def add_patch(self,
labels,
predicted,
weights,
coord=None,
volname=None,
patches=None):
"""Evaluates single-object segmentation quality."""
predicted = mask.crop_and_pad(predicted, (0, 0, 0), self._eval_shape)
weights = mask.crop_and_pad(weights, (0, 0, 0), self._eval_shape)
labels = mask.crop_and_pad(labels, (0, 0, 0), self._eval_shape)
loss, = self.sess.run([self.eval_loss], {
self.eval_labels: labels,
self.eval_preds: predicted
})
self.loss += loss
self.total_voxels += labels.size
self.masked_voxels += np.sum(weights == 0.0)
pred_mask = predicted >= self.eval_threshold
true_mask = labels > 0.5
pred_bg = np.logical_not(pred_mask)
true_bg = np.logical_not(true_mask)
self.tp += np.sum(pred_mask & true_mask)
self.fp += np.sum(pred_mask & true_bg)
self.fn += np.sum(pred_bg & true_mask)
self.tn += np.sum(pred_bg & true_bg)
self.num_patches += 1
predicted = expit(predicted)
self.images_xy.append(self.slice_image(labels, predicted, weights, 0))
self.images_xz.append(self.slice_image(labels, predicted, weights, 1))
self.images_yz.append(self.slice_image(labels, predicted, weights, 2))
def add_patch(self, labels, predicted, weights,
coord=None, volname=None, patches=None):
"""Evaluates single-object segmentation quality."""
options = tf.RunOptions(trace_level=tf.RunOptions.FULL_TRACE) #profiling
run_metadata = tf.RunMetadata()
predicted = mask.crop_and_pad(predicted, (0, 0, 0), self._eval_shape)
weights = mask.crop_and_pad(weights, (0, 0, 0), self._eval_shape)
labels = mask.crop_and_pad(labels, (0, 0, 0), self._eval_shape)
loss, = self.sess.run([self.eval_loss], {self.eval_labels: labels,
self.eval_preds: predicted},options=options, run_metadata=run_metadata)
fetched_timeline = timeline.Timeline(run_metadata.step_stats)
chrome_trace = fetched_timeline.generate_chrome_trace_format()
with open('timeline_01.json', 'w') as f:
f.write(chrome_trace)
self.loss += loss
self.total_voxels += labels.size
self.masked_voxels += np.sum(weights == 0.0)
pred_mask = predicted >= self.eval_threshold
true_mask = labels > 0.5
pred_bg = np.logical_not(pred_mask)
true_bg = np.logical_not(true_mask)
self.tp += np.sum(pred_mask & true_mask)
self.fp += np.sum(pred_mask & true_bg)
self.fn += np.sum(pred_bg & true_mask)
self.tn += np.sum(pred_bg & true_bg)
self.num_patches += 1
predicted = expit(predicted)
self.images_xy.append(self.slice_image(labels, predicted, weights, 0))
self.images_xz.append(self.slice_image(labels, predicted, weights, 1))
self.images_yz.append(self.slice_image(labels, predicted, weights, 2))
def get_example(load_example, eval_tracker, model, get_offsets):
"""Generates individual training examples.
Args:
load_example: callable returning a tuple of image and label ndarrays
as well as the seed coordinate and volume name of the example
eval_tracker: EvalTracker object
model: FFNModel object
get_offsets: iterable of (x, y, z) offsets to investigate within the
training patch
Yields:
tuple of:
seed array, shape [1, z, y, x, 1]
image array, shape [1, z, y, x, 1]
label array, shape [1, z, y, x, 1]
"""
seed_shape = train_canvas_size(model).tolist()[::-1]
while True:
full_patches, full_labels, loss_weights, coord, volname = load_example(
)
