def unet_multitask(config, train_mode):
if train_mode:
save_output = True
load_saved_output = False
preprocessing = preprocessing_multitask_train(config)
else:
save_output = True
load_saved_output = False
preprocessing = preprocessing_multitask_inference(config)
unet_multitask = Step(name='unet_multitask',
transformer=PyTorchUNetMultitask(**config.unet),
input_steps=[preprocessing],
cache_dirpath=config.env.cache_dirpath,
save_output=save_output,
load_saved_output=load_saved_output)
mask_resize = Step(name='mask_resize',
transformer=Resizer(),
input_data=['input'],
input_steps=[unet_multitask],
adapter={
'images':
([(unet_multitask.name, 'mask_prediction')]),
'target_sizes': ([('input', 'target_sizes')]),
},
cache_dirpath=config.env.cache_dirpath,
save_output=save_output)
contour_resize = Step(name='contour_resize',
transformer=Resizer(),
input_data=['input'],
input_steps=[unet_multitask],
adapter={
'images':
([(unet_multitask.name, 'contour_prediction')]),
'target_sizes': ([('input', 'target_sizes')]),
},
cache_dirpath=config.env.cache_dirpath,
save_output=save_output)
detached = Step(name='detached',
transformer=Postprocessor(),
input_steps=[mask_resize, contour_resize],
adapter={
'images': ([(mask_resize.name, 'resized_images')]),
'contours':
([(contour_resize.name, 'resized_images')]),
},
cache_dirpath=config.env.cache_dirpath,
save_output=save_output)
output = Step(name='output',
transformer=Dummy(),
input_steps=[detached],
adapter={
'y_pred': ([(detached.name, 'labeled_images')]),
},
cache_dirpath=config.env.cache_dirpath)
return output
def seq_conv_train(config):
xy_train = Step(name='xy_train',
transformer=XYSplit(**config.xy_splitter),
input_data=['input'],
adapter={'meta': ([('input', 'meta')]),
'train_mode': ([('input', 'train_mode')])
},
cache_dirpath=config.env.cache_dirpath)
xy_inference = Step(name='xy_inference',
transformer=XYSplit(**config.xy_splitter),
input_data=['input'],
adapter={'meta': ([('input', 'meta_valid')]),
'train_mode': ([('input', 'train_mode')])
},
cache_dirpath=config.env.cache_dirpath)
loader_train = Step(name='loader',
transformer=MetadataImageSegmentationLoader(**config.loader),
input_data=['input'],
input_steps=[xy_train, xy_inference],
adapter={'X': ([('xy_train', 'X')], squeeze_inputs),
'y': ([('xy_train', 'y')], squeeze_inputs),
'train_mode': ([('input', 'train_mode')]),
'X_valid': ([('xy_inference', 'X')], squeeze_inputs),
'y_valid': ([('xy_inference', 'y')], squeeze_inputs),
},
cache_dirpath=config.env.cache_dirpath)
sequential_convnet = Step(name='sequential_convnet',
transformer=SequentialConvNet(**config.sequential_convnet),
input_steps=[loader_train],
cache_dirpath=config.env.cache_dirpath)
mask_resize = Step(name='mask_resize',
transformer=Resizer(),
input_data=['input'],
input_steps=[sequential_convnet],
adapter={'images': ([('sequential_convnet', 'predicted_masks')]),
'target_sizes': ([('input', 'target_sizes')]),
},
cache_dirpath=config.env.cache_dirpath)
thresholding = Step(name='thresholding',
transformer=Thresholder(**config.thresholder),
input_steps=[mask_resize],
adapter={'images': ([('mask_resize', 'resized_images')]),
},
cache_dirpath=config.env.cache_dirpath)
output = Step(name='output',
transformer=Dummy(),
input_steps=[thresholding],
adapter={'y_pred': ([('thresholding', 'binarized_images')]),
},
cache_dirpath=config.env.cache_dirpath)
return output
def unet_inference(config):
xy_inference = Step(name='xy_inference',
transformer=XYSplit(**config.xy_splitter),
input_data=['input'],
adapter={'meta': ([('input', 'meta')]),
'train_mode': ([('input', 'train_mode')])
},
cache_dirpath=config.env.cache_dirpath)
loader_inference = Step(name='loader',
transformer=MetadataImageSegmentationLoader(**config.loader),
input_data=['input'],
input_steps=[xy_inference, xy_inference],
adapter={'X': ([('xy_inference', 'X')], squeeze_inputs),
'y': ([('xy_inference', 'y')], squeeze_inputs),
'train_mode': ([('input', 'train_mode')]),
},
cache_dirpath=config.env.cache_dirpath)
unet_network = Step(name='unet_network',
transformer=PyTorchUNet(**config.unet_network),
input_steps=[loader_inference],
cache_dirpath=config.env.cache_dirpath)
mask_resize = Step(name='mask_resize',
transformer=Resizer(),
input_data=['input'],
input_steps=[unet_network],
adapter={'images': ([('unet_network', 'predicted_masks')]),
