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 get_statistics(tensor): rc = lt.ReshapeCoder(list(tensor.axes.keys())[:-1], ['batch']) return rc.decode(ops.distribution_statistics(rc.encode(tensor)))
def error_panel(
target_lt: lt.LabeledTensor,
predicted_lt: lt.LabeledTensor,
name: str = None,
) -> lt.LabeledTensor:
"""Creates a big informative error panel image.
Args:
target_lt: The ground truth values in canonical order.
predicted_lt: The predicted values in canonical prediction order as a
probability distribution.
name: Optional op name.
Returns:
The error panel.
"""
with tf.name_scope(name, 'error_panel',
[target_lt, predicted_lt]) as scope:
target_lt = lt.transpose(target_lt, util.CANONICAL_AXIS_ORDER)
predicted_lt = lt.transpose(predicted_lt,
util.CANONICAL_PREDICTION_AXIS_ORDER)
assert list(target_lt.axes.items())[:-1] == list(
predicted_lt.axes.items())[:-1], (target_lt.axes,
predicted_lt.axes)
rc = lt.ReshapeCoder(list(predicted_lt.axes.keys())[:-1], ['batch'])
statistic_lt = rc.decode(
ops.distribution_statistics(rc.encode(predicted_lt)))
columns = []
def get_column(
labeled_tensor: lt.LabeledTensor,
color: Tuple[float, float, float],
):
labeled_tensor = add_border(color, PAD_WIDTH, labeled_tensor)
labeled_tensor = lt.transpose(
labeled_tensor,
['batch', 'z', 'channel', 'row', 'column', 'color'])
columns.append(
lt.reshape(labeled_tensor, ['z', 'channel', 'row'], ['row']))
# We only show these statistics.
statistics = [
'mode', 'median', 'mean', 'standard_deviation', 'entropy'
]
# Show the statistics on the predictions.
for s in statistics:
get_column(lt.select(statistic_lt, {'statistic': s}), PURPLE)
# Show the ground truth target image.
get_column(lt.select(target_lt, {'mask': False}), TURQUOISE)
# Show the cross entropy error.
num_classes = len(predicted_lt.axes['class'])
cross_entropy_lt = cross_entropy_error(
util.onehot(num_classes, lt.select(target_lt, {'mask': False})),
predicted_lt)
get_column(cross_entropy_lt, RED)
# Show the additive error visualizations.
for s in statistics[:3]:
rc = lt.ReshapeCoder(['z', 'channel'], ['channel'])
error_lt = rc.decode(
additive_error(
rc.encode(lt.select(target_lt, {'mask': False})),
rc.encode(lt.select(statistic_lt, {'statistic': s}))))
get_column(error_lt, WHITE)
# Show the subtractive error visualizations.
for s in statistics[:3]:
rc = lt.ReshapeCoder(['z', 'channel'], ['channel'])
error_lt = rc.decode(
subtractive_error(
rc.encode(lt.select(target_lt, {'mask': False})),
rc.encode(lt.select(statistic_lt, {'statistic': s}))))
get_column(error_lt, BLACK)
# Show the pixel presence / absence masks.
get_column(lt.select(target_lt, {'mask': True}), MANGO)
panel_lt = lt.concat(columns, 'column', name=scope)
return panel_lt