# Write a random fraction of paired examples to images and make sure they have
# matching and correct orientations.
if FLAGS.debug:
if random.uniform(0, 1) > 0.999:
write_patch_and_label_to_img(
patch=full_patches[0, 0, :, :, 0] * FLAGS.image_stddev +
FLAGS.image_mean,
label=full_labels[0, 0, :, :, 0] * 255,
unique_id='_'.join(coord[0].astype(str).tolist()),
dirname="./debug")
# Always start with a clean seed.
seed = logit(mask.make_seed(seed_shape, 1, pad=FLAGS.seed_pad))
for off in get_offsets(model, seed):
predicted = mask.crop_and_pad(seed, off,
model.input_seed_size[::-1])
patches = mask.crop_and_pad(full_patches, off,
model.input_image_size[::-1])
labels = mask.crop_and_pad(full_labels, off,
model.pred_mask_size[::-1])
weights = mask.crop_and_pad(loss_weights, off,
model.pred_mask_size[::-1])
# Necessary, since the caller is going to update the array and these
# changes need to be visible in the following iterations.
assert predicted.base is seed
yield predicted, patches, labels, weights
# TODO(jpgard): track volname in eval_tracker. Currently nothing is done with
# volname, but it should be monitored to ensure coverage of all training
# volumes. Similar for coord; would be good to check coordinates covered,
# or at least a sample of them.
eval_tracker.add_patch(full_labels, seed, loss_weights, coord, volname,
full_patches)
def get_example(load_example, eval_tracker, model, get_offsets):
"""Generates individual training examples.
Args:
load_example: callable returning a tuple of image and label ndarrays
as well as the seed coordinate and volume name of the example
eval_tracker: EvalTracker object
model: FFNModel object
get_offsets: iterable of (x, y, z) offsets to investigate within the
training patch
Yields:
tuple of:
seed array, shape [1, z, y, x, 1]
image array, shape [1, z, y, x, 1]
label array, shape [1, z, y, x, 1]
"""
seed_shape = train_canvas_size(model).tolist()[::-1]
while True:
full_patches, full_labels, loss_weights, coord, volname = load_example(
)
# Always start with a clean seed.
seed = logit(mask.make_seed(seed_shape, 1, pad=FLAGS.seed_pad))
for off in get_offsets(model, seed):
predicted = mask.crop_and_pad(seed, off,
model.input_seed_size[::-1])
patches = mask.crop_and_pad(full_patches, off,
model.input_image_size[::-1])
labels = mask.crop_and_pad(full_labels, off,
model.pred_mask_size[::-1])
weights = mask.crop_and_pad(loss_weights, off,
model.pred_mask_size[::-1])
# Necessary, since the caller is going to update the array and these
# changes need to be visible in the following iterations.
assert predicted.base is seed
yield predicted, patches, labels, weights
# eval_tracker.add_patch(
# full_labels, seed, loss_weights, coord, volname, full_patches)
eval_tracker.add_patch_v2(full_labels, seed, loss_weights, coord,
volname, full_patches)
def max_pred_offsets(model, seed):
"""Generates offsets with the policy used for inference."""
# Always start at the center.
queue = deque([(0, 0, 0)])
done = set()
train_image_radius = train_image_size(model) // 2
input_image_radius = np.array(model.input_image_size) // 2
while queue:
offset = queue.popleft()
# Drop any offsets that would take us beyond the image fragment we
# loaded for training.
if np.any(np.abs(np.array(offset)) + input_image_radius >
train_image_radius):
continue
# Ignore locations that were visited previously.
quantized_offset = (
offset[0] // max(model.deltas[0], 1),
offset[1] // max(model.deltas[1], 1),
offset[2] // max(model.deltas[2], 1))
if quantized_offset in done:
continue
done.add(quantized_offset)
yield offset
# Look for new offsets within the updated seed.
curr_seed = mask.crop_and_pad(seed, offset, model.pred_mask_size[::-1])
todos = sorted(
movement.get_scored_move_offsets(
model.deltas[::-1],
curr_seed[0, ..., 0],
threshold=logit(FLAGS.threshold)), reverse=True)
queue.extend((x[2] + offset[0],
x[1] + offset[1],
x[0] + offset[2]) for _, x in todos)
