app.py

import torch

from transformers import AutoImageProcessor, AutoModelForObjectDetection
#from transformers import pipeline

from PIL import Image
import matplotlib.pyplot as plt
import matplotlib.patches as patches

import io
from random import choice


image_processor_tiny = AutoImageProcessor.from_pretrained("hustvl/yolos-tiny")
model_tiny = AutoModelForObjectDetection.from_pretrained("hustvl/yolos-tiny")

image_processor_small = AutoImageProcessor.from_pretrained("hustvl/yolos-small")
model_small = AutoModelForObjectDetection.from_pretrained("hustvl/yolos-small")


import gradio as gr


COLORS = ["#ff7f7f", "#ff7fbf", "#ff7fff", "#bf7fff",
            "#7f7fff", "#7fbfff", "#7fffff", "#7fffbf",
            "#7fff7f", "#bfff7f", "#ffff7f", "#ffbf7f"]

fdic = {
    "family" : "Impact",
    "style" : "italic",
    "size" : 15,
    "color" : "yellow",
    "weight" : "bold"
}


def get_figure(in_pil_img, in_results):
    plt.figure(figsize=(16, 10))
    plt.imshow(in_pil_img)
    ax = plt.gca()

    for score, label, box in zip(in_results["scores"], in_results["labels"], in_results["boxes"]):
        selected_color = choice(COLORS)

        #box = [round(i, 2) for i in box.tolist()]
        x, y, w, h = int(box[0]), int(box[1]), int(box[2]-box[0]), int(box[3]-box[1])
        print(x, y, w, h)
        ax.add_patch(plt.Rectangle((x, y), w, h, fill=False, color=selected_color, linewidth=3))
        ax.text(x, y, f"{model_tiny.config.id2label[label.item()]}: {round(score.item()*100, 1)}%", fontdict=fdic)
        #print(
        #    f"Detected {model_tiny.config.id2label[label.item()]} with confidence "
        #    f"{round(score.item(), 3)} at location {box}"
        #)

    plt.axis("off")

    return plt.gcf()


def infer(in_model, in_threshold, in_pil_img):
    print(type(in_pil_img))
    print(threshold)
    inputs = image_processor_tiny(images=in_pil_img, return_tensors="pt")
    outputs = model_tiny(**inputs)

    # convert outputs (bounding boxes and class logits) to COCO API
    target_sizes = torch.tensor([in_pil_img.size[::-1]])
    results = image_processor_tiny.post_process_object_detection(outputs, threshold=in_threshold, target_sizes=target_sizes)[
        0
    ]
    print(results)

    figure = get_figure(in_pil_img, results)

    buf = io.BytesIO()
    figure.savefig(buf, bbox_inches='tight')
    buf.seek(0)
    output_pil_img = Image.open(buf)

    return output_pil_img


#from transformers.models.flava import modeling_flava
with gr.Blocks(css=".gradio-container {background:lightyellow;color:red;}", title="テスト"
               ) as demo:
    #sample_index = gr.State([])

    gr.HTML('<div style="font-size:12pt; text-align:center; color:yellow;">MNIST 分類器</div>')

    model = gr.Radio(["detr-resnet-50", "detr-resnet-101"], value="detr-resnet-50")

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

    threshold = gr.Slider(0, 1.0, value=0.9, label='threshold')

    send_btn = gr.Button("予測する")
    send_btn.click(fn=infer, inputs=[model, threshold, input_image], outputs=[output_image])

#demo.queue()
demo.launch(debug=True)




### EOF ###