app.py

"""
Gradio requires input to be fed in a very peculiar way and does not provide too much flexibility - don't expect from this demo too much. The backbone had to be adjusted to work on hugging face spaces. Go see https://github.com/PiotrAntoniak/QuestionAnswering for a prettier version utilizing streamlit. 
"""


import gradio as gr
 
description = """Do you have a long document and a bunch of questions that can be answered given the data in this file? 
Fear not for this demo is for you. 
Upload your pdf, ask your questions and wait for the magic to happen.
DISCLAIMER: I do no have idea what happens to the pdfs that you upload and who has access to them so make sure there is nothing confidential there. 
""" 
title = "QA answering from a pdf."

from datetime import datetime
import numpy as np
import time
import hashlib
import torch
from transformers import AutoTokenizer, AutoModel, AutoModelForQuestionAnswering, pipeline
from tqdm import tqdm
import os
device = "cuda:0" if torch.cuda.is_available() else "cpu"
import textract
from scipy.special import softmax
import pandas as pd
from datetime import datetime


tokenizer = AutoTokenizer.from_pretrained("sentence-transformers/multi-qa-mpnet-base-dot-v1")
model = AutoModel.from_pretrained("sentence-transformers/multi-qa-mpnet-base-dot-v1").to(device).eval()
tokenizer_ans = AutoTokenizer.from_pretrained("deepset/roberta-large-squad2")
model_ans = AutoModelForQuestionAnswering.from_pretrained("deepset/roberta-large-squad2").to(device).eval()

if device == 'cuda:0':
    pipe = pipeline("question-answering",model_ans,tokenizer =tokenizer_ans,device = 0)
else:
    pipe = pipeline("question-answering",model_ans,tokenizer =tokenizer_ans)
    
def cls_pooling(model_output):
    return model_output.last_hidden_state[:,0]

def encode_query(query):
    encoded_input = tokenizer(query, truncation=True, return_tensors='pt').to(device)

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

    embeddings = cls_pooling(model_output)

    return embeddings.cpu()


def encode_docs(docs,maxlen = 64, stride = 32):
    encoded_input = []
    embeddings = []
    spans = []
    file_names = []
    name, text = docs
    temp_text = ""
    text = text.split(" ")
    if len(text) < maxlen:
        text = " ".join(text)
        
        encoded_input.append(tokenizer(temp_text, return_tensors='pt', truncation = True).to(device))
        spans.append(temp_text)
        file_names.append(name)

    else:
        num_iters = int(len(text)/maxlen)+1
        for i in range(num_iters):
            if i == 0:
                temp_text = " ".join(text[i*maxlen:(i+1)*maxlen+stride])
            else:
                temp_text = " ".join(text[(i-1)*maxlen:(i)*maxlen][-stride:] + text[i*maxlen:(i+1)*maxlen])

            encoded_input.append(tokenizer(temp_text, return_tensors='pt', truncation = True).to(device))
            spans.append(temp_text)
            file_names.append(name)

    with torch.no_grad():
        for encoded in tqdm(encoded_input): 
            model_output = model(**encoded, return_dict=True)
            embeddings.append(cls_pooling(model_output))
    
    embeddings = np.float32(torch.stack(embeddings).transpose(0, 1).cpu())
    
    np.save("emb_{}.npy".format(name),dict(zip(list(range(len(embeddings))),embeddings))) 
    np.save("spans_{}.npy".format(name),dict(zip(list(range(len(spans))),spans)))
    np.save("file_{}.npy".format(name),dict(zip(list(range(len(file_names))),file_names)))
    
    return embeddings, spans, file_names
   
def predict(query,data):
    print(datetime.today().strftime('%Y-%m-%d %H:%M:%S'))
    name_to_save = data.name.split("/")[-1].split(".")[0][:-8]
    k=20
    st = str([query,name_to_save])
    st_hashed = str(hashlib.sha256(st.encode()).hexdigest()) #just to speed up examples load
    hist = st + " " + st_hashed 
    now = datetime.now()
    current_time = now.strftime("%H:%M:%S")
    
