def create_input_tensors_dict(self, batch_size):
if self.batching_mode == "pair":
imgs, labels, tags, original_classes = self._create_inputs_for_pair(
batch_size)
imgs_raw = unnormalize(imgs)
elif self.batching_mode == "group":
imgs, labels, tags, original_images = self._create_inputs_for_group(
batch_size)
original_classes = labels
imgs_raw = original_images
# imgs_raw = None
elif self.batching_mode == "eval":
imgs, imgs_raw, tags, labels = self._create_inputs_for_eval(
batch_size)
original_classes = labels
else:
raise ValueError("Incorrect batching mode error")
#summary = tf.get_collection(tf.GraphKeys.SUMMARIES)[-1]
# self.summaries.append(summary)
if self.use_summaries:
summ = tf.summary.image("imgs", unnormalize(imgs))
self.summaries.append(summ)
tensors = {"inputs": imgs, "labels": labels,
"tags": tags, "original_labels": original_classes}
if imgs_raw is not None:
tensors["imgs_raw"] = imgs_raw
return tensors
def create_input_tensors_dict(self, batch_size):
self._load_inputfile_lists()
resize_mode, input_size = self._get_resize_params(self.subset, self.image_size, ResizeMode.Unchanged)
augmentors, shuffle = self._parse_augmentors_and_shuffle()
inputfile_tensors = [tf.convert_to_tensor(l, dtype=tf.string) for l in self.inputfile_lists]
queue = tf.train.slice_input_producer(inputfile_tensors, shuffle=shuffle)
tensors_dict, summaries = self._read_inputfiles(queue, resize_mode, input_size, augmentors)
tensors_dict = create_batch_dict(batch_size, tensors_dict)
if self.use_summaries:
inputs = tensors_dict["inputs"]
input_img = unnormalize(inputs[:, :, :, :3])
# Add clicks to the image so that they can be viewed in tensorboard
if inputs.get_shape()[-1] > 4:
[input_img] = tf.py_func(self.add_clicks, [tf.concat([input_img, inputs[:, :, :, 3:4]],
axis=3), 'r'], [tf.float32])
[input_img] = tf.py_func(self.add_clicks, [tf.concat([input_img, inputs[:, :, :, 4:5]],
axis=3), 'g'], [tf.float32])
summ0 = tf.summary.image("inputs", input_img)
summ1 = tf.summary.image("labels", tensors_dict["labels"] * 255) # will only work for binary segmentation
summaries = [summ0, summ1]
# count is incremented after each summary creation. This helps us to keep track of the number of channels
count = 0
# Old label would be present either as a single channel, or along with 2 other distance transform channels.
if inputs.get_shape()[-1] == 4 or inputs.get_shape == 6:
summ2 = tf.summary.image("old_labels", inputs[:, :, :, 3:4] * 255)
summaries.append(summ2)
count += 1
# Append the distance transforms, if they are available.
if inputs.get_shape()[-1] > 4:
# Get negative distance transform from the extra input channels.
start = 3 + count
end = start + 1
start = tf.constant(start)
end = tf.constant(end)
summ3 = tf.summary.image(Constants.DT_NEG, inputs[:, :, :, 3:4])
start = end
end = start + 1
summ4 = tf.summary.image(Constants.DT_POS, inputs[:, :, :, 4:5])
summaries.append(summ3)
summaries.append(summ4)
self.summaries += summaries
return tensors_dict
def create_input_tensors_dict(self, batch_size):
use_index_img = self.subset != "train"
tensors = create_tensor_dict(
unnormalized_img=self.img_placeholder,
label=self.label_placeholder,
tag=self.tag_placeholder,
raw_label=self.label_placeholder,
old_label=self.old_label_placeholder,
flow_past=self.flow_into_past_placeholder,
flow_future=self.flow_into_future_placeholder,
use_index_img=use_index_img,
u0=self.u0_placeholder,
u1=self.u1_placeholder)
# TODO: need to set shape here?
resize_mode, input_size = self._get_resize_params(
self.subset, self.image_size)
tensors = resize(tensors, resize_mode, input_size)
if len(input_size) == 3:
input_size = input_size[1:]
tensors = self._prepare_augmented_batch(tensors,
batch_size,
image_size=input_size)
if self.use_summaries:
inputs = tensors["inputs"]
summ0 = tf.summary.image("inputs",
unnormalize(inputs[:, :, :, :3]))
summ1 = tf.summary.image(
"labels", tensors["labels"] *
255) # will only work well for binary segmentation
self.summaries.append(summ0)
self.summaries.append(summ1)
# Old label would be present either as a single channel, or along with 2 other distance transform channels.
if inputs.get_shape()[-1] == 4 or inputs.get_shape == 6:
summ2 = tf.summary.image("old_labels",
inputs[:, :, :, 3:4] * 255)
self.summaries.append(summ2)
return tensors
def create_input_tensors_dict(self, batch_size):
use_index_img = self.subset != "train"
tensors = create_tensor_dict(
unnormalized_img=self.img_placeholder,
label=self.label_placeholder,
tag=self.tag_placeholder,
raw_label=self.label_placeholder,
old_label=self.old_label_placeholder,
flow_past=self.flow_into_past_placeholder,
flow_future=self.flow_into_future_placeholder,
use_index_img=use_index_img,
u0=self.u0_placeholder,
u1=self.u1_placeholder,
flow=self.flow_placeholder)
