def save_images(autoencoder, args, test_folds):
assert (args.threshold != -1)
array_files = util.load_array_of_files(args.path, test_folds)
for fname in array_files:
print('Processing image', fname)
img = cv2.imread(fname, cv2.IMREAD_GRAYSCALE)
img = np.asarray(img)
rows = img.shape[0]
cols = img.shape[1]
if img.shape[0] < args.window or img.shape[1] < args.window:
new_rows = args.window if img.shape[
0] < args.window else img.shape[0]
new_cols = args.window if img.shape[
1] < args.window else img.shape[1]
img = cv2.resize(img, (new_cols, new_rows),
interpolation=cv2.INTER_CUBIC)
img = np.asarray(img).astype('float32')
img = 255. - img
finalImg = np.zeros(img.shape, dtype=bool)
for (x, y, window) in utilDataGenerator.sliding_window(
img, stepSize=args.step,
windowSize=(args.window, args.window)):
if window.shape[0] != args.window or window.shape[1] != args.window:
continue
roi = img[y:(y + args.window), x:(x + args.window)].copy()
roi = roi.reshape(1, args.window, args.window, 1)
roi = roi.astype('float32') #/ 255.
prediction = autoencoder.predict(roi)
prediction = (prediction > args.threshold)
finalImg[y:(y + args.window),
x:(x + args.window)] = prediction[0].reshape(
args.window, args.window)
finalImg = 1 - finalImg
finalImg *= 255
finalImg = finalImg.astype('uint8')
if finalImg.shape[0] != rows or finalImg.shape[1] != cols:
finalImg = cv2.resize(finalImg, (cols, rows),
interpolation=cv2.INTER_CUBIC)
outFilename = fname.replace('_GR/', '_PR-' + args.modelpath + '/')
util.mkdirp(os.path.dirname(outFilename))
cv2.imwrite(outFilename, finalImg)
weights_filename = 'model_weights_' + BASE_LOG_NAME + '.h5'
print('Loading data...')
x_sufix = '_GR'
y_sufix = '_GT'
train_folds, test_folds = load_dataset_folds(args.db, args.dbp)
# augmentation ?
if args.aug == True: # Add the augmented folders
for f in list(train_folds):
train_folds.append( util.rreplace(f, "/", "/aug_", 1) )
array_test_files = util.load_array_of_files(args.path, test_folds)
x_test, y_test = utilDataGenerator.generate_chunks(array_test_files, x_sufix, y_sufix, args.window, args.window)
if args.test == False:
array_train_files = util.load_array_of_files(args.path, train_folds)
train_data_generator = utilDataGenerator.LazyChunkGenerator(array_train_files, x_sufix, y_sufix, args.page, args.window, args.step)
train_data_generator.shuffle()
if args.start_from > 0:
train_data_generator.set_pos(args.start_from)
print('# Processing path:', args.path)
print('# Database:', args.db)
print('# Db param:', args.dbp)
print('# Train data:', len(train_data_generator) if args.test == False else '--')
def main(args=None):
args = parse_menu()
x_sufix = '_GR'
y_sufix = '_GT'
weights_filename = define_weights_filename(args)
print('Loading data...')
train_folds, test_folds = load_dataset_folds(args.db, args.dbp)
# Run data augmentation ?
if args.aug == True: # Add the augmented folders
for f in list(train_folds):
train_folds.append(util.rreplace(f, '/', '/aug_', 1))
array_test_files = util.load_array_of_files(args.path, test_folds)
x_test, y_test = utilDataGenerator.generate_chunks(array_test_files,
x_sufix, y_sufix,
args.window,
args.window)
if args.test == False:
array_train_files = util.load_array_of_files(args.path, train_folds)
train_data_generator = utilDataGenerator.LazyChunkGenerator(
array_train_files, x_sufix, y_sufix, args.page, args.window,
args.step)
train_data_generator.shuffle()
if args.start_from > 0:
train_data_generator.set_pos(args.start_from)
print('# Processing path:', args.path)
print('# Database:', args.db)
print('# Db param:', args.dbp)
print('# Train data:',
len(train_data_generator) if args.test == False else '--')
print('# Test data:', x_test.shape)
print('# Augmentation:', args.aug)
print('# Window size:', args.window)
print('# Step size:', args.step)
print('# Init weights:', args.loadmodel)
print('# nb_filters:', args.nb_filters)
print('# kernel size:', args.kernel)
print('# Dropout:', args.dropout)
print('# nb_super_epoch:', args.nb_super_epoch)
print('# nb_pages:', args.page)
print('# nb_epoch:', args.nb_epoch)
print('# batch:', args.batch)
print('# early_stopping_mode:', args.early_stopping_mode)
print('# early_stopping_patience:', args.early_stopping_patience)
print('# Threshold:', args.threshold)
print('# Weights filename:', weights_filename)
autoencoder = build_SAE_network(args, weights_filename)
best_th = args.threshold
if args.test == False:
args.monitor = 'min'
best_th = utilFit.batch_fit_with_data_generator(
autoencoder, train_data_generator, x_test, y_test, args,
weights_filename)
# Re-Load last weights
autoencoder.load_weights(weights_filename)
# Save output images
args.modelpath = weights_filename
args.threshold = best_th
save_images(autoencoder, args, test_folds)