app.py

from transformers import ViTImageProcessor, ViTForImageClassification
import gradio as gr
from datasets import load_dataset
import torch
import random
import numpy as np
import pandas as pd
# tabulate in dependencies



def get_predictions(image):
    inputs = processor(image, return_tensors="pt")
    with torch.no_grad():
        outputs = model(**inputs)
        logits = outputs.logits

    # Get top n predictions
    top_indices = logits[0].argsort(dim=-1, descending=True)
    probabilities = torch.softmax(logits, dim=-1)[0, top_indices]
    labels = [model.config.id2label[idx.item()] for idx in top_indices]

    predictions = {}
    for i, label in enumerate(labels):
        predictions[label] = probabilities[i]

    return predictions

data = load_dataset("marcelomoreno26/geoguessr",split="test")
model_name = "marcelomoreno26/vit-base-patch-16-384-geoguessr"

# Initialize the feature extractor
processor = ViTImageProcessor.from_pretrained(model_name)
# Initialize the model
model = ViTForImageClassification.from_pretrained(model_name)


length = len(data)
countries = []

with open("countries.txt", "r") as file:
    for line in file:
        countries.append(line.strip())


def get_result(selection):
    global correct_country
    global model_prediction
    global filtered_predictions
    if selection == correct_country and correct_country == model_prediction:
        result = "It's a draw!"
    elif selection == correct_country:
        result = "Congratulations! You won!"
    elif correct_country == model_prediction:
        result = "Sorry, you lost. The AI guessed it right!"
    else:
        result = "Sorry, you both lost."

    total_prob = sum([(float(value)) for value in filtered_predictions.values()])
    prob_per_country = [(key,np.round(float(value)/total_prob,3)*100) for key,value in filtered_predictions.items()]
    df = pd.DataFrame(prob_per_country,columns=["Country","Model Confidence (%)"]).sort_values(by="Model Confidence (%)",ascending=False)
    ai_confidence = f"The AI's guess was {model_prediction}\n\nAI's Results:\n"+ df.to_markdown(index=False)



    return f"The correct country was: {correct_country}\n{result}", ai_confidence


def load():
    global filtered_predictions
    # Randomly select an image
    i = random.randint(0, len(data) - 1)
    image = data[i]['image']
    correct_country = data[i]['label']

    # Randomly sample 4 countries as options
    options = [country for country in random.sample(countries, 4) if country != correct_country]
    options.append(correct_country)
    random.shuffle(options)

    # Get model predictions
    predictions = get_predictions(image)
    filtered_predictions = {country: predictions[country] for country in options}
    model_prediction = max(filtered_predictions, key=filtered_predictions.get)

    return image, options, correct_country, model_prediction


def reload():
    global correct_country
    global model_prediction
    global filtered_predictions
    # Randomly select an image
    i = random.randint(0, len(data) - 1)
    image = data[i]['image']
    correct_country = data[i]['label']

    # Randomly sample 4 countries as options
    options = [country for country in random.sample(countries, 4) if country != correct_country]
    options.append(correct_country)
    random.shuffle(options)

    # Get model predictions
    predictions = get_predictions(image)
    filtered_predictions = {country: predictions[country] for country in options}
    model_prediction = max(filtered_predictions, key=filtered_predictions.get)


    return gr.Image(image), gr.Radio(choices=options, label ="Select the country:"),  "", ""



with gr.Blocks() as demo:
    
    image, options, correct_country, model_prediction = load()

    gr.Markdown("# GeoGuessr - Can You Beat the AI?")
    gr.Markdown("Try to guess the country in the image. Can you beat the AI?")
    img = gr.Image(image)
    radio = gr.Radio(choices=options, label ="Select the country:")
    ai_pred = gr.Markdown()
    text = gr.Text(label="Result")
    radio.select(fn=get_result, inputs=radio, outputs=[text,ai_pred])

    btn = gr.Button(value="Get New Image")
    btn.click(reload, None,outputs=[img,radio,text,ai_pred])

demo.launch()