app.py

from huggingface_hub import from_pretrained_fastai
import gradio as gr
from fastai.vision.all import *
import PIL
import torchvision.transforms as transforms

##Extras por si pudiera reconstruir la imagen en HF también
import numpy as np
import os
import cv2

def extract_subimages(image : np.ndarray, wwidth, wheight, overlap_fraction):
    """
    Extracts subimages of the input image using a moving window of size (wwidth, wheight)
    with the specified overlap fraction. Returns a tuple (subimages, coords) where subimages
    is a list of subimages and coords is a list of tuples (x, y) indicating the top left corner
    coordinates of each subimage in the input image.
    """
    subimages = []
    coords = []
    height, width, channels = image.shape
    if channels > 3:
      image = image[:,:,0:3]
      channels = 3
    overlap = int(max(0, min(overlap_fraction, 1)) * min(wwidth, wheight))
    y = 0
    while y + wheight <= height:
        x = 0
        while x + wwidth <= width:
            subimage = image[y:y+wheight, x:x+wwidth, :]
            subimages.append(subimage)
            coords.append((x, y))
            x += wwidth - overlap
        y += wheight - overlap
    if y < height:
        y = height - wheight
        x = 0
        while x + wwidth <= width:
            subimage = image[y:y+wheight, x:x+wwidth, :]
            subimages.append(subimage)
            coords.append((x, y))
            x += wwidth - overlap
        if x < width:
            x = width - wwidth
            subimage = image[y:y+wheight, x:x+wwidth, :]
            subimages.append(subimage)
            coords.append((x, y))
    if x < width:
        x = width - wwidth
        y = 0
        while y + wheight <= height:
            subimage = image[y:y+wheight, x:x+wwidth, :]
            subimages.append(subimage)
            coords.append((x, y))
            y += wheight - overlap
        if y < height:
            y = height - wheight
            subimage = image[y:y+wheight, x:x+wwidth, :]
            subimages.append(subimage)
            coords.append((x, y))
    return subimages, coords

# Si no hay archivos tif (labels) no se tratan, no hace falta considerarlo
def generate_and_save_subimages(path, output_dir_images, output_dir_labels = None):
  if output_dir_labels:
    if not os.path.exists(output_dir_labels):
        os.makedirs(output_dir_labels)

  if not os.path.exists(output_dir_images):
      os.makedirs(output_dir_images)

  for filename in os.listdir(path):
      if filename.endswith(".png") or filename.endswith(".tif"):
          filepath = os.path.join(path, filename)
          image = cv2.imread(filepath)
          subimages, coords = extract_subimages(image, 400, 400, 0.66)
          for i, subimage in enumerate(subimages):
              if filename.endswith(".png"):
                  output_filename = os.path.join(output_dir_images, f"{filename.rsplit('.', 1)[0]}_{coords[i][0]}_{coords[i][1]}.png")
                  cv2.imwrite(output_filename, subimage)
              else:
                  if output_dir_labels:
                    output_filename = os.path.join(output_dir_labels, f"{filename.rsplit('.', 1)[0]}_{coords[i][0]}_{coords[i][1]}.tif")
                    cv2.imwrite(output_filename, subimage)

def generate_and_save_subimages_nolabel(path, output_dir_images, olverlap=0.0, imagesformat="png", split_in_dirs=True):
  for entry in os.scandir(path):
    if entry.is_file() and entry.name.lower().endswith(imagesformat):
      filepath = entry.path
      gss_single(filepath, output_dir_images, olverlap, imagesformat, split_in_dirs)

def gss_single(filepath, output_dir_images, olverlap=0.0, imagesformat="png", split_in_dirs=True):
      image = cv2.imread(filepath)

      if split_in_dirs:
        dir_this_image = Path(output_dir_images)/filepath.rsplit('.', 1)[0]
        os.makedirs(dir_this_image, exist_ok=True)
      else:
        os.makedirs(output_dir_images, exist_ok=True)

