def load_dataset(source, output_size, chunk_size=3, autoencoder=False):
config_reset()
(image_path, label_path) = source[0]
config.load(yaml_str=
f'''
io:
cache:
dir: {os.path.dirname(image_path)}
dataset:
images:
type: {source[2]}
directory: {os.path.dirname(image_path)}
extension: {source[1]}
preprocess: ~
labels:
type: {source[4]}
directory: {os.path.dirname(label_path)}
extension: {source[3]}
preprocess: ~''')
if autoencoder:
return imagery_dataset.AutoencoderDataset(config.dataset.images(), (chunk_size, chunk_size),
tile_shape=config.io.tile_size(),
stride=config.train.spec().stride)
return imagery_dataset.ImageryDataset(config.dataset.images(), config.dataset.labels(),
(output_size, output_size),
(chunk_size, chunk_size), tile_shape=config.io.tile_size(),
stride=config.train.spec().stride)
def main(options):
images = config.dataset.images()
if not images:
print('No images specified.', file=sys.stderr)
return 1
tc = config.train.spec()
if options.autoencoder:
ids = imagery_dataset.AutoencoderDataset(
images, config.train.network.chunk_size(), tc.chunk_stride)
else:
labels = config.dataset.labels()
if not labels:
print('No labels specified.', file=sys.stderr)
return 1
ids = imagery_dataset.ImageryDataset(
images, labels, config.train.network.chunk_size(),
config.train.network.output_size(), tc.chunk_stride)
try:
if options.resume is not None:
model = tf.keras.models.load_model(options.resume,
custom_objects=ALL_LAYERS)
else:
model = config_model(ids.num_bands())
model, _ = train(model, ids, tc)
if options.model is not None:
model.save(options.model)
except KeyboardInterrupt:
print()
print('Training cancelled.')
return 0
def autoencoder(all_sources):
source = all_sources[0]
config.reset() # don't load any user files
(image_path, _) = source[0]
config.load(yaml_str=
'''
io:
cache:
dir: %s
dataset:
images:
type: %s
directory: %s
extension: %s
preprocess:
enabled: false
train:
network:
chunk_size: 3
mlflow:
enabled: false''' %
(os.path.dirname(image_path), source[2], os.path.dirname(image_path), source[1]))
dataset = imagery_dataset.AutoencoderDataset(config.dataset.images(),
config.train.network.chunk_size(), config.train.spec().chunk_stride)
return dataset
def main(options):
log_folder = config.dataset.log_folder()
if log_folder:
if not options.resume: # Start fresh and clear the read logs
os.system('rm ' + log_folder + '/*')
print('Dataset progress recording in: ' + log_folder)
else:
print('Resuming dataset progress recorded in: ' + log_folder)
start_time = time.time()
images = config.dataset.images()
if not images:
print('No images specified.', file=sys.stderr)
return 1
tc = config.train.spec()
if options.autoencoder:
ids = imagery_dataset.AutoencoderDataset(images, config.train.network.chunk_size(),
tc.chunk_stride, resume_mode=options.resume,
log_folder=log_folder)
else:
labels = config.dataset.labels()
if not labels:
print('No labels specified.', file=sys.stderr)
return 1
ids = imagery_dataset.ImageryDataset(images, labels, config.train.network.chunk_size(),
config.train.network.output_size(), tc.chunk_stride,
resume_mode=options.resume,
log_folder=log_folder)
try:
if options.resume is not None:
model = tf.keras.models.load_model(options.resume, custom_objects=ALL_LAYERS)
else:
model = config_model(ids.num_bands())
model, _ = train(model, ids, tc)
if options.model is not None:
save_model(model, options.model)
except KeyboardInterrupt:
print()
print('Training cancelled.')
stop_time = time.time()
print('Elapsed time = ', stop_time-start_time)
return 0
def autoencoder(all_sources):
source = all_sources[0]
conftest.config_reset()
(image_path, _) = source[0]
config.load(yaml_str='''
io:
cache:
dir: %s
dataset:
images:
type: %s
directory: %s
extension: %s
preprocess: ~''' %
(os.path.dirname(image_path), source[2],
os.path.dirname(image_path), source[1]))
dataset = imagery_dataset.AutoencoderDataset(
config.dataset.images(), (3, 3), stride=config.train.spec().stride)
return dataset
def main(options):
log_folder = config.train.log_folder()
if log_folder:
if not options.resume: # Start fresh and clear the read logs
os.system('rm -f ' + log_folder + '/*')
print('Dataset progress recording in: ' + log_folder)
else:
print('Resuming dataset progress recorded in: ' + log_folder)
images = config.dataset.images()
if not images:
print('No images specified.', file=sys.stderr)
return 1
img = images.load(0)
model = config_model(img.num_bands())
if options.resume is not None:
temp_model = tf.keras.models.load_model(options.resume, custom_objects=custom_objects())
else:
# this one is not built with proper scope, just used to get input and output shapes
temp_model = model()
start_time = time.time()
tile_size = config.io.tile_size()
tile_overlap = None
stride = config.train.spec().stride
# compute input and output sizes
if temp_model.input_shape[1] is None:
in_shape = None
out_shape = temp_model.compute_output_shape((0, tile_size[0], tile_size[1], temp_model.input_shape[3]))
out_shape = out_shape[1:3]
tile_overlap = (tile_size[0] - out_shape[0], tile_size[1] - out_shape[1])
else:
in_shape = temp_model.input_shape[1:3]
out_shape = temp_model.output_shape[1:3]
if options.autoencoder:
ids = imagery_dataset.AutoencoderDataset(images, in_shape, tile_shape=tile_size,
tile_overlap=tile_overlap, stride=stride)
else:
labels = config.dataset.labels()
if not labels:
print('No labels specified.', file=sys.stderr)
return 1
ids = imagery_dataset.ImageryDataset(images, labels, out_shape, in_shape,
tile_shape=tile_size, tile_overlap=tile_overlap,
stride=stride)
if log_folder is not None:
ids.set_resume_mode(options.resume, log_folder)
assert temp_model.input_shape[1] == temp_model.input_shape[2], 'Must have square chunks in model.'
