import gradio as gr
import os
import torch

from model import create_vit
from timeit import default_timer as timer
from typing import Tuple, Dict

class_names = ["NORMAL", "PNEUMONIA"]

vit_model, vit_transforms = create_vit(seed=42)

vit_model.load_state_dict(
    torch.load(
        f="finetuned_vit_b_16_pneumonia_feature_extractor.pth", 
        map_location=torch.device("cpu")
    )
)

def predict(img):
    start_timer = timer()
    
    img = vit_transforms(img).unsqueeze(0)
    
    vit_model.eval()
    with torch.inference_mode():
        pred_prob_int = torch.sigmoid(vit_model(img)).round().int().squeeze()
        
    if pred_prob_int.item() == 1:
        class_name = class_names[1]
    else:
        class_name = class_names[0]
            
    pred_time = round(timer() - start_timer, 5)
    
    return class_name, pred_time

title = "Detect Pneumonia from chest X-Ray"
description = "A ViT feature extractor Computer Vision model to detect Pneumonia from X-Ray Images."
article = "Access project repository at [GitHub](https://github.com/Ammar2k/pneumonia_detection)"

example_list = [["examples/" + example] for example in os.listdir("examples")]

demo = gr.Interface(fn=predict, 
                    inputs=gr.Image(type="pil"),
                    outputs=[gr.Label(num_top_classes=6, label="Predictions"), 
                    gr.Number(label="Prediction time(s)")],
                    examples=example_list,
                    title=title,
                    description=description,
                    article=article
                   )

demo.launch()