      subimages, coords = extract_subimages(image, 400, 400, olverlap)
      for i, subimage in enumerate(subimages):
        if split_in_dirs:
          output_filename = os.path.join(dir_this_image, f"{filepath.rsplit('.', 1)[0]}_{coords[i][0]}_{coords[i][1]}.png")
        else:
          output_filename = os.path.join(output_dir_images, f"{filepath.rsplit('.', 1)[0]}_{coords[i][0]}_{coords[i][1]}.png")
        cv2.imwrite(output_filename, subimage)

def split_windows_in_folders(input_images_folder, output_images_folder):
  for filename in os.listdir(input_images_folder):
    dir_this_image = Path(output_images_folder)/filename.rsplit('.', 1)[0]
    os.makedirs(dir_this_image, exist_ok=True)
    if filename.endswith(".png"):
      print(str(dir_this_image))
      filepath = os.path.join(path, filename)
      image = cv2.imread(filepath)
      subimages, coords = extract_subimages(image, 400, 400, 0)
      for i, subimage in enumerate(subimages):
          output_filename = os.path.join(dir_this_image, f"{filename.rsplit('.', 1)[0]}_{coords[i][0]}_{coords[i][1]}.png")
          cv2.imwrite(output_filename, subimage)


def subimages_from_directory(directorio):
  # Define el directorio a recorrer
  directorio = directorio

  # Define la expresión regular para buscar los números X e Y en el nombre de archivo
  patron = re.compile(r"(.*)_(\d+)_(\d+)\.(png|jpg|tif)")

  windowlist = []
  coords = []

  # Recorre el directorio en busca de imágenes
  for filename in os.listdir(directorio):
    match = patron.search(filename)
    if match:
        origname = match.group(1)
        x = int(match.group(2))
        y = int(match.group(3))
        #print(f"El archivo {filename} tiene los números X={x} e Y={y}")
        img = cv2.imread(os.path.join(directorio, filename))
        windowlist.append(img)
        coords.append((x, y))

  # Ordena las listas por coordenadas X e Y
  windowlist, coords = zip(*sorted(zip(windowlist, coords), key=lambda pair: (pair[1][0], pair[1][1])))
  wh, ww, chan = windowlist[0].shape
  origsize = tuple(elem1 + elem2 for elem1, elem2 in zip(coords[-1], (wh,ww)))

  return windowlist, coords, wh, ww, chan, origsize

def subimages_onlypath(directorio):
  # Define el directorio a recorrer
  directorio = directorio
  pathlist = []

  patron = re.compile(r"(.*)_(\d+)_(\d+)\.(png|jpg|tif)")

  for filename in os.listdir(directorio):
    match = patron.search(filename)
    if match:
        pathlist.append(os.path.join(directorio, filename))

  return pathlist

def ReconstructFromMW(windowlist, coords, wh, ww, chan, origsize):
  canvas = np.zeros((origsize[1], origsize[0], chan), dtype=np.uint8)
  for idx, window in enumerate(windowlist):
    canvas[coords[idx][1]:coords[idx][1]+wh, coords[idx][0]:coords[idx][0]+ww, :] = window
  return canvas

def get_list_tp(path):
  list_to_process = []  # Inicializar la lista que contendrá los nombres de los subdirectorios
  list_names = []
  # Recorrer los elementos del directorio
  for element in os.scandir(path):
      # Verificar si el elemento es un directorio
      if element.is_dir():
          # Agregar el nombre del subdirectorio a la lista
          windowlist, coords, wh, ww, chan, origsize = subimages_from_directory(element)
          list_to_process.append(ReconstructFromMW(windowlist, coords, wh, ww, chan, origsize))
          list_names.append(element.name)
  return list_to_process, list_names

def get_paths_tp(path):
  list_to_process = []  # Inicializar la lista que contendrá los nombres de los subdirectorios
  # Recorrer los elementos del directorio
  for element in os.scandir(path):
      # Verificar si el elemento es un directorio
      if element.is_dir():
          # Agregar el nombre del subdirectorio a la lista
          list_to_process.append(subimages_onlypath(element))
  return list_to_process

