transform_tif_images.py

from PIL import Image
import os
import numpy as np
import pathlib

def prepare_tif(path,output,img_height,img_widht,it='tif'):
    images = []
    labels = []
    categories = os.listdir(path)
    print(categories)
    for c in categories:
      data_dir = pathlib.Path(os.path.join(path,c))
      tif_data = list(data_dir.glob(f'./*.{it}'))
      for image in tif_data:
        im = Image.open(str(image))
        imarray = np.array(im)
        imarray = np.reshape(imarray,(img_height,img_widht,1))
        images.append(imarray)
        labels.append(c)

    print(f"Founded {len(images)} images from {len(set(labels))} classes")

    encoded = np.zeros((len(labels), 9))

    encoding = dict(zip(np.unique(labels),range(len(np.unique(labels)))))
    for i in range(len(encoded)):
      position = encoding[labels[i]]
      encoded[i][position] = 1

    np.savez_compressed(output+'.npz', x=np.array(images), y=np.array(encoded))

""" Example: use in keras

prepare_tif('images/train',output='images_np/train',img_height=200,img_widht=200)
prepare_tif('images/val',output='images_np/val',img_height=200,img_widht=200)

train_data = np.load('images_np/train.npz')
val_data = np.load('images_np/val.npz')

X = train_data['x']
y = train_data['y']

X_val = val_data['x']
y_val = val_data['y']

num_classes = len(y[0])
batch_size = 32
step_size_train = len(X)/batch_size
step_size_val = len(X_val)/batch_size

model = Sequential()

# create model with some layers
model.add(Conv2D(32, (3, 3), padding='same', input_shape=(200, 200,1), activation='relu'))
      ...
      ...
      ...
      ...
      ...
model.add(Dense(num_classes, activation='softmax'))
model.compile(loss='categorical_crossentropy',optimizer='adam',metrics=['accuracy'])



history = model.fit(X,y,
                   steps_per_epoch = step_size_train,
                   validation_data = (X_val,y_val),
                   validation_steps = step_size_val,
                   epochs=50)

"""