def test_set_normalization_if_no_data(self):
img_path = os.path.join(base_path, '../test_data/shapes/pixels/*')
label_path = os.path.join(base_path, '../test_data/shapes/labels.ilp')
t = Session()
# t.load_training_data(img_path, label_path)
t.set_normalization('local')
def test_set_normalization(self):
img_path = os.path.join(base_path, '../test_data/shapes/pixels/*')
label_path = os.path.join(base_path, '../test_data/shapes/labels.ilp')
t = Session()
t.load_training_data(img_path, label_path)
t.make_model('unet_2d', (1, 572, 572))
t.set_normalization('local')
assert t.data.normalize_mode == 'local'
assert t.data.global_norm_minmax is None
t.set_normalization('global_0+255')
assert t.data.normalize_mode == 'global'
assert t.data.global_norm_minmax == (0, 255)
def test_set_normalization_if_no_data(self):
t = Session()
# t.load_training_data(img_path, label_path)
t.set_normalization('local')
class DeepimagejExporter(object):
'''
A DeepImageJ exporter provides methods to deploy Keras models
trained with YAPiC to the ImageJ plugin DeepImageJ.
Parameters
----------
model_path : string
Path to Keras model in h5 format.
save_path : string
Path to directory where the exported model and metadata is saved.
example_image_path: string
Path to example input image in tif format.
'''
def __init__(self, model_path, save_path, example_image_path):
save_dir = os.path.dirname(save_path)
if len(save_dir) == 0:
save_dir = './'
msg = '{} does not exist'.format(save_dir)
assert os.path.isdir(save_dir), msg
self.save_path = save_path
self.s = Session()
self.s.load_prediction_data(example_image_path, self.save_path)
self.s.load_model(model_path)
self.model_path = model_path
self.example_image_path = example_image_path
msg = 'model is not unet_2d, cannot be exported to deepimagej'
assert self._is_model_unet_2d(), msg
self.model_reshaped = None
self.metadata = None
self.template_dir = os.path.join(os.path.dirname(__file__),
'templates/deepimagej101')
def export_as_deepimagej(self,
author='n/a',
url='http://',
credit='n.a',
version='n.a',
reference='n/a',
size='small',
applymodel=True):
self._reshape_unet_2d(size=size)
self._update_metadata(author=author,
version=version,
url=url,
credit=credit,
reference=reference)
self._export_as_tensorflow_model()
self._format_xml()
shutil.copyfile(os.path.join(self.template_dir, 'postprocessing.txt'),
os.path.join(self.save_path, 'postprocessing.txt'))
shutil.copyfile(os.path.join(self.template_dir, 'preprocessing.txt'),
os.path.join(self.save_path, 'preprocessing.txt'))
if applymodel:
self.apply_model('local')
shutil.copyfile(self.example_image_path,
os.path.join(self.save_path, 'exampleImage.tiff'))
def apply_model(self, normalization_mode):
self.s.set_normalization(normalization_mode)
self.s.predict(multichannel=True)
result_img_name = os.path.basename(self.example_image_path)
save_path = os.path.join(self.save_path, result_img_name)
new_save_path = save_path.replace(result_img_name, 'resultImage.tiff')
os.rename(save_path, new_save_path)
def _is_model_unet_2d(self):
return self.s.model.name == 'unet_2d'
# return self.s.model.count_params() == 32424323
def _reshape_unet_2d(self, size='middle'):
if size == 'small':
shape_xy = 112
elif size == 'middle':
shape_xy = 224
elif size == 'large':
shape_xy = 368
else:
shape_xy = 112
print('reshape to {}'.format(shape_xy))
N_classes = self.s.model.output_shape[-1]
N_channels = self.s.model.input_shape[-1]
self.model_reshaped = unet_2d.build_network(
N_classes, (N_channels, 1, shape_xy, shape_xy),
squeeze=True,
padding='same')
self.model_reshaped.set_weights(self.s.model.get_weights())
def _update_metadata(self,
author='n/a',
url='http://',
credit='n.a',
version='n.a',
reference='n/a'):
if self.model_reshaped is None:
return
if self.metadata is None:
self.metadata = {}
self.metadata['name'] = os.path.basename(self.save_path)
self.metadata['author'] = author
self.metadata['url'] = url
self.metadata['credit'] = credit
self.metadata['version'] = version
self.metadata['reference'] = reference
date_format = '%a %b %d %H:%M:%S %Z %Y'
self.metadata['date'] = time.strftime(date_format, time.localtime())
N_channels = self.model_reshaped.input_shape[-1]
size_xy = self.model_reshaped.input_shape[2]
self.metadata['channels'] = N_channels
self.metadata['input_tensor_dimensions'] = (-1, size_xy, size_xy,
N_channels)
self.metadata['patch_size'] = size_xy
self.metadata['padding'] = int(size_xy * 0.19)
# metadata = {'name': 'my_model',
# 'author': 'n/a',
# 'url': 'http://',
# 'credit': 'n/a',
# 'version': 'n/a',
# 'reference': 'n/a',
# 'date': 'Tue Mar 31 17:18:06 CEST 2020',
# 'test_image_size_xy': (512, 512),
# 'input_tensor_dimensions': (-1, 112, 112, 3),
# 'patch_size': (112),
# 'padding': 10}
def _format_xml(self):
if self.metadata is None:
return
xml_path = os.path.join(self.template_dir, 'config.xml')
tree = ET.parse(xml_path)
key_mapping = (
(('ModelInformation', 'Name'), 'name'),
(('ModelInformation', 'Author'), 'author'),
(('ModelInformation', 'URL'), 'url'),
(('ModelInformation', 'Credit'), 'credit'),
(('ModelInformation', 'Version'), 'version'),
