app.py

import os
import copy
import time

import numpy as np
import onnxruntime

from PIL import Image, ImageOps

import gradio

def run_inference(onnx_session, input_size, image):
    # Resize
    temp_image = image.copy()
    resize_image = temp_image.resize((input_size, input_size), Image.ANTIALIAS)
    x = ImageOps.exif_transpose(resize_image)
    x = np.array(x)

    # Preprocessing
    x = x.astype(np.float32)
    mean = np.array([0.485, 0.456, 0.406], dtype=np.float32)
    std = np.array([0.229, 0.224, 0.225], dtype=np.float32)
    x = (x / 255 - mean) / std
    x = x.transpose(2, 0, 1).astype('float32')
    x = x.reshape(-1, 3, input_size, input_size)

    # Inference
    input_name = onnx_session.get_inputs()[0].name
    output_name = onnx_session.get_outputs()[0].name
    onnx_result = onnx_session.run([output_name], {input_name: x})

    # Postprocessing
    onnx_result = np.array(onnx_result).squeeze()
    min_value = np.min(onnx_result)
    max_value = np.max(onnx_result)
    onnx_result = (onnx_result - min_value) / (max_value - min_value)
    onnx_result *= 255
    onnx_result = onnx_result.astype('uint8')

    return onnx_result

# Load model
onnx_session = onnxruntime.InferenceSession("u2net.onnx")

def create_rgba(mode, image):
    image = Image.fromarray(image).convert('RGB')
    out = run_inference(
        onnx_session,
        320,
        image,
    )
    resize_image = Image.fromarray(out).resize((image.size[0], image.size[1]), Image.ANTIALIAS)

    if mode == "binary":
        resize_image = resize_image.point(lambda x: 255 if x > 125 else 0)

    mask = resize_image

    rgba_image = image.convert('RGBA')
    rgba_image.putalpha(mask)

    return rgba_image

inputs = [gradio.inputs.Radio(["binary", "smooth"]), gradio.inputs.Image()]
outputs = gradio.outputs.Image(type='numpy')
gradio.Interface(fn=create_rgba, inputs=inputs, outputs=outputs).launch()