app.py

import os
import numpy as np
import gradio as gr
import torch
from torchvision import models, transforms
from PIL import Image


# -- install detectron2 from source ------------------------------------------------------------------------------
os.system('pip install git+https://github.com/facebookresearch/detectron2.git')
os.system('pip install pyyaml==5.1')

import detectron2

from detectron2.utils.logger import setup_logger

from detectron2 import model_zoo
from detectron2.engine import DefaultPredictor
from detectron2.config import get_cfg
from detectron2.utils.visualizer import Visualizer
from detectron2.data import MetadataCatalog, DatasetCatalog
import cv2


setup_logger()

# -- load rcnn model ---------------------------------------------------------------------------------------------
cfg = get_cfg()
# load model config
cfg.merge_from_file(model_zoo.get_config_file("COCO-InstanceSegmentation/mask_rcnn_R_50_FPN_3x.yaml"))
cfg.MODEL.ROI_HEADS.SCORE_THRESH_TEST = 0.5  # set threshold for this model
# set model weights
cfg.MODEL.WEIGHTS = model_zoo.get_checkpoint_url("COCO-InstanceSegmentation/mask_rcnn_R_50_FPN_3x.yaml")
cfg.MODEL.DEVICE= 'cpu'  # move to cpu
predictor = DefaultPredictor(cfg)  # create model

# -- load design modernity model for classification --------------------------------------------------------------
DesignModernityModel = torch.load("DesignModernityModel.pt")

DesignModernityModel.eval() # set state of the model to inference

# Set class labels
LABELS = ['2000-2003', '2006-2008', '2009-2011', '2012-2014', '2015-2018']
n_labels = len(LABELS)

# define maéan and std dev for normalization
MEAN = [0.485, 0.456, 0.406]
STD = [0.229, 0.224, 0.225]

# define image transformation steps
carTransforms = transforms.Compose([transforms.Resize(224),
                                    transforms.ToTensor(),
                                    transforms.Normalize(mean=MEAN, std=STD)])

     
# -- define a function for extraction of the detected car ---------------------------------------------------------                                                                                                                                          
def cropImage(outputs, im, boxes, car_class_true):
    # Get the masks
    masks = list(np.array(outputs["instances"].pred_masks[car_class_true]))
    max_idx = torch.tensor([(x[2] - x[0])*(x[3] - x[1]) for x in boxes]).argmax().item()

    # Pick an item to mask
    item_mask = masks[max_idx]
    
    # Get the true bounding box of the mask
    segmentation = np.where(item_mask == True) # return a list of different position in the bow, which are the actual detected object
    x_min = int(np.min(segmentation[1])) # minimum x position
    x_max = int(np.max(segmentation[1]))
    y_min = int(np.min(segmentation[0]))
    y_max = int(np.max(segmentation[0]))
    
    # Create cropped image from the just portion of the image we want
    cropped = Image.fromarray(im[y_min:y_max, x_min:x_max, :], mode = 'RGB')
    # Create a PIL Image out of the mask
    mask = Image.fromarray((item_mask * 255).astype('uint8')) ###### change 255 
    # Crop the mask to match the cropped image
    cropped_mask = mask.crop((x_min, y_min, x_max, y_max))
    
    # Load in a background image and choose a paste position
    height = y_max-y_min
    width = x_max-x_min
    background = Image.new(mode='RGB', size=(width, height), color=(255, 255, 255, 0))
    
    # Create a new foreground image as large as the composite and paste the cropped image on top
    new_fg_image = Image.new('RGB', background.size)
    new_fg_image.paste(cropped)
    
    # Create a new alpha mask as large as the composite and paste the cropped mask
    new_alpha_mask = Image.new('L', background.size, color=0)
    new_alpha_mask.paste(cropped_mask)
    
    #composite the foreground and background using the alpha mask
    composite = Image.composite(new_fg_image, background, new_alpha_mask)
    
    return composite
    
# -- define function for image segmentation and classification --------------------------------------------------------
def classifyCar(im):
    # read image
    #im = cv2.imread(im)
    # perform segmentation
    outputs = predictor(im)
    v = Visualizer(im[:, :, ::-1], MetadataCatalog.get(cfg.DATASETS.TRAIN[0]), scale=1)
    out = v.draw_instance_predictions(outputs["instances"])
    # check if a car was detected in the image
    car_class_true = outputs["instances"].pred_classes == 2
    boxes = list(outputs["instances"].pred_boxes[car_class_true])
    
    # if a car was detected, extract the car and perform modernity score classification
    if len(boxes) != 0:
        im2 = cropImage(outputs, im, boxes, car_class_true)

      
        with torch.no_grad():
            scores = torch.nn.functional.softmax(DesignModernityModel(carTransforms(im2).unsqueeze(0))[0])
        label = {LABELS[i]: float(scores[i]) for i in range(n_labels)}
      
    # if no car was detected, show original image and print "No car detected"  
    else:
        im2 = Image.fromarray(np.uint8(im)).convert('RGB')
        label = "No car detected"

    return im2, label

# -- create interface for model ----------------------------------------------------------------------------------------
interface = gr.Interface(classifyCar, inputs='image', outputs=['image','label'], cache_examples=False, title='Modernity car classification')
interface.launch()