app.py

import gradio as gr
import torch
from utils import colorize
from utils import get_min_and_max
from PIL import Image
import tempfile
import numpy as np

import json

DEVICE = 'cuda' if torch.cuda.is_available() else 'cpu'
model = torch.hub.load('isl-org/ZoeDepth', "ZoeD_N", pretrained=True).to(DEVICE).eval()

def split_numpy_16_bit(numpy_image_16bit):
  result_numpy_split = np.zeros((numpy_image_16bit.shape[0], numpy_image_16bit.shape[1], 3));

  i = 0;
  for row in numpy_image_16bit:
    j = 0
    for pixel in row:
      firstPart = np.array(pixel).astype(np.uint8)
      secondPart = np.array(pixel>>8).astype(np.uint8)

      result_numpy_split[i, j, 0] = firstPart
      result_numpy_split[i, j, 1] = secondPart
      j += 1
    i += 1

  return result_numpy_split
    
def predict(image):
    image.thumbnail((1024,1024))
    depth = model.infer_pil(image)
    colored_depth, vmin, vmax = colorize(depth, cmap='gray_r')

    
    tmp = tempfile.NamedTemporaryFile(suffix='.png', delete=False)
    raw_depth = Image.fromarray((depth*256).astype('uint16'))
    raw_depth.save(tmp.name)

    depth_uint16 = (depth*256).astype('uint16')
    numpy_image_split = split_numpy_16_bit(depth_uint16)

    depth_image_two_components = Image.fromarray(np.uint8(numpy_image_split)).convert('RGB')
    
    return colored_depth, tmp.name, depth_image_two_components
    
iface = gr.Interface(fn=predict, inputs=gr.Image(label="Input Image", type='pil'), outputs=["image", "file", "image"])
iface.launch()