assert temp_model.input_shape[3] == ids.num_bands(), 'Model takes wrong number of bands.'
tf.keras.backend.clear_session()
try:
model, _ = train(model, ids, config.train.spec(), options.resume)
if options.model is not None:
save_model(model, options.model)
except KeyboardInterrupt:
print('Training cancelled.')
stop_time = time.time()
print('Elapsed time = ', stop_time-start_time)
return 0
def main(options):
images = config.dataset.images()
if not images:
print('No images specified.', file=sys.stderr)
return 1
img = images.load(0)
model = config_model(img.num_bands())
temp_model = model()
tile_size = config.io.tile_size()
tile_overlap = None
stride = config.train.spec().stride
# compute input and output sizes
if temp_model.input_shape[1] is None:
in_shape = None
out_shape = temp_model.compute_output_shape((0, tile_size[0], tile_size[1], temp_model.input_shape[3]))
out_shape = out_shape[1:3]
tile_overlap = (tile_size[0] - out_shape[0], tile_size[1] - out_shape[1])
else:
in_shape = temp_model.input_shape[1:3]
out_shape = temp_model.output_shape[1:3]
if options.autoencoder:
ids = imagery_dataset.AutoencoderDataset(images, in_shape, tile_shape=tile_size,
tile_overlap=tile_overlap, stride=stride)
else:
labels = config.dataset.labels()
if not labels:
print('No labels specified.', file=sys.stderr)
return 1
ids = imagery_dataset.ImageryDataset(images, labels, out_shape, in_shape,
tile_shape=tile_size, tile_overlap=tile_overlap,
stride=stride)
assert temp_model.input_shape[1] == temp_model.input_shape[2], 'Must have square chunks in model.'
assert temp_model.input_shape[3] == ids.num_bands(), 'Model takes wrong number of bands.'
tf.keras.backend.clear_session()
#colormap = np.zeros(dtype=np.uint8, shape=(len(config.dataset.classes), 3))
#for c in config.dataset.classes:
# print(len(config.dataset.classes), c.value)
# colormap[c.value][0] = (c.color >> 32) & 0xFF
# colormap[c.value][1] = (c.color >> 16) & 0xFF
# colormap[c.value][2] = c.color & 0xFF
images = []
labels = []
PLOT_AT_ONCE=7
for result in ids.dataset(config.dataset.classes.weights(), config_augmentation()):
image = result[0].numpy()
label = result[1].numpy()
pw = (image.shape[0] - label.shape[0]) // 2
ph = (image.shape[1] - label.shape[1]) // 2
label = np.pad(label, ((pw, pw), (ph, ph), (0, 0)))
images.append(image)
labels.append(label)
if len(images) == PLOT_AT_ONCE:
plot_images(images, labels)
images = []
labels = []
return 0
def main(options):
if mixed_policy_device_compatible(
) and not config.train.disable_mixed_precision():
mixed_precision.set_global_policy('mixed_float16')
print(
'Tensorflow Mixed Precision is enabled. This improves training performance on compatible GPUs. '
'However certain precautions should be taken and several additional changes can be made to improve '
'performance further. Details: https://www.tensorflow.org/guide/mixed_precision#summary'
)
images = config.dataset.images()
if not images:
print('No images specified.', file=sys.stderr)
return 1
img = images.load(0)
model = config_model(img.num_bands())
if options.resume is not None and not options.resume.endswith('.h5'):
temp_model = load_model(options.resume)
else:
# this one is not built with proper scope, just used to get input and output shapes
temp_model = model()
start_time = time.time()
tile_size = config.io.tile_size()
tile_overlap = None
stride = config.train.spec().stride
# compute input and output sizes
if temp_model.input_shape[1] is None:
in_shape = None
out_shape = temp_model.compute_output_shape(
(0, tile_size[0], tile_size[1], temp_model.input_shape[3]))
out_shape = out_shape[1:3]
tile_overlap = (tile_size[0] - out_shape[0],
tile_size[1] - out_shape[1])
else:
in_shape = temp_model.input_shape[1:3]
out_shape = temp_model.output_shape[1:3]
if options.autoencoder:
ids = imagery_dataset.AutoencoderDataset(
images,
in_shape,
tile_shape=tile_size,
tile_overlap=tile_overlap,
stride=stride,
max_rand_offset=config.train.spec().max_tile_offset)
else:
labels = config.dataset.labels()
if not labels:
print('No labels specified.', file=sys.stderr)
return 1
ids = imagery_dataset.ImageryDataset(
images,
labels,
out_shape,
in_shape,
tile_shape=tile_size,
tile_overlap=tile_overlap,
stride=stride,
max_rand_offset=config.train.spec().max_tile_offset)
assert temp_model.input_shape[1] == temp_model.input_shape[
2], 'Must have square chunks in model.'
assert temp_model.input_shape[3] == ids.num_bands(
), 'Model takes wrong number of bands.'
tf.keras.backend.clear_session()
# Try to have the internal model format we use match the output model format
internal_model_extension = '.savedmodel'
if options.model and ('.h5' in options.model):
internal_model_extension = '.h5'
try:
model, _ = train(model, ids, config.train.spec(), options.resume,
internal_model_extension)
if options.model is not None:
save_model(model, options.model)
except KeyboardInterrupt:
print('Training cancelled.')
stop_time = time.time()
print('Elapsed time = ', stop_time - start_time)
return 0