from icevision.all import *
import PIL
import torch
from torchvision import transforms
import gradio as gr

# Load model
class_map = ClassMap(['selected_variant'])
backbone = faster_rcnn.backbones.resnet_fpn.resnet18(pretrained=True)
model = faster_rcnn.model(backbone=backbone, num_classes=len(class_map))
model.load_state_dict(torch.load('object_localization_full-ancestry.model.pth', map_location=torch.device('cpu')))
#model_type = models.torchvision.faster_rcnn

def predict(model, image, detection_threshold: float = 0.5):
    # Whenever you have images in memory (numpy arrays) you can use `Dataset.from_images`
    infer_ds = Dataset.from_images([image])

    batch, samples = faster_rcnn.build_infer_batch(infer_ds)
    preds = faster_rcnn.predict(
        model=model,
        batch=batch,
        detection_threshold=detection_threshold
    )
    return samples[0]["img"], preds[0]
    

def show_preds(input_image, detection_threshold=0.5):

    img, pred = predict(model=model, image=input_image, detection_threshold=detection_threshold)
    # print(pred)
    img = draw_pred(img=img, pred=pred, class_map=class_map, display_label=False, display_bbox=True)
    img = PIL.Image.fromarray(img)
    pred_bbox = pred['bboxes']
    pred_score = pred['scores']
    # print("Output Image: ", img.size, type(img))
    return img, pred_bbox, pred_score

# Populate examples in Gradio interface
examples = [
    ['1.png'],
    ['2.png'],
    ['3.png']
]

description = "An object detection framework to localize regions under post-admixture selection from images of ancestry-painted chromosomes!"

gr_interface = gr.Interface(
    fn=show_preds,
    inputs=[gr.Image(label="Upload 200x200 B&W image of ancestry-painted chromosomes:"), gr.Slider(minimum=0, maximum=1, step=0.1, default=0.5, label="Detection Threshold")],
    outputs=[gr.Image(type="pil"), gr.Textbox(label="Predicted BBox:"), gr.Textbox(label="Predicted BBox Score:")],
    title='Detect adaptive variants in admixed populations',
    description=description,
    examples=examples)

gr_interface.launch(inline=False, share=False, debug=True)