def process_multifolder(process_folders, result_folder):
  for folder in process_folders:
    folname = os.path.basename(os.path.dirname(folder[0]))
    destname = Path(result_folder)/folname
    os.makedirs(destname, exist_ok=True)
    for subimagepath in folder:
      img = PIL.Image.open(subimagepath)
      image = transforms.Resize((400,400))(img)
      tensor = transform_image(image=image)
      with torch.no_grad():
          outputs = model(tensor)
      outputs = torch.argmax(outputs,1)
      mask = np.array(outputs.cpu())
      mask[mask==1]=255
      mask=np.reshape(mask,(400,400))
      mask_img = Image.fromarray(mask.astype('uint8'))

      filename = os.path.basename(subimagepath)
      new_image_path = os.path.join(result_folder, folname, filename)
      mask_img.save(new_image_path)

def recombine_windows(results_folder_w, result_f_rec):
  imgs, nombres = get_list_tp(results_folder_w)
  os.makedirs(result_f_rec, exist_ok=True)

  for idx, image in enumerate(imgs):
    img = Image.fromarray(image)
    new_image_path = os.path.join(result_f_rec, nombres[idx] + '.tif')
    img.save(new_image_path, compression='tiff_lzw')
  return new_image_path

def process_single_image(single_image_path, base_f, pro_f, rsw_f, rsd_f):
  gss_single(single_image_path, pro_f, 0, "tif", True)
  process_multifolder(get_paths_tp(pro_f),rsw_f)
  pt = recombine_windows(rsw_f,rsd_f)
  shutil.rmtree(pro_f)
  shutil.rmtree(rsw_f)
  #copiar_info_georref(single_image_path, pt)
  return pt

# from osgeo import gdal, osr

# def copiar_info_georref(entrada, salida):
#     try:
#         # Abrir el archivo GeoTIFF original
#         original_dataset = gdal.Open(entrada)

#         # Obtener la información de georreferenciación del archivo original
#         original_projection = original_dataset.GetProjection()
#         original_geotransform = original_dataset.GetGeoTransform()

#         # Abrir la imagen resultado
#         result_dataset = gdal.Open(salida, gdal.GA_Update)

#         # Copiar la información de georreferenciación del archivo original a la imagen resultado
#         result_dataset.SetProjection(original_projection)
#         result_dataset.SetGeoTransform(original_geotransform)

#         # Cerrar los archivos
#         original_dataset = None
#         result_dataset = None

#     except Exception as e:
#         print("Error: ", e)

###FIN de extras



#repo_id = "Ignaciobfp/segmentacion-dron-marras"
#learner = from_pretrained_fastai(repo_id)

device = torch.device("cpu") 
#model = learner.model
model = torch.jit.load("modelo_marras.pth")
model = model.cpu()

def transform_image(image):
    my_transforms = transforms.Compose([transforms.ToTensor(),
                                        transforms.Normalize(
                                            [0.485, 0.456, 0.406],
                                            [0.229, 0.224, 0.225])])
    image_aux = image
    return my_transforms(image_aux).unsqueeze(0).to(device)


# Definimos una función que se encarga de llevar a cabo las predicciones
def predict(img):
    img_pil = PIL.Image.fromarray(img, 'RGB')
    image = transforms.Resize((400,400))(img_pil)
    tensor = transform_image(image=image)
    model.to(device)
    with torch.no_grad():
        outputs = model(tensor)
    outputs = torch.argmax(outputs,1)
    mask = np.array(outputs.cpu())
    mask[mask==1]=255
    mask=np.reshape(mask,(400,400))
    return Image.fromarray(mask.astype('uint8'))

def predict_full(img):
    single_image_path = "tmp.tif"
    base_f = "."
    pro_f = "processing"
    rsw_f = "results_windows"
    rsd_f = "results_together"
    destpath = process_single_image(single_image_path, base_f, pro_f, rsw_f, rsd_f)
    im = Image.open(destpath)
    return im
    
# Creamos la interfaz y la lanzamos.
gr.Interface(fn=predict, inputs=gr.inputs.Image(shape=(400, 400)), outputs=gr.outputs.Image(type="pil"), examples=['examples/1CA SUR_1200_800.png', 'examples/1CA SUR_4000_1200.png', 'examples/1CA SUR_4800_2000.png']).launch(share=False)
# gr.Interface(fn=predict_full, inputs=gr.inputs.Image(), outputs=gr.outputs.Image(type="pil")).launch(share=False)