app.py

import gradio as gr
import timm
import torch
from cods.classif.cp import ClassificationConformalizer
from cods.classif.data import ClassificationDataset
from cods.classif.data.predictions import ClassificationPredictions
from cods.classif.models import ClassificationModel
from datasets import load_dataset

# from ultralytics import YOLO
from PIL import Image

# from transformers import AutoImageProcessor, AutoModelForImageClassification
from dataset import DatasetWrapper

DATASETS = {
    "miniimagenet": "timm/mini-imagenet",
    "imagenette": "frgfm/imagenette",
    "imagenet": "imagenet-1k",
}

MODELS = {
    "miniimagenet": [
        "QuentinJG/ResNet18-miniimagenet",
        "shahrukhx01/vit-base-patch16-miniimagenet",
    ],
}

classification_conformalizer = ClassificationConformalizer(method="lac", preprocess="softmax")


def calibrate(dataset_name, model_name):
    global model
    # processor = AutoImageProcessor.from_pretrained(model_name)
    # model = AutoModelForImageClassification.from_pretrained(model_name)
    # model = #lambda x: model(processor(x))#**processor(x, return_tensors="pt"))
    # model = timm.create_model(model_name, pretrained=True)#, num_classes=100)
    model_name = "resnet34"
    global pretrained_resnet_34
    pretrained_resnet_34 = timm.create_model(model_name, pretrained=True)
    classifier = ClassificationModel(model=pretrained_resnet_34, model_name=model_name)
    global dataset
    dataset = load_dataset(DATASETS[dataset_name], split="validation")
    dataset = DatasetWrapper(dataset)

    val_preds = classifier.build_predictions(
        dataset,
        dataset_name=dataset_name,
        split_name="cal",
        batch_size=512,
        shuffle=False,
    )
    classification_conformalizer.calibrate(val_preds, alpha=0.1)
    return f"Calibrated on {dataset_name} with model {model_name}"


def predict_image(img):
    img_old = img.copy()
    img = dataset.transforms(img).unsqueeze(0)
    pred = pretrained_resnet_34(img)
    inference_pred = ClassificationPredictions(
        dataset_name="uploaded",
        split_name="test",
        image_paths=[None],
        idx_to_cls=dataset.idx_to_cls,
        true_cls=torch.tensor([-1]),  # Placeholder for true class
        pred_cls=pred,  # Placeholder for predicted class probabilities
    )

    result = classification_conformalizer.conformalize(inference_pred)
    list_of_classes = [dataset.idx_to_cls[i] for i in result[0].detach().numpy()]
    result = f"Predicted classes with 90% confidence: {list_of_classes}"
    return img_old, result


# Load a pretrained YOLOv8n model
# model = YOLO("yolov8n.pt")


def main_function(lbd, img):
    # results = model(img)  # predict on an image
    # r = results[0]
    # im_bgr = r.plot()  # BGR-order numpy array
    # im_rgb = Image.fromarray(im_bgr[..., ::-1])  # RGB-order PIL image
    # new_img = im_rgb
    # res = results[0].save(filename="output.jpg")  # save the image
    # # load image
    # new_img = Image.open("output.jpg")
    new_img = img
    return new_img


with gr.Blocks() as demo:
    gr.Markdown("# Image Classification with Conformal Prediction")
    gr.Markdown("## Upload an image and get conformalized classification predictions.")

    with gr.Row():
        dataset_dropdown = gr.Dropdown(
            choices=DATASETS.keys(), label="Select Dataset", value=list(DATASETS.keys())[0]
        )
        model_dropdown = gr.Dropdown(
            choices=MODELS[dataset_dropdown.value],
            label="Select Model",
            value=MODELS[dataset_dropdown.value][0],
        )

    calibrate_btn = gr.Button("Calibrate")
    status_text = gr.Textbox(label="Status", interactive=False)

    gr.Markdown("---")

    with gr.Row():
        input_image = gr.Image(label="Upload Image", type="pil")
        output_image = gr.Image(label="Processed Image")

    predict_btn = gr.Button("Predict")
    result_text = gr.Textbox(label="Prediction Result")

    # Connect components
    calibrate_btn.click(
        fn=calibrate, inputs=[dataset_dropdown, model_dropdown], outputs=status_text
    )

    predict_btn.click(fn=predict_image, inputs=input_image, outputs=[output_image, result_text])

demo.launch()