from transformers import AutoFeatureExtractor, YolosForObjectDetection
import gradio as gr
from PIL import Image
import torch
import matplotlib.pyplot as plt
import io
import numpy as np


COLORS = [[0.000, 0.447, 0.741], [0.850, 0.325, 0.098], [0.929, 0.694, 0.125],
          [0.494, 0.184, 0.556], [0.466, 0.674, 0.188], [0.301, 0.745, 0.933]]


def get_class_list_from_input(classes_string: str):
    if classes_string == "":
        return []
    classes_list = classes_string.split(",")
    classes_list = [x.strip() for x in classes_list]
    return classes_list

def infer(img, model_name: str, prob_threshold: int, classes_to_show = str):
    feature_extractor = AutoFeatureExtractor.from_pretrained(f"hustvl/{model_name}")
    model = YolosForObjectDetection.from_pretrained(f"hustvl/{model_name}")

    img = Image.fromarray(img)
    
    pixel_values = feature_extractor(img, return_tensors="pt").pixel_values

    with torch.no_grad():
        outputs = model(pixel_values, output_attentions=True)
    
    probas = outputs.logits.softmax(-1)[0, :, :-1]
    keep = probas.max(-1).values > prob_threshold

    target_sizes = torch.tensor(img.size[::-1]).unsqueeze(0)
    postprocessed_outputs = feature_extractor.post_process(outputs, target_sizes)
    bboxes_scaled = postprocessed_outputs[0]['boxes']
    
    classes_list = get_class_list_from_input(classes_to_show)
    res_img = plot_results(img, probas[keep], bboxes_scaled[keep], model, classes_list)
    
    return res_img

def plot_results(pil_img, prob, boxes, model, classes_list):
    plt.figure(figsize=(16,10))
    plt.imshow(pil_img)
    ax = plt.gca()
    colors = COLORS * 100
    for p, (xmin, ymin, xmax, ymax), c in zip(prob, boxes.tolist(), colors):
        cl = p.argmax()
        object_class = model.config.id2label[cl.item()]
        
        if len(classes_list) > 0 :
            if object_class not in classes_list:
                continue
            
        ax.add_patch(plt.Rectangle((xmin, ymin), xmax - xmin, ymax - ymin,
                                fill=False, color=c, linewidth=3))
        text = f'{object_class}: {p[cl]:0.2f}'
        ax.text(xmin, ymin, text, fontsize=15,
                bbox=dict(facecolor='yellow', alpha=0.5))
    plt.axis('off')
    return fig2img(plt.gcf())
    
def fig2img(fig):
    buf = io.BytesIO()
    fig.savefig(buf)
    buf.seek(0)
    img = Image.open(buf)
    return img

description = """Object Detection with YOLOS. Choose https://github.com/amikelive/coco-labels/blob/master/coco-labels-2014_2017.txtyour model and you're good to go.

You can adapt the minimum probability threshold with the slider.

Additionally you can restrict the classes that will be shown by putting in a comma separated list of 
[COCO classes](https://github.com/amikelive/coco-labels/blob/master/coco-labels-2014_2017.txt). 
Leaving the field empty will show all classes"""

image_in = gr.components.Image()
image_out = gr.components.Image()
model_choice = gr.components.Dropdown(["yolos-tiny", "yolos-small", "yolos-base", "yolos-small-300", "yolos-small-dwr"], value="yolos-small", label="YOLOS Model")
prob_threshold_slider = gr.components.Slider(minimum=0, maximum=1.0, step=0.01, value=0.9, label="Probability Threshold")
classes_to_show = gr.components.Textbox(placeholder="e.g. person, boat", label="Classes to use (empty means all classes)")

Iface = gr.Interface(
    fn=infer,
    inputs=[image_in,model_choice, prob_threshold_slider, classes_to_show],
    outputs=image_out,
    #examples=[["examples/10_People_Marching_People_Marching_2_120.jpg"], ["examples/12_Group_Group_12_Group_Group_12_26.jpg"], ["examples/43_Row_Boat_Canoe_43_247.jpg"]],
    title="Object Detection with YOLOS",
    description=description,
).launch()