game3.py

import requests
import random
import time
import pandas as pd
import gradio as gr
import numpy as np
from transformers import AutoTokenizer, AutoModelForSequenceClassification
from transformers import pipeline
import torch

def read3(num_selected_former):
    fname = 'data3_convai2_inferred.txt'
    with open(fname, encoding='utf-8') as f:
        content = f.readlines()
        index_selected = random.randint(0,len(content)/2-1)
        while index_selected == num_selected_former:
            index_selected = random.randint(0,len(content)/2-1)
        text = eval(content[index_selected*2])
        interpretation = eval(content[int(index_selected*2+1)])
        
        min_len = 5

        tokens = [i[0] for i in interpretation]
        tokens = tokens[1:-1]
        while len(tokens) <= min_len or '\\' in text['text'] or '//' in text['text']:
            index_selected = random.randint(0,len(content)/2-1)
            text = eval(content[int(index_selected*2)])
        res_tmp = [(i, 0) for i in text['text'].split(' ')]
        res = {"original": text['text'], "interpretation": res_tmp}    
    return res, index_selected
    
def func3(num_selected, human_predict, num1, num2, user_important):
    chatbot = []
    # num1: Human score; num2: AI score
    fname = 'data3_convai2_inferred.txt'

    with open(fname) as f:
        content = f.readlines()
        text = eval(content[int(num_selected*2)])
        interpretation = eval(content[int(num_selected*2+1)])
        
        if text['binary_label'] == 1:
            golden_label = int(5 * (1 - text['binary_score']))
        else:
            golden_label = int(5 * (1 + text['binary_score']))
            
    # (START) off-the-shelf version -- slow at the beginning
    # Load model directly
    # Use a pipeline as a high-level helper

    device = torch.device("cuda") if torch.cuda.is_available() else torch.device("cpu")
    classifier = pipeline("text-classification", model="padmajabfrl/Gender-Classification", device=device)
    output = classifier([text['text']])

    print(output)
    out = output[0]
    
    # (END) off-the-shelf version
    
    if out['label'] == 'Female':
        ai_predict = int(10 * out['score'])
    else:
        ai_predict = 10 - int(10 * out['score'])
    
    user_select = "You focused on "
    flag_select = False
    if user_important == "":
        user_select += "nothing. Interesting! "
    else:
        user_select += "'" + user_important + "'. "
    # for i in range(len(user_marks)):
    #     if user_marks[i][1] != None and h1[i][0] not in ["P", "N"]:
    #         flag_select = True
    #         user_select += "'" + h1[i][0] + "'"
    #         if i == len(h1) - 1:
    #             user_select += ". "
    #         else:
    #             user_select += ", "
    # if not flag_select:
    #     user_select += "nothing. Interesting! "
    user_select += "Wanna see how the AI made the guess? Click here. ⬅️"
    
    if golden_label > 6:
        gender = ' (female)'
    elif golden_label < 4:
        gender = ' (male)'    
    else:
        gender = ' (neutral)'
        
    if abs(golden_label - human_predict) <= 2  and abs(golden_label - ai_predict) <= 2:
        chatbot.append(("The correct answer is " + str(golden_label) + gender + ". Congratulations! 🎉 Both of you get the correct answer!", user_select))
        num1 += 1
        num2 += 1
    elif abs(golden_label - human_predict) > 2 and abs(golden_label - ai_predict) > 2:
        chatbot.append(("The correct answer is " + str(golden_label) + gender + ". Sorry.. No one gets the correct answer. But nice try! 😉", user_select))
    elif abs(golden_label - human_predict) <= 2 and abs(golden_label - ai_predict) > 2:
        chatbot.append(("The correct answer is " + str(golden_label) + gender + ". Great! 🎉 You are closer to the answer and better than AI!", user_select))
        num1 += 1
    else:
        chatbot.append(("The correct answer is " + str(golden_label) + gender + ". Sorry.. AI wins in this round.", user_select))
        num2 += 1
    
    # tot_scores = ''' ### <p style="text-align: center;"> 🤖 Machine &ensp; ''' + str(int(num2)) + ''' &ensp; VS &ensp; ''' + str(int(num1)) + ''' &ensp; Human 👨👩 </p>'''
    
    # tot_scores = ''' #### <p style="text-align: center;"> Today's Scores:</p>
    #                 #### <p style="text-align: center;"> 🤖 Machine &ensp; <span style="color: red;">''' + str(int(num2)) + '''</span> &ensp; VS &ensp; <span style="color: red;">''' + str(int(num1)) + '''</span> &ensp; Human 🙋 </p>'''
    tot_scores = ''' #### <p style="text-align: center;"> Today's Scores: &ensp; &ensp; 🤖 Machine &ensp; <span style="color: red;">''' + str(int(num2)) + '''</span> &ensp; VS &ensp; <span style="color: red;">''' + str(int(num1)) + '''</span> &ensp; Human 🙋 </p>'''