'target_sizes': ([('input', 'target_sizes')]),
},
cache_dirpath=config.env.cache_dirpath)
thresholding = Step(name='thresholding',
transformer=Thresholder(**config.thresholder),
input_steps=[mask_resize],
adapter={'images': ([('mask_resize', 'resized_images')]),
},
cache_dirpath=config.env.cache_dirpath)
output = Step(name='output',
transformer=Dummy(),
input_steps=[thresholding],
adapter={'y_pred': ([('thresholding', 'binarized_images')]),
},
cache_dirpath=config.env.cache_dirpath)
return output
def mask_postprocessing(model, config, save_output=False):
mask_resize = Step(name='mask_resize',
transformer=Resizer(),
input_data=['input'],
input_steps=[model],
adapter={
'images':
([(model.name, 'multichannel_map_prediction')]),
'target_sizes': ([('input', 'target_sizes')]),
},
cache_dirpath=config.env.cache_dirpath,
save_output=save_output)
category_mapper = Step(name='category_mapper',
transformer=CategoryMapper(),
input_steps=[mask_resize],
adapter={
'images': ([('mask_resize', 'resized_images')]),
},
cache_dirpath=config.env.cache_dirpath,
save_output=save_output)
return category_mapper
def center_postprocessing(model, config, save_output=True):
center_resize = Step(name='center_resize',
transformer=Resizer(),
input_data=['input'],
input_steps=[model],
adapter={
'images': ([(model.name, 'center_prediction')]),
'target_sizes': ([('input', 'target_sizes')]),
},
cache_dirpath=config.env.cache_dirpath,
save_output=save_output)
center_thresholding = Step(name='center_thresholding',
transformer=Thresholder(**config.thresholder),
input_steps=[center_resize],
adapter={
'images':
([('center_resize', 'resized_images')]),
},
cache_dirpath=config.env.cache_dirpath,
save_output=save_output)
return center_thresholding
def mask_postprocessing(loader, model, config, save_output=False):
if config.postprocessor.crf.apply_crf:
dense_crf = Step(name='dense_crf',
transformer=DenseCRFStream(**config.postprocessor.crf) if config.execution.stream_mode \
else DenseCRF(**config.postprocessor.crf),
input_steps=[loader, model],
adapter={'images': ([(model.name, 'multichannel_map_prediction')]),
'raw_images_generator': ([(loader.name, 'datagen')]),
},
cache_dirpath=config.env.cache_dirpath,
save_output=save_output)
mask_resize = Step(name='mask_resize',
transformer=ResizerStream() if config.execution.stream_mode else Resizer(),
input_data=['input'],
input_steps=[dense_crf],
adapter={'images': ([('dense_crf', 'crf_images')]),
'target_sizes': ([('input', 'target_sizes')]),
},
cache_dirpath=config.env.cache_dirpath,
save_output=save_output)
else:
mask_resize = Step(name='mask_resize',
transformer=ResizerStream() if config.execution.stream_mode else Resizer(),
input_data=['input'],
input_steps=[model],
adapter={'images': ([(model.name, 'multichannel_map_prediction')]),
'target_sizes': ([('input', 'target_sizes')]),
},
cache_dirpath=config.env.cache_dirpath,
save_output=save_output)
category_mapper = Step(name='category_mapper',
transformer=CategoryMapperStream() if config.execution.stream_mode else CategoryMapper(),
input_steps=[mask_resize],
adapter={'images': ([('mask_resize', 'resized_images')]),
},
cache_dirpath=config.env.cache_dirpath,
save_output=save_output)
mask_erosion = Step(name='mask_erosion',
transformer=MaskEroderStream(**config.postprocessor) if config.execution.stream_mode
else MaskEroder(**config.postprocessor),
input_steps=[category_mapper],
adapter={'images': ([(category_mapper.name, 'categorized_images')]),
},
cache_dirpath=config.env.cache_dirpath,
load_saved_output=False)
detached = multiclass_object_labeler(mask_erosion, config, save_output=save_output)
mask_dilation = Step(name='mask_dilation',
transformer=LabeledMaskDilatorStream(**config.postprocessor) if config.execution.stream_mode
else LabeledMaskDilator(**config.postprocessor),
input_steps=[detached],
adapter={'images': ([(detached.name, 'labeled_images')]),
},
cache_dirpath=config.env.cache_dirpath,
load_saved_output=False)
score_builder = Step(name='score_builder',
transformer=ScoreBuilder(),
input_steps=[mask_dilation, mask_resize],
adapter={'images': ([(mask_dilation.name, 'dilated_images')]),
'probabilities': ([(mask_resize.name, 'resized_images')]),
},
cache_dirpath=config.env.cache_dirpath,
save_output=save_output)
return score_builder