app.py

import json
import gradio as gr
import numpy as np
import time
import csv
import json
import os
import random
import string
import sys
import time
import gradio as gr
import numpy as np
import pandas as pd
from huggingface_hub import (
    CommitScheduler,
    HfApi,
    InferenceClient,
    login,
    snapshot_download,
)
from PIL import Image
from utils import string_to_image
import matplotlib.backends.backend_agg as agg
import math
from pathlib import Path

np.random.seed(int(time.time()))
csv.field_size_limit(sys.maxsize)
np.random.seed(int(time.time()))


NUMBER_OF_IMAGES = 30
intro_screen = Image.open("./images/intro.jpg")

meta_top1 = json.load(open("./data/top1-exp-materials/metadata.json"))
meta_topK = json.load(open("./data/topK-exp-materials/metadata.json"))

all_data = {}
all_data["top1"] = meta_top1
all_data["topK"] = meta_topK


# for data in  all_data["top1"] and all_data["topK"] add a key to show which type they are
for k in all_data["top1"].keys():
    all_data["top1"][k]["type"] = "top1"

for k in all_data["topK"].keys():
    all_data["topK"][k]["type"] = "topK"


###############################################################################################################
session_token = os.environ.get("SessionToken")
login(token=session_token)

REPO_URL = "xairesearch2023-advnet/HumanStudyData"
JSON_DATASET_DIR = Path("responses")

################################################################################################################

scheduler = CommitScheduler(
    repo_id=REPO_URL,
    repo_type="dataset",
    folder_path=JSON_DATASET_DIR,
    path_in_repo="./data",
    every=1,
    private=True,
)


if not JSON_DATASET_DIR.exists():
    JSON_DATASET_DIR.mkdir()


def generate_data(type_of_nns):
    global NUMBER_OF_IMAGES
    
    if type_of_nns == "topQ":
        type_of_nns = "topK"
    
    # randomly pick NUMBER_OF_IMAGES from the dataset with type type_of_nns
    keys = list(all_data[type_of_nns].keys())
    sample_data = random.sample(keys, NUMBER_OF_IMAGES)

    data = []
    for k in sample_data:
        new_datapoint = all_data[type_of_nns][k]
        new_datapoint["image-path"] = f"./data/{type_of_nns}-exp-materials/{k}.jpeg"
        data.append(new_datapoint)

    return data


def load_sample(data, current_index):
    current_datapoint = data[current_index]

    image_path = current_datapoint["image-path"]
    image = Image.open(image_path)
    top_1 = current_datapoint["top1-label"]
    top_1_score = current_datapoint["top1-score"]

    q_template = (
        "<div style='font-size: 24px;'>Sam guessed the Input image is "
        "<span style='font-weight: bold;'>{}</span> "
        "with <span style='font-weight: bold;'>{}%</span> "
        "confidence. Is this bird a <span style='font-weight: bold;'>{}</span>?"
        "</div>"
    )

    q_template = (
        "<div style='font-size: 24px;'>Sam guessed the Input image is "
        "<span style='font-weight: bold;'>{}</span> "
        "with <span style='font-weight: bold;'>{}%</span> "
        "confidence.<br>Is this bird a <span style='font-weight: bold;'>{}</span>?"
        "</div>"
    )

    top_1_score = top_1_score * 100
    top_1_score = round(top_1_score, 2)

    rounded_up_score = math.ceil(top_1_score)
    rounded_up_score = int(rounded_up_score)
    question = q_template.format(top_1, str(rounded_up_score), top_1)

    accept_reject = current_datapoint["Accept/Reject"]

    return image, top_1, rounded_up_score, question, accept_reject


def preprocessing(data, type_of_nns, current_index, history, username):
    print("preprocessing")
    data = generate_data(type_of_nns)
    print("data generated")

    # append a random text to the username
    random_text = "".join(
        random.choice(string.ascii_lowercase + string.digits) for _ in range(8)
    )

    if username == "":
        username = "username"

    username = f"{username}-{random_text}"

    current_index = 0
    print("loading sample ....")
    qimage, top_1, top_1_score, question, accept_reject = load_sample(
        data, current_index
    )

    return (
        qimage,
        top_1,
        top_1_score,
        question,
        accept_reject,
        current_index,
        history,
        data,
        username,
    )


def update_app(decision, data, current_index, history, username):
    global NUMBER_OF_IMAGES
    if current_index == -1:
        gr.Error("Please Enter your username and load samples")

        fake_plot = string_to_image("Please Enter your username and load samples")
        canvas = agg.FigureCanvasAgg(fake_plot)
        canvas.draw()
        empty_image = Image.frombytes(
            "RGBA", canvas.get_width_height(), canvas.tostring_argb()
        )

        return (
            empty_image,
            "",
            "",
            "",
            "",
            current_index,
            history,
            data,
            0,
            gr.update(interactive=False),
            gr.update(interactive=False),
            "",
        )

    # Done, let's save and upload
    if current_index == NUMBER_OF_IMAGES - 1:
        time_stamp = int(time.time())

        # Add decision to the history
        current_dicitonary = data[current_index].copy()
        current_dicitonary["user_decision"] = decision
        current_dicitonary["user_id"] = username
        accept_reject_string = "Accept" if decision == "YES" else "Reject"
        current_dicitonary["is_user_correct"] = (
            current_dicitonary["Accept/Reject"] == accept_reject_string
        )
        history.append(current_dicitonary)

        # convert to percentage
        final_decision_data = {
            "user_id": username,
            "time": time_stamp,
            "history": history,
        }

