def setup_losses(
gitapp: GetInputTargetAndPredictedParameters,
name: str = None,
) -> Tuple[Dict[str, lt.LabeledTensor], Dict[str, lt.LabeledTensor]]:
"""Creates cross entropy losses.
Args:
gitapp: GetInputTargetAndPredictedParameters.
name: Optional op name.
Returns:
A dictionary of tensors with the input reconstruction losses.
A dictionary of tensors with the target prediction losses.
"""
logging.info('Setting up losses')
with tf.name_scope(name, 'setup_losses', []) as scope:
(_, input_lt, target_lt, predict_input_lt,
predict_target_lt) = get_input_target_and_predicted(gitapp)
predicted_size = len(predict_input_lt.axes['row'])
visualize.summarize_image(
visualize.error_panel(util.crop_center(predicted_size, input_lt),
visualize.to_softmax(predict_input_lt),
name=scope + 'input_patch_error_panel'))
visualize.summarize_image(
visualize.error_panel(util.crop_center(predicted_size, target_lt),
visualize.to_softmax(predict_target_lt),
name=scope + 'target_patch_error_panel'))
def mean(lts: Dict[str, lt.LabeledTensor]) -> tf.Tensor:
sum_op = tf.add_n([t.tensor for t in lts.values()])
return sum_op / float(len(lts))
tag = 'input'
input_loss_lts = itemize_losses(gitapp.loss,
input_lt,
predict_input_lt,
name=scope + tag)
tf.summary.scalar(name='loss/' + tag, tensor=mean(input_loss_lts))
tag = 'target'
target_loss_lts = itemize_losses(gitapp.loss,
target_lt,
predict_target_lt,
name=scope + tag)
tf.summary.scalar(name='loss/' + tag, tensor=mean(target_loss_lts))
variables = tf.global_variables()
for v in variables:
tf.summary.histogram(name='variable/' + v.name, values=v)
return input_loss_lts, target_loss_lts
def get_statistics(t: lt.LabeledTensor) -> lt.LabeledTensor:
t = visualize.to_softmax(t)
rc = lt.ReshapeCoder(list(t.axes.keys())[:-1], ['batch'])
return rc.decode(ops.distribution_statistics(rc.encode(t)))
def setup_stitch(
gitapp: GetInputTargetAndPredictedParameters,
name=None,
) -> Dict[str, lt.LabeledTensor]:
"""Creates diagnostic images.
All diagnostic images are registered as summaries.
Args:
gitapp: GetInputTargetAndPredictedParameters.
name: Optional op name.
Returns:
A mapping where the keys are names of summary images and the values
are image tensors.
"""
logging.info('Setting up stitch')
with tf.name_scope(name, 'setup_stitch', []) as scope:
(patch_centers, input_lt, target_lt, predict_input_lt,
predict_target_lt) = get_input_target_and_predicted(gitapp)
predicted_size = len(predict_input_lt.axes['row'])
assert predicted_size == len(predict_input_lt.axes['column'])
input_lt = util.crop_center(predicted_size, input_lt)
target_lt = util.crop_center(predicted_size, target_lt)
# For now, we're not handling overlap or missing data.
assert gitapp.stride == predicted_size
if gitapp.bp is not None:
# Rebatch so a single tensor is all the patches in a single image.
[input_lt, target_lt, predict_input_lt,
predict_target_lt] = util.entry_point_batch(
[input_lt, target_lt, predict_input_lt, predict_target_lt],
bp=util.BatchParameters(size=len(patch_centers),
num_threads=1,
capacity=1),
enqueue_many=True,
entry_point_names=[
'input_stitch', 'target_stitch', 'predict_input_stitch',
'predict_target_stitch'
],
name='stitch')
rc = lt.ReshapeCoder(util.CANONICAL_AXIS_ORDER[3:], ['channel'])
input_lt = rc.decode(
ops.patches_to_image(patch_centers, rc.encode(input_lt)))
rc = lt.ReshapeCoder(util.CANONICAL_AXIS_ORDER[3:], ['channel'])
target_lt = rc.decode(
ops.patches_to_image(patch_centers, rc.encode(target_lt)))
rc = lt.ReshapeCoder(util.CANONICAL_PREDICTION_AXIS_ORDER[3:],
['channel'])
predict_input_lt = rc.decode(
ops.patches_to_image(patch_centers, rc.encode(predict_input_lt)))
rc = lt.ReshapeCoder(util.CANONICAL_PREDICTION_AXIS_ORDER[3:],
['channel'])
predict_target_lt = rc.decode(
ops.patches_to_image(patch_centers, rc.encode(predict_target_lt)))
def get_statistics(t: lt.LabeledTensor) -> lt.LabeledTensor:
t = visualize.to_softmax(t)
rc = lt.ReshapeCoder(list(t.axes.keys())[:-1], ['batch'])
return rc.decode(ops.distribution_statistics(rc.encode(t)))
# C++ entry points .
with tf.name_scope(''):
input_lt = lt.identity(input_lt, name='entry_point_stitched_input')
target_lt = lt.identity(target_lt,
name='entry_point_stitched_target')
# The nodes are used purely to export data to C++.
lt.identity(get_statistics(predict_input_lt),
name='entry_point_stitched_predicted_input')
lt.identity(get_statistics(predict_target_lt),
name='entry_point_stitched_predicted_target')
predict_input_lt = visualize.to_softmax(predict_input_lt)
predict_target_lt = visualize.to_softmax(predict_target_lt)
input_summary_lt = visualize.error_panel(input_lt, predict_input_lt)
target_summary_lt = visualize.error_panel(target_lt, predict_target_lt)
if gitapp.bp is not None:
input_summary_lt, target_summary_lt = lt.batch(
[input_summary_lt, target_summary_lt],
# We'll see 3 images in the visualizer.
batch_size=3,
enqueue_many=True,
num_threads=1,
capacity=1,
name='group')
input_summary_lt = lt.identity(input_summary_lt,
name=scope + 'input_error_panel')
target_summary_lt = lt.identity(target_summary_lt,
name=scope + 'target_error_panel')
visualize_op_dict = {}
visualize_op_dict['input'] = input_lt
visualize_op_dict['predict_input'] = predict_input_lt
visualize_op_dict['target'] = target_lt
visualize_op_dict['predict_target'] = predict_target_lt
def summarize(tag, labeled_tensor):
visualize.summarize_image(labeled_tensor,
name=scope + 'summarize/' + tag)
visualize_op_dict[tag] = labeled_tensor
summarize('input_error_panel', input_summary_lt)
summarize('target_error_panel', target_summary_lt)
return visualize_op_dict