app.py

import torch
import cv2
import torch.nn as nn
import numpy as np
from torchvision import models, transforms
import time
import os
import copy
import pickle
from PIL import Image
import datetime
import gdown
import zipfile
import urllib.request
import gradio as gr

IMG_SIZE = 512
CLASSES = [ "No DR", "Mild", "Moderate", "Severe", "Proliferative DR" ]

checkpoint = "./demo_checkpoint_convnext.pth"

device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")

model = torch.load(checkpoint).to(device)

global_transforms = transforms.Compose([
    transforms.ToPILImage(),
    transforms.Lambda(lambda image: image.convert('RGB')),
    transforms.Resize(IMG_SIZE),
    transforms.ToTensor(),
    transforms.Normalize([0.2786802, 0.2786802, 0.2786802], [0.16637428, 0.16637428, 0.16637428])
    ])
    
def crop_image_from_gray(img,tol=7):
  mask = img>tol

  img1=img[np.ix_(mask.any(1),mask.any(0))]
  img2=img[np.ix_(mask.any(1),mask.any(0))]
  img3=img[np.ix_(mask.any(1),mask.any(0))]
  img = np.stack([img1,img2,img3],axis=-1)

  return img

def circle_crop(img):   
    height, width = img.shape    
    
    x = int(width/2)
    y = int(height/2)
    r = np.amin((x,y))
    
    circle_img = np.zeros((height, width), np.uint8)
    cv2.circle(circle_img, (x,y), int(r), 1, thickness=-1)
    img = cv2.bitwise_and(img, img, mask=circle_img)
    img = crop_image_from_gray(img)
    return img 
    
def preprocess(img):
  # Extract Green Channel
  img = img[:,:,1]
  #CLAHE
  clahe = cv2.createCLAHE(clipLimit=2.0, tileGridSize=(8, 8))
  img = clahe.apply(img)
  # Circle crop
  img = circle_crop(img)
  # Resize
  img = cv2.resize(img, (IMG_SIZE,IMG_SIZE))

  return img
  
def do_inference(img):
  img = preprocess(img)
  img_t = global_transforms(img)
  batch_t = torch.unsqueeze(img_t, 0)
  model.eval()
  # We don't need gradients for test, so wrap in 
  # no_grad to save memory
  with torch.no_grad():
      batch_t = batch_t.to(device)
      # forward propagation
      output = model( batch_t)
      # get prediction
      probs = torch.nn.functional.softmax(output, dim=1)
      output = torch.argsort(probs, dim=1, descending=True).cpu().numpy()[0].astype(int)
      probs = probs.cpu().numpy()[0]
      probs = probs[output]
      labels = np.array(CLASSES)[output]
      return {labels[i]: round(float(probs[i]),2) for i in range(len(labels))}

im = gr.inputs.Image(shape=(512, 512), image_mode='RGB', 
                         invert_colors=False, source="upload",
                         type="pil")     
title = "ConvNeXt for Diabetic Retinopathy Detection"
description = ""
examples = [['./noDr.png'],['./severe.png']]
#article="<p style='text-align: center'><a href='https://github.com/mawady/colab-recipes-cv' target='_blank'>Colab Recipes for Computer Vision - Dr. Mohamed Elawady</a></p>"
iface = gr.Interface(
    do_inference, 
    im, 
    gr.outputs.Label(num_top_classes=5),
    live=False,
    interpretation=None,
    title=title,
    description=description,
    examples=examples
)

#iface.test_launch()

iface.launch()