import streamlit as st
import torch
from transformers import AutoTokenizer, AutoModel, AutoModelForSequenceClassification
from safetensors.torch import load_file as safe_load

target_to_ind = {'cs': 0, 'econ': 1, 'eess': 2, 'math': 3, 'phys': 4, 'q-bio': 5, 'q-fin': 6, 'stat': 7}
target_to_label = {'cs': 'Computer Science', 'econ': 'Economics', 'eess': 'Electrical Engineering and Systems Science', 'math': 'Mathematics', 'phys': 'Physics', 
                  'q-bio': 'Quantitative Biology', 'q-fin': 'Quantitative Finance', 'stat': 'Statistics'}
ind_to_target = {ind: target for target, ind in target_to_ind.items()}


![image/gif](https://cdn-uploads.huggingface.co/production/uploads/67f0e021e17e5360e7c18a52/oNcjUz7MbWW0Bymah31y7.gif)

st.title('papers_classifier  - your best friend in 🤓')
st.text("You need to give me paper's title and (if you want to) it's abstract. Also you need to choose classification mode - there are 2 of them:\
best prediction and top 95% which means that you'll see as many classes as model needs to show to be confident with probability at least 0.95 that the correct one is among them. \
After that you need to press the Get prediction button and I'll tell you to which fields of study this paper is related. \
According to arXiv there are 8 different fields - Computer Science, Economics, Electrical Engineering and Systems Science, Mathematics, Physics, Quantitative Biology, \
Quantitative Finance and Statistics.\
")

@st.cache_resource
def load_model_and_tokenizer():
    model_name = 'distilbert/distilbert-base-cased'
    tokenizer = AutoTokenizer.from_pretrained(model_name)
    model = AutoModelForSequenceClassification.from_pretrained(model_name, num_labels=len(target_to_ind))
    
    state_dict = safe_load("model.safetensors")
    model.load_state_dict(state_dict)
    
    return model, tokenizer


model, tokenizer = load_model_and_tokenizer()


def get_predict(title: str, abstract: str) -> (str, float, dict):
    text = [title + tokenizer.sep_token + abstract[:128]]

    tokens_info = tokenizer(
        text,
        padding=True,
        truncation=True,
        return_tensors="pt",
    )
    
    with torch.no_grad():
        out = model(**tokens_info)
        probs = torch.nn.functional.softmax(out.logits, dim=-1).tolist()[0]

        return list(sorted([(p, ind_to_target[i]) for i, p in enumerate(probs)]))[::-1]


title = st.text_area("Title ", "", height=100)
abstract = st.text_area("Abstract ", "", height=150)


mode = st.radio("Mode: ", ("Best prediction", "Top 95%"))

if st.button("Get prediction", key="manual"):
    if len(title) == 0:
        st.error("Please, provide paper's title")
    else:
        with st.spinner("Be patient, I'm doing my best"):
            predict = get_predict(title, abstract)

        tags = []
        threshold = 0 if mode == "Best prediction" else 0.95
        sum_p = 0
        for p, tag in predict:
            sum_p += p
            tags.append(target_to_label[tag])

            if sum_p >= threshold:
                break
        tags = '\n'.join(tags)
        st.success(tags)