(('ModelInformation', 'Date'), 'date'),
(('ModelInformation', 'Reference'), 'reference'),
(('ModelCharacteristics', 'Channels'), 'channels'),
(('ModelCharacteristics', 'InputTensorDimensions'),
'input_tensor_dimensions'),
(('ModelCharacteristics', 'PatchSize'), 'patch_size'),
(('ModelCharacteristics', 'Padding'), 'padding'),
)
for item in key_mapping:
value = str(self.metadata[item[1]])
if item[1] == 'input_tensor_dimensions':
value = value.replace('(', ',')\
.replace(')', ',')\
.replace(' ', '')
tree.find(item[0][0]).find(item[0][1]).text = value
save_path = os.path.join(self.save_path, 'config.xml')
tree.write(save_path)
def _export_as_tensorflow_model(self):
model = self.model_reshaped
builder = tf.saved_model.builder.SavedModelBuilder(self.save_path)
signature = tf.saved_model.signature_def_utils.predict_signature_def(
inputs={'input': model.input}, outputs={'output': model.output})
signature_def_map = {
tf.saved_model.signature_constants.DEFAULT_SERVING_SIGNATURE_DEF_KEY:
signature
}
builder.add_meta_graph_and_variables(
K.get_session(), [tf.saved_model.tag_constants.SERVING],
signature_def_map=signature_def_map)
builder.save()
def main(args):
if args['--cpu']:
# deactivate gpu for tensorflow
os.environ['CUDA_VISIBLE_DEVICES'] = ''
if args['--gpu']:
# define gpu hardware
os.environ['CUDA_VISIBLE_DEVICES'] = args['--gpu']
image_path = os.path.abspath(os.path.expanduser(args['<image_path>']))
model_name = args['<network>']
if args['deploy']:
output_path = args['<output_path>']
predict_example_image = not args['--skip-predict']
assert os.path.isfile(model_name)
exp = DeepimagejExporter(model_name, output_path, image_path)
exp.export_as_deepimagej(author=args['--author'],
version=args['--version'],
url=args['--url'],
credit=args['--credit'],
reference=args['--reference'],
size=args['--size'],
applymodel=predict_example_image)
return
norm_string = args['--normalize']
s = Session()
if args['train']:
label_path = os.path.abspath(os.path.expanduser(args['<label_path>']))
aug_string = args['--augment']
max_epochs = int(args['--epochs'])
steps_per_epoch = int(args['--steps'])
log_filename = args['--csvfile']
model_export_filename = args['--file']
valfraction = float(args['--valfraction'])
if args['--batchsize']:
batch_size = int(args['--batchsize'])
else:
batch_size = None
s.load_training_data(image_path, label_path, batch_size)
models_available = ['unet_2d', 'unet_multi_z', 'convnet_for_unittest']
if os.path.isfile(model_name):
s.load_model(model_name)
elif model_name in models_available:
size_xy = 572
if model_name == 'unet_2d' or model_name == 'convnet_for_unittest':
size_z = 1
if model_name == 'unet_multi_z':
size_z = 5
if model_name == 'convnet_for_unittest':
size_xy = 100
s.make_model(model_name, (size_z, size_xy, size_xy))
s.set_normalization(norm_string)
s.set_augmentation(aug_string)
if valfraction > 0:
s.define_validation_data(valfraction)
s.train(max_epochs=max_epochs,
steps_per_epoch=steps_per_epoch,
log_filename=log_filename,
model_filename=model_export_filename)
f = h5py.File(model_export_filename, 'r+')
lbl_map_list = [[key, val] for key, val in s.lbl_map.items()]
f.create_dataset('lbl_map', data=lbl_map_list)
if args['predict']:
output_path = args['<output_path>']
assert os.path.isfile(model_name), '<network> must be a h5 model file'
multichannel = not (args['--split'])
s.load_prediction_data(image_path, output_path)
s.load_model(model_name)
s.set_normalization(norm_string)
s.predict(multichannel=multichannel)
def main(args):
s = Session()
if args['--cpu']:
# deactivate gpu for tensorflow
os.environ['CUDA_VISIBLE_DEVICES'] = ''
if args['--gpu']:
# define gpu hardware
os.environ['CUDA_VISIBLE_DEVICES'] = args['--gpu']
image_path = os.path.abspath(os.path.expanduser(args['<image_path>']))
model_name = args['<network>']
norm_string = args['--normalize']
if args['train']:
label_path = os.path.abspath(os.path.expanduser(args['<label_path>']))
aug_string = args['--augment']
max_epochs = int(args['--epochs'])
steps_per_epoch = int(args['--steps'])
log_filename = args['--csvfile']
model_export_filename = args['--file']
valfraction = float(args['--valfraction'])
s.load_training_data(image_path, label_path)
models_available = ['unet_2d',
'unet_multi_z',
'convnet_for_unittest']
if os.path.isfile(model_name):
s.load_model(model_name)
elif model_name in models_available:
size_xy = 572
if model_name == 'unet_2d' or model_name == 'convnet_for_unittest':
size_z = 1
if model_name == 'unet_multi_z':
size_z = 5
if model_name == 'convnet_for_unittest':
size_xy = 100
s.make_model(model_name, (size_z, size_xy, size_xy))
s.set_normalization(norm_string)
s.set_augmentation(aug_string)
if valfraction > 0:
s.define_validation_data(valfraction)
s.train(max_epochs=max_epochs,
steps_per_epoch=steps_per_epoch,
log_filename=log_filename,
model_filename=model_export_filename)
if args['predict']:
output_path = args['<output_path>']
assert os.path.isfile(model_name), '<network> must be a h5 model file'
s.load_prediction_data(image_path, output_path)
s.load_model(model_name)
s.set_normalization(norm_string)
s.predict()