        # upload the decision to the server
        temp_filename = f"./responses/results_{username}.json"
        # convert decision_dict to json and save it on the disk
        with open(temp_filename, "w") as f:
            json.dump(final_decision_data, f)

        fake_plot = string_to_image("Thank you for your time!")
        canvas = agg.FigureCanvasAgg(fake_plot)
        canvas.draw()
        empty_image = Image.frombytes(
            "RGBA", canvas.get_width_height(), canvas.tostring_argb()
        )

        # TODO, Call the accuracy and show it to the user
        # calcualte the mean of is_user_correct
        all_is_user_correct = [d["is_user_correct"] for d in history]
        accuracy = np.mean(all_is_user_correct) * 100   
        accuracy = round(accuracy, 2)

        return (
            empty_image,
            "",
            "",
            "",
            "",
            current_index,
            history,
            data,
            current_index + 1,
            gr.update(interactive=False),
            gr.update(interactive=False),
            f"User Accuracy: {accuracy}",
        )

    if current_index >= 0 and current_index < NUMBER_OF_IMAGES - 1:
        current_dicitonary = data[current_index].copy()
        current_dicitonary["user_decision"] = decision
        current_dicitonary["user_id"] = username
        accept_reject_string = True if decision == "YES" else False
        current_dicitonary["is_user_correct"] = (
            current_dicitonary["Accept/Reject"] == accept_reject_string
        )

        print(f" accept/reject : {current_dicitonary['Accept/Reject'] }")
        print(
            f" accept/reject status: {current_dicitonary['Accept/Reject'] == accept_reject_string}"
        )

        history.append(current_dicitonary)

        current_index += 1
        qimage, top_1, top_1_score, question, accept_reject = load_sample(
            data, current_index
        )

        return (
            qimage,
            top_1,
            top_1_score,
            question,
            accept_reject,
            current_index,
            history,
            data,
            current_index,
            gr.update(interactive=True),
            gr.update(interactive=True),
            "",
        )


def disable_component():
    return gr.update(interactive=False)


def enable_component():
    return gr.update(interactive=True)


def hide_component():
    return gr.update(visible=False)


with gr.Blocks(theme=gr.themes.Soft()) as demo:
    data_state = gr.State({})
    current_index = gr.State(-1)
    history = gr.State([])

    gr.Markdown("# Advising Networks")
    gr.Markdown("## Accept/Reject AI predicted label using Explanations")

    with gr.Column():
        with gr.Row():
            username_textbox = gr.Textbox(label="Username", value=f"username")
            labeled_images_textbox = gr.Textbox(label="Labeled Images", value="0")
            total_images_textbox = gr.Textbox(
                label="Total Images", value=NUMBER_OF_IMAGES
            )
            type_of_nns_dropdown = gr.Dropdown(
                label="Type of NNs",
                choices=["top1", "topQ"],
                value="top1",
            )

        prepare_btn = gr.Button(value="Start The Experiment")

    with gr.Column():
        with gr.Row():
            question_textbox = gr.HTML("")
            # question_textbox = gr.Markdown("")

        with gr.Column(elem_id="parent_row"):
            query_image = gr.Image(
                type="pil", label="Query", show_label=False, value="./images/intro.jpg"
            )

        with gr.Row():
            accept_btn = gr.Button(value="YES", interactive=False)
            reject_btn = gr.Button(value="NO", interactive=False)

        with gr.Column(elem_id="parent_row"):
            top_1_textbox = gr.Textbox(label="Top 1", value="", visible=False)
            top_1_score_textbox = gr.Textbox(
                label="Top 1 Score", value="", visible=False
            )
            accept_reject_textbox = gr.Textbox(
                label="Accept/Reject", value="", visible=False
            )

    with gr.Column():
        with gr.Row():
            final_results = gr.HTML("")

    # data, type_of_nns, current_index, history
    prepare_btn.click(
        preprocessing,
        inputs=[
            data_state,
            type_of_nns_dropdown,
            current_index,
            history,
            username_textbox,
        ],
        outputs=[
            query_image,
            top_1_textbox,
            top_1_score_textbox,
            question_textbox,
            accept_reject_textbox,
            current_index,
            history,
            data_state,
            username_textbox,
        ],
    ).then(fn=disable_component, outputs=[prepare_btn]).then(
        fn=disable_component, outputs=[type_of_nns_dropdown]
    ).then(
        fn=disable_component, outputs=[username_textbox]
    ).then(
        fn=disable_component, outputs=[prepare_btn]
    ).then(
        fn=enable_component, outputs=[accept_btn]
    ).then(
        fn=enable_component, outputs=[reject_btn]
    ).then(
        fn=hide_component, outputs=[prepare_btn]
    )

    accept_btn.click(
        update_app,
        inputs=[accept_btn, data_state, current_index, history, username_textbox],
        outputs=[
            query_image,
            top_1_textbox,
            top_1_score_textbox,
            question_textbox,
            accept_reject_textbox,
            current_index,
            history,
            data_state,
            labeled_images_textbox,
            accept_btn,
            reject_btn,
            final_results,
        ],
    )

    reject_btn.click(
        update_app,
        inputs=[reject_btn, data_state, current_index, history, username_textbox],
        outputs=[
            query_image,
            top_1_textbox,
            top_1_score_textbox,
            question_textbox,
            accept_reject_textbox,
            current_index,
            history,
            data_state,
            labeled_images_textbox,
            accept_btn,
            reject_btn,
            final_results,
        ],
    )


demo.launch(debug=False, server_name="0.0.0.0")
# demo.launch(debug=False)