karolmajek/Axial-DeepLab-SWideRNet

1import collections2import os3import tempfile4from matplotlib import gridspec5from matplotlib import pyplot as plt6import numpy as np7from PIL import Image8import urllib9import tensorflow as tf10import gradio as gr11from subprocess import call12import sys13import requests14url1 = 'https://cdn.pixabay.com/photo/2014/09/07/21/52/city-438393_1280.jpg'15r = requests.get(url1, allow_redirects=True)16open("city1.jpg", 'wb').write(r.content)17url2 = 'https://cdn.pixabay.com/photo/2016/02/19/11/36/canal-1209808_1280.jpg'18r = requests.get(url2, allow_redirects=True)19open("city2.jpg", 'wb').write(r.content)20DatasetInfo = collections.namedtuple(21    'DatasetInfo',22    'num_classes, label_divisor, thing_list, colormap, class_names')23def _cityscapes_label_colormap():24  """Creates a label colormap used in CITYSCAPES segmentation benchmark.25  See more about CITYSCAPES dataset at https://www.cityscapes-dataset.com/26  M. Cordts, et al. "The Cityscapes Dataset for Semantic Urban Scene Understanding." CVPR. 2016.27  Returns:28    A 2-D numpy array with each row being mapped RGB color (in uint8 range).29  """30  colormap = np.zeros((256, 3), dtype=np.uint8)31  colormap[0] = [128, 64, 128]32  colormap[1] = [244, 35, 232]33  colormap[2] = [70, 70, 70]34  colormap[3] = [102, 102, 156]35  colormap[4] = [190, 153, 153]36  colormap[5] = [153, 153, 153]37  colormap[6] = [250, 170, 30]38  colormap[7] = [220, 220, 0]39  colormap[8] = [107, 142, 35]40  colormap[9] = [152, 251, 152]41  colormap[10] = [70, 130, 180]42  colormap[11] = [220, 20, 60]43  colormap[12] = [255, 0, 0]44  colormap[13] = [0, 0, 142]45  colormap[14] = [0, 0, 70]46  colormap[15] = [0, 60, 100]47  colormap[16] = [0, 80, 100]48  colormap[17] = [0, 0, 230]49  colormap[18] = [119, 11, 32]50  return colormap51def _cityscapes_class_names():52  return ('road', 'sidewalk', 'building', 'wall', 'fence', 'pole',53          'traffic light', 'traffic sign', 'vegetation', 'terrain', 'sky',54          'person', 'rider', 'car', 'truck', 'bus', 'train', 'motorcycle',55          'bicycle')56def cityscapes_dataset_information():57  return DatasetInfo(58      num_classes=19,59      label_divisor=1000,60      thing_list=tuple(range(11, 19)),61      colormap=_cityscapes_label_colormap(),62      class_names=_cityscapes_class_names())63def perturb_color(color, noise, used_colors, max_trials=50, random_state=None):64  """Pertrubs the color with some noise.65  If `used_colors` is not None, we will return the color that has66  not appeared before in it.67  Args:68    color: A numpy array with three elements [R, G, B].69    noise: Integer, specifying the amount of perturbing noise (in uint8 range).70    used_colors: A set, used to keep track of used colors.71    max_trials: An integer, maximum trials to generate random color.72    random_state: An optional np.random.RandomState. If passed, will be used to73      generate random numbers.74  Returns:75    A perturbed color that has not appeared in used_colors.76  """77  if random_state is None:78    random_state = np.random79  for _ in range(max_trials):80    random_color = color + random_state.randint(81        low=-noise, high=noise + 1, size=3)82    random_color = np.clip(random_color, 0, 255)83    if tuple(random_color) not in used_colors:84      used_colors.add(tuple(random_color))85      return random_color86  print('Max trial reached and duplicate color will be used. Please consider '87        'increase noise in `perturb_color()`.')88  return random_color89def color_panoptic_map(panoptic_prediction, dataset_info, perturb_noise):90  """Helper method to colorize output panoptic map.91  Args:92    panoptic_prediction: A 2D numpy array, panoptic prediction from deeplab93      model.94    dataset_info: A DatasetInfo object, dataset associated to the model.95    perturb_noise: Integer, the amount of noise (in uint8 range) added to each96      instance of the same semantic class.97  Returns:98    colored_panoptic_map: A 3D numpy array with last dimension of 3, colored99      panoptic prediction map.100    used_colors: A dictionary mapping semantic_ids to a set of colors used101      in `colored_panoptic_map`.102  """103  if panoptic_prediction.ndim != 2:104    raise ValueError('Expect 2-D panoptic prediction. Got {}'.format(105        panoptic_prediction.shape))106  semantic_map = panoptic_prediction // dataset_info.label_divisor107  instance_map = panoptic_prediction % dataset_info.label_divisor108  height, width = panoptic_prediction.shape109  colored_panoptic_map = np.zeros((height, width, 3), dtype=np.uint8)110  used_colors = collections.defaultdict(set)111  # Use a fixed seed to reproduce the same visualization.112  random_state = np.random.RandomState(0)113  unique_semantic_ids = np.unique(semantic_map)114  for semantic_id in unique_semantic_ids:115    semantic_mask = semantic_map == semantic_id116    if semantic_id in dataset_info.thing_list:117      # For `thing` class, we will add a small amount of random noise to its118      # correspondingly predefined semantic segmentation colormap.119      unique_instance_ids = np.unique(instance_map[semantic_mask])120      for instance_id in unique_instance_ids:121        instance_mask = np.logical_and(semantic_mask,122                                       instance_map == instance_id)123        random_color = perturb_color(124            dataset_info.colormap[semantic_id],125            perturb_noise,126            used_colors[semantic_id],127            random_state=random_state)128        colored_panoptic_map[instance_mask] = random_color129    else:130      # For `stuff` class, we use the defined semantic color.131      colored_panoptic_map[semantic_mask] = dataset_info.colormap[semantic_id]132      used_colors[semantic_id].add(tuple(dataset_info.colormap[semantic_id]))133  return colored_panoptic_map, used_colors134def vis_segmentation(image,135                     panoptic_prediction,136                     dataset_info,137                     perturb_noise=60):138  """Visualizes input image, segmentation map and overlay view."""139  plt.figure(figsize=(30, 20))140  grid_spec = gridspec.GridSpec(2, 2)141  ax = plt.subplot(grid_spec[0])142  plt.imshow(image)143  plt.axis('off')144  ax.set_title('input image', fontsize=20)145  ax = plt.subplot(grid_spec[1])146  panoptic_map, used_colors = color_panoptic_map(panoptic_prediction,147                                                 dataset_info, perturb_noise)148  plt.imshow(panoptic_map)149  plt.axis('off')150  ax.set_title('panoptic map', fontsize=20)151  ax = plt.subplot(grid_spec[2])152  plt.imshow(image)153  plt.imshow(panoptic_map, alpha=0.7)154  plt.axis('off')155  ax.set_title('panoptic overlay', fontsize=20)156  ax = plt.subplot(grid_spec[3])157  max_num_instances = max(len(color) for color in used_colors.values())158  # RGBA image as legend.159  legend = np.zeros((len(used_colors), max_num_instances, 4), dtype=np.uint8)160  class_names = []161  for i, semantic_id in enumerate(sorted(used_colors)):162    legend[i, :len(used_colors[semantic_id]), :3] = np.array(163        list(used_colors[semantic_id]))164    legend[i, :len(used_colors[semantic_id]), 3] = 255165    if semantic_id < dataset_info.num_classes:166      class_names.append(dataset_info.class_names[semantic_id])167    else:168      class_names.append('ignore')169  plt.imshow(legend, interpolation='nearest')170  ax.yaxis.tick_left()171  plt.yticks(range(len(legend)), class_names, fontsize=15)172  plt.xticks([], [])173  ax.tick_params(width=0.0, grid_linewidth=0.0)174  plt.grid('off')175  return plt176def run_cmd(command):177    try:178        print(command)179        call(command, shell=True)180    except KeyboardInterrupt:181        print("Process interrupted")182        sys.exit(1)183MODEL_NAME = 'axial_swidernet_1_1_4.5_os16_axial_deeplab_cityscapes_trainfine_saved_model'184_MODELS = ('resnet50_os32_panoptic_deeplab_cityscapes_crowd_trainfine_saved_model',185           'resnet50_beta_os32_panoptic_deeplab_cityscapes_trainfine_saved_model',186           'wide_resnet41_os16_panoptic_deeplab_cityscapes_trainfine_saved_model',187           'swidernet_sac_1_1_1_os16_panoptic_deeplab_cityscapes_trainfine_saved_model',188           'swidernet_sac_1_1_3_os16_panoptic_deeplab_cityscapes_trainfine_saved_model',189           'swidernet_sac_1_1_4.5_os16_panoptic_deeplab_cityscapes_trainfine_saved_model',190           'axial_swidernet_1_1_1_os16_axial_deeplab_cityscapes_trainfine_saved_model',191           'axial_swidernet_1_1_3_os16_axial_deeplab_cityscapes_trainfine_saved_model',192           'axial_swidernet_1_1_4.5_os16_axial_deeplab_cityscapes_trainfine_saved_model',193           'max_deeplab_s_backbone_os16_axial_deeplab_cityscapes_trainfine_saved_model',194           'max_deeplab_l_backbone_os16_axial_deeplab_cityscapes_trainfine_saved_model')195_DOWNLOAD_URL_PATTERN = 'https://storage.googleapis.com/gresearch/tf-deeplab/saved_model/%s.tar.gz'196_MODEL_NAME_TO_URL_AND_DATASET = {197    model: (_DOWNLOAD_URL_PATTERN % model, cityscapes_dataset_information())198    for model in _MODELS199}200MODEL_URL, DATASET_INFO = _MODEL_NAME_TO_URL_AND_DATASET[MODEL_NAME]201model_dir = tempfile.mkdtemp()202download_path = os.path.join(model_dir, MODEL_NAME + '.gz')203urllib.request.urlretrieve(MODEL_URL, download_path)204run_cmd("tar -xzvf " + download_path + " -C " + model_dir)205LOADED_MODEL = tf.saved_model.load(os.path.join(model_dir, MODEL_NAME))206def inference(image):207    image = image.resize(size=(512, 512))208    im = np.array(image)209    output = LOADED_MODEL(tf.cast(im, tf.uint8))210    return vis_segmentation(im, output['panoptic_pred'][0], DATASET_INFO)211title = "Deeplab2 - Axial Deeplab SWideRNet-(1, 1, 4.5)"212description = "demo for Deeplab2. To use it, simply upload your image, or click one of the examples to load them. Read more at the links below.\nModel: 'axial_swidernet_1_1_4.5_os16_axial_deeplab_cityscapes_trainfine_saved_model'"213article = "<p style='text-align: center'><a href='https://arxiv.org/abs/2106.09748'>DeepLab2: A TensorFlow Library for Deep Labeling</a> | <a href='https://github.com/google-research/deeplab2'>Github Repo</a></p>"214gr.Interface(215    inference, 216    [gr.inputs.Image(type="pil", label="Input")], 217    gr.outputs.Image(type="plot", label="Output"),218    title=title,219    description=description,220    article=article,221    examples=[222            ["city1.jpg"],223            ["city2.jpg"]224        ]).launch()225