    return ai_predict, chatbot, num1, num2, tot_scores

def interpre3(num_selected):
    fname = 'data3_convai2_inferred.txt'
    tokenizer = AutoTokenizer.from_pretrained("padmajabfrl/Gender-Classification")

    with open(fname) as f:
        content = f.readlines()
        text = eval(content[int(num_selected*2)])
        interpretation = eval(content[int(num_selected*2+1)])
    
    print(interpretation)

    encodings = tokenizer(text['text'], return_offsets_mapping=True)

    print(encodings['offset_mapping'])
    is_subword = [False, False]
    for i in range(2, len(encodings['offset_mapping'])):
        if encodings['offset_mapping'][i][0] == encodings['offset_mapping'][i-1][1]:
            is_subword.append(True)
        else:
            is_subword.append(False)
    print(is_subword)
    interpretation_combined = []

    index_tmp = 0
    while index_tmp < (len(interpretation) - 1):
        if not is_subword[index_tmp+1]:
            interpretation_combined.append(interpretation[index_tmp])
            index_tmp += 1
        else:
            text_combined = interpretation[index_tmp][0]
            score_combinded = interpretation[index_tmp][1]
            length = 1
            while is_subword[index_tmp+length]:
                text_combined += interpretation[index_tmp+length][0]
                score_combinded += interpretation[index_tmp+length][1]
                length += 1
            interpretation_combined.append((text_combined, score_combinded/length))
            index_tmp += length

    interpretation_combined.append(('', 0.0))
    print(interpretation_combined)
    
    res = {"original": text['text'], "interpretation": interpretation_combined}
    # pos = []
    # neg = []
    # res = []
    # for i in interpretation:
    #     if i[1] > 0:
    #         pos.append(i[1])
    #     elif i[1] < 0:
    #         neg.append(i[1])
    #     else:
    #         continue
    # median_pos = np.median(pos)
    # median_neg = np.median(neg)


    # res.append(("P", "+"))
    # res.append(("/", None))
    # res.append(("N", "-"))
    # res.append(("Review:", None))
    # for i in interpretation:
    #     if i[1] > median_pos:
    #         res.append((i[0], "+"))
    #     elif i[1] < median_neg:
    #         res.append((i[0], "-"))
    #     else:
    #         res.append((i[0], None))
    return res

    
def func3_written(text_written, human_predict, lang_written):
    chatbot = []
    # num1: Human score; num2: AI score

    # (START) off-the-shelf version

    # tokenizer = AutoTokenizer.from_pretrained("nlptown/bert-base-multilingual-uncased-sentiment")
    # model = AutoModelForSequenceClassification.from_pretrained("nlptown/bert-base-multilingual-uncased-sentiment")

    device = torch.device("cuda") if torch.cuda.is_available() else torch.device("cpu")
    classifier = pipeline("text-classification", model="padmajabfrl/Gender-Classification", device=device)
    tokenizer = AutoTokenizer.from_pretrained("padmajabfrl/Gender-Classification")

    output = classifier([text_written])

    print(output)
    out = output[0]
    # (END) off-the-shelf version

    if out['label'] == 'Female':
        ai_predict = int(10 * out['score'])
    else:
        ai_predict = 10 - int(10 * out['score'])
    
    if abs(ai_predict - human_predict) <= 2:
        chatbot.append(("AI gives it a close score! 🎉", "⬅️ Feel free to try another one! This time let’s see if you can trick the AI into giving a wrong rating. ⬅️"))
    else:
        chatbot.append(("AI thinks in a different way from human. 😉", "⬅️ Feel free to try another one! ⬅️"))

    import shap

    gender_classifier = pipeline("text-classification", model="padmajabfrl/Gender-Classification", return_all_scores=True, device=device)

    explainer = shap.Explainer(gender_classifier)

    shap_values = explainer([text_written])
    interpretation = list(zip(shap_values.data[0], shap_values.values[0, :, 1]))

    
    encodings = tokenizer(text['text'], return_offsets_mapping=True)

    print(encodings['offset_mapping'])
    is_subword = [False, False]
    for i in range(2, len(encodings['offset_mapping'])):
        if encodings['offset_mapping'][i][0] == encodings['offset_mapping'][i-1][1]:
            is_subword.append(True)
        else:
            is_subword.append(False)
    print(is_subword)
    interpretation_combined = []

    index_tmp = 0
    while index_tmp < (len(interpretation) - 1):
        if not is_subword[index_tmp+1]:
            interpretation_combined.append(interpretation[index_tmp])
            index_tmp += 1
        else:
            text_combined = interpretation[index_tmp][0]
            score_combinded = interpretation[index_tmp][1]
            length = 1
            while is_subword[index_tmp+length]:
                text_combined += interpretation[index_tmp+length][0]
                score_combinded += interpretation[index_tmp+length][1]
                length += 1
            interpretation_combined.append((text_combined, score_combinded/length))
            index_tmp += length

    interpretation_combined.append(('', 0.0))
    print(interpretation_combined)

    res = {"original": text_written, "interpretation": interpretation_combined}
    print(res)

    return res, ai_predict, chatbot