import gradio as gr
import numpy as np
from PIL import Image
from transformers import AutoImageProcessor, AutoModelForImageClassification

model_names = [
    "0-ma/swin-geometric-shapes-tiny",
    "0-ma/mobilenet-v2-geometric-shapes",
    "0-ma/focalnet-geometric-shapes-tiny",
    "0-ma/efficientnet-b2-geometric-shapes",
    "0-ma/beit-geometric-shapes-base",
    "0-ma/mit-b0-geometric-shapes",
    "0-ma/vit-geometric-shapes-base",
    "0-ma/resnet-geometric-shapes",
    "0-ma/vit-geometric-shapes-tiny",
]

example_images = [
    'example/1_None.jpg',
    'example/2_Circle.jpg',
    'example/3_Triangle.jpg',
    'example/4_Square.jpg',
    'example/5_Pentagone.jpg',
    'example/6_Hexagone.jpg'
]

labels = [example.split("_")[1].split(".")[0] for example in example_images]

feature_extractors = {model_name: AutoImageProcessor.from_pretrained(model_name) for model_name in model_names}
classification_models = {model_name: AutoModelForImageClassification.from_pretrained(model_name) for model_name in model_names}

def predict(image, selected_model):
    feature_extractor = feature_extractors[selected_model]
    model = classification_models[selected_model]

    inputs = feature_extractor(images=[image], return_tensors="pt")
    logits = model(**inputs)['logits'].cpu().detach().numpy()[0]
    logits_positive = logits
    logits_positive[logits < 0] = 0
    logits_positive = logits_positive/np.sum(logits_positive)

    confidences = {}
    for i in range(len(labels)):
        if logits[i] > 0:
            confidences[labels[i]] = float(logits_positive[i])
    return confidences

title = "Geometric Shape Classifier"
description = "Select a model and upload an image to classify geometric shapes."

with gr.Blocks() as demo:
    gr.Markdown(f"# {title}")
    gr.Markdown(description)
    
    with gr.Row():
        model_dropdown = gr.Dropdown(choices=model_names, label="Selected Model", value=model_names[0])
        gr.Examples(
            examples=example_images,
            inputs=image_input,
            outputs=output,
            fn=lambda img: predict(img, model_dropdown.value),
            cache_examples=True,
        )
    image_input = gr.Image(type="pil")
    
    output = gr.Label()

    image_input.change(fn=lambda img: predict(img, model_dropdown.value), inputs=[image_input], outputs=output)
    model_dropdown.change(fn=lambda img, model: predict(img, model), inputs=[image_input, model_dropdown], outputs=output)

demo.launch()