    try: #if the same question was already asked for this document, upload question and answer
        df = pd.read_csv("{}.csv".format(hash(st)))
        list_outputs = []
        for i in range(k):
            temp = [df.iloc[n] for n in range(k)][i]
            text = ''
            text += 'PROBABILITIES: '+ temp.Probabilities + '\n\n' 
            text += 'ANSWER: ' +temp.Answer + '\n\n' 
            text += 'PASSAGE: '+temp.Passage + '\n\n' 
            list_outputs.append(text)
        return list_outputs
    except Exception as e:
        print(e)
        print(st)

    if name_to_save+".txt" in os.listdir(): #if the document was already used, load its embeddings
        doc_emb = np.load('emb_{}.npy'.format(name_to_save),allow_pickle='TRUE').item()
        doc_text = np.load('spans_{}.npy'.format(name_to_save),allow_pickle='TRUE').item()
        file_names_dicto = np.load('file_{}.npy'.format(name_to_save),allow_pickle='TRUE').item()
        
        doc_emb = np.array(list(doc_emb.values())).reshape(-1,768)
        doc_text = list(doc_text.values())
        file_names = list(file_names_dicto.values())
    
    else:
        text = textract.process("{}".format(data.name)).decode('utf8')
        text = text.replace("\r", " ")
        text = text.replace("\n", " ")
        text = text.replace(" . "," ")
        
        doc_emb, doc_text, file_names = encode_docs((name_to_save,text),maxlen = 64, stride = 32)
        
        doc_emb = doc_emb.reshape(-1, 768)
        with open("{}.txt".format(name_to_save),"w",encoding="utf-8") as f:
            f.write(text)
    
    #once embeddings are calculated, run MIPS
    start = time.time()
    query_emb = encode_query(query)
    
    scores = np.matmul(query_emb, doc_emb.transpose(1,0))[0].tolist()
    doc_score_pairs = list(zip(doc_text, scores, file_names))
    doc_score_pairs = sorted(doc_score_pairs, key=lambda x: x[1], reverse=True)
    
    probs_sum = 0
    probs = softmax(sorted(scores,reverse = True)[:k])
    table = {"Passage":[],"Answer":[],"Probabilities":[]}
    
    
    #get answers for each pair of question (from user) and top best passages
    for i, (passage, _, names) in enumerate(doc_score_pairs[:k]):
        passage = passage.replace("\n","")
        #passage = passage.replace(" . "," ")
        
        if probs[i] > 0.1 or (i < 3 and probs[i] > 0.05): #generate answers for more likely passages but no less than 2
            QA = {'question':query,'context':passage}
            ans = pipe(QA)
            probabilities = "P(a|p): {}, P(a|p,q): {}, P(p|q): {}".format(round(ans["score"],5), 
                                                                          round(ans["score"]*probs[i],5), 
                                                                          round(probs[i],5))
            table["Passage"].append(passage)
            table["Answer"].append(str(ans["answer"]).upper())
            table["Probabilities"].append(probabilities)
        else:
            table["Passage"].append(passage)
            table["Answer"].append("no_answer_calculated")
            table["Probabilities"].append("P(p|q): {}".format(round(probs[i],5)))
            
        
    #format answers for ~nice output and save it for future (if the same question is asked again using same pdf)
    df = pd.DataFrame(table)
    print(df)
    print("time: "+ str(time.time()-start))
    
    
    with open("HISTORY.txt","a", encoding = "utf-8") as f:
        f.write(hist)
        f.write(" " + str(current_time))
        f.write("\n")
        f.close()
    df.to_csv("{}.csv".format(hash(st)), index=False)
    
    list_outputs = []
    for i in range(k):
        text = ''
        temp = [df.iloc[n] for n in range(k)][i]
        text += 'PROBABILITIES: '+ temp.Probabilities + '\n\n' 
        text += 'ANSWER: ' +temp.Answer + '\n\n' 
        text += 'PASSAGE: '+temp.Passage + '\n\n' 
  
        list_outputs.append(text)
    
    return list_outputs



iface = gr.Interface(examples = [
        ["How high is the highest mountain?","China.pdf"], 
        ["Where is the highest  mountain?","China.pdf"]
    ],
    
    fn =predict,
    inputs = [gr.Textbox(),
              gr.File(),
    ],
    outputs = 'text',
    description=description,
    title = title
                     )

iface.launch()