# TODO: need to set shape here?
resize_mode, input_size = self._get_resize_params(
self.subset, self.image_size)
tensors = resize(tensors, resize_mode, input_size)
if len(input_size) == 3:
input_size = input_size[1:]
tensors = self._prepare_augmented_batch(tensors,
batch_size,
image_size=input_size)
if self.use_summaries:
inputs = tensors["inputs"]
summ0 = tf.summary.image("inputs",
unnormalize(inputs[:, :, :, :3]))
summ1 = tf.summary.image(
"labels", tensors["labels"] *
255) # will only work well for binary segmentation
self.summaries.append(summ0)
self.summaries.append(summ1)
return tensors
def _create_summaries(self, tensors_dict):
inputs = tensors_dict["inputs"]
input_imgs = unnormalize(inputs[:, :, :, :3])
# count is incremented after each summary creation. This helps us to keep track of the number of channels
start = 3
summaries = []
# Old label would be present either as a single channel, or along with 2 other distance transform channels.
if inputs.get_shape()[-1] in [4, 6, 8]:
if Constants.OLD_LABEL_AS_DT in tensors_dict:
[input_imgs] = tf.py_func(self.add_clicks, [
tf.concat([input_imgs, inputs[:, :, :, start:start + 1]],
axis=3), 'g'
], [tf.float32])
else:
[old_label, input_imgs
] = tf.py_func(self.get_masked_image,
[input_imgs, inputs[:, :, :, start:start + 1]],
[tf.float32, tf.float32])
summ = tf.summary.image("old_labels", inputs[:, :, :,
start:start + 1])
summaries.append(summ)
start += 1
# Add clicks to the image so that they can be viewed in tensorboard
if inputs.get_shape()[-1] > 6:
[input_imgs] = tf.py_func(self.add_clicks, [
tf.concat([input_imgs, inputs[:, :, :, start:start + 1]],
axis=3), 'r'
], [tf.float32])
[input_imgs] = tf.py_func(self.add_clicks, [
tf.concat([input_imgs, inputs[:, :, :, start + 1:start + 2]],
axis=3), 'r'
], [tf.float32])
[input_imgs] = tf.py_func(self.add_clicks, [
tf.concat([input_imgs, inputs[:, :, :, start + 2:start + 3]],
axis=3), 'g'
], [tf.float32])
[input_imgs] = tf.py_func(self.add_clicks, [
tf.concat([input_imgs, inputs[:, :, :, start + 3:start + 4]],
axis=3), 'g'
], [tf.float32])
elif inputs.get_shape()[-1] > 4:
[input_imgs] = tf.py_func(self.add_clicks, [
tf.concat([input_imgs, inputs[:, :, :, start:start + 1]],
axis=3), 'r'
], [tf.float32])
[input_imgs] = tf.py_func(self.add_clicks, [
tf.concat([input_imgs, inputs[:, :, :, start + 1:start + 2]],
axis=3), 'g'
], [tf.float32])
# bounding boxes
if Constants.BBOXES in tensors_dict:
bboxes = tensors_dict[Constants.BBOXES]
# permute y1, y2, x1, x2 -> y1, x1, y2, x1
bboxes = tf.stack([
bboxes[..., 0], bboxes[..., 2], bboxes[..., 1], bboxes[..., 3]
],
axis=-1)
# normalize bboxes to [0..1]
height = tf.shape(input_imgs)[1]
width = tf.shape(input_imgs)[2]
bboxes = tf.cast(bboxes, tf.float32) / tf.cast(
tf.stack([height, width, height, width], axis=0), tf.float32)
imgs_with_bboxes = tf.image.draw_bounding_boxes(input_imgs, bboxes)
summ = tf.summary.image("inputs", imgs_with_bboxes)
summaries.append(summ)
else:
summ = tf.summary.image("inputs", input_imgs)
summaries.append(summ)
if not isinstance(tensors_dict["labels"], tuple):
# will only work well for binary segmentation
summ = tf.summary.image("labels", tensors_dict["labels"] * 255)
summaries.append(summ)
# Append the distance transforms, if they are available.
if inputs.get_shape()[-1] > 6:
# Get negative distance transform from the extra input channels.
summ = tf.summary.image(Constants.DT_NEG, inputs[:, :, :,
start:start + 1])
summaries.append(summ)
summ = tf.summary.image(Constants.DT_NEG,
inputs[:, :, :, start + 1:start + 2])
summaries.append(summ)
summ = tf.summary.image(Constants.DT_POS,
inputs[:, :, :, start + 2:start + 3])
summaries.append(summ)
summ = tf.summary.image(Constants.DT_POS,
inputs[:, :, :, start + 3:start + 4])
summaries.append(summ)
elif inputs.get_shape()[-1] > 4:
# Get negative distance transform from the extra input channels.
summ = tf.summary.image(Constants.DT_NEG, inputs[:, :, :,
start:start + 1])
summaries.append(summ)
summ = tf.summary.image(Constants.DT_POS,
inputs[:, :, :, start + 1:start + 2])
summaries.append(summ)
return summaries