5,
        min_value=2)

    if st.button('Give me my search.'):
        results = {model: inference.text_search(anchor, n_texts, model, QA_MODELS_ID) for model in select_models}
        st.table(pd.DataFrame(results[select_models[0]]).T)

    if st.button('3D Clustering of 1000 search results using T-SNE on generated embeddings'):
        st.write("Currently only works at local due to Spaces / plotly integration.")
        st.write("Demonstration : https://gyazo.com/1ff0aa438ae533de3b3c63382af7fe80")
        # fig = inference.text_cluster(anchor, 1000, select_models[0], QA_MODELS_ID)
        # fig.show()

elif menu == "Gender Bias Evaluation":
    st.header("Gender Bias Evaluation")
    st.markdown('''
**Instructions**: Here we can observe **inherent gender bias** in training set via random sampling of the sentences.

Input 3 texts, one without any mention of gender for target occupation and 2 others with gendered pronouns. 

Hopefully the evaluation performed here can proceed towards improving Gender-neutrality of datasets.

For more cool information on sentence embeddings, see the [sBert project](https://www.sbert.net/examples/applications/computing-embeddings/README.html).
''')

    select_models = st.multiselect("Choose models", options=list(MODELS_ID), default=list(MODELS_ID))

    samples = st.radio("Samples", options=["President of United States", "Professor", "Nurse", "Custom"])

    if samples == "President of United States":
        base_text = st.text_input("Gender Neutral Text", "President of the United States promised relief to Hurricane survivors.")
        male_text = st.text_input("Male-assumed Text", "He promised relief to Hurricane survivors.")
        female_text = st.text_input("Female-assumed Text", "She promised relief to Hurricane survivors.")
    elif samples == "Professor":
        base_text = st.text_input("Gender Neutral Text", "Professor ended the class earlier than usual.")
        male_text = st.text_input("Male-assumed Text", "He ended the class earlier than usual.")
        female_text = st.text_input("Female-assumed Text", "She ended the class earlier than usual.")
    elif samples == "Nurse":
        base_text = st.text_input("Gender Neutral Text", "Nurse administered the vaccine and rubbed alcohol.")
        male_text = st.text_input("Male-assumed Text", "He administered the vaccine and rubbed alcohol.")
        female_text = st.text_input("Female-assumed Text", "She administered the vaccine and rubbed alcohol.")
    else:
        base_text = st.text_input("Gender Neutral Text", "<Occupation> \"did something....\"")
        male_text = st.text_input("Male-assumed Text", "He \"did something....\"")
        female_text = st.text_input("Female-assumed Text", "She \"did something....\"")


    enter = st.button("Compare")
    if enter:
        results = {model: inference.text_similarity(base_text, [male_text, female_text], model, MODELS_ID) for model in select_models}

        index = ["male", "female", "gender_bias"]
        df_total = pd.DataFrame(index=index)
        for key, value in results.items():
            softmax = [round(ts.item(), 4) for ts in torch.nn.functional.softmax(torch.from_numpy(value['score'].values))]
            if softmax[0] > softmax[1]:
                gender = "male"
            elif abs(softmax[0] - softmax[1]) < 1e-3:
                gender = "neutral"
            else:
                gender = "female"
            softmax.append(gender)
            df_total[key] = softmax

        st.write('Here are the results for selected models:')
        st.write(df_total)