def _make_data_gen(hypes, phase, data_dir):
"""Return a data generator that outputs image samples.
@ Returns
image: integer array of shape [width, height, 3].
Representing RGB value of each pixel.
gt_image: boolean array of shape [width, height, num_classes].
Set `gt_image[i,j,k] == 1` if and only if pixel i,j
is assigned class k. `gt_image[i,j,k] == 0` otherwise.
[Alternativly make gt_image[i,j,*] a valid propability
distribution.]
"""
if phase == 'train':
data_file = hypes['data']["train_file"]
elif phase == 'val':
data_file = hypes['data']["val_file"]
else:
assert False, "Unknown Phase %s" % phase
data_file = os.path.join(data_dir, data_file)
data = _load_gt_file(hypes, data_file)
for image, gt_image in data:
if phase == 'val':
yield image, utils.load_segmentation_mask(hypes,
gt_image)[:, :, :-1]
elif phase == 'train':
image, gt_image = jitter_input(hypes, image, gt_image)
yield image, utils.load_segmentation_mask(hypes,
gt_image)[:, :, :-1]
image, gt_image = jitter_input(hypes, np.fliplr(image),
np.fliplr(gt_image))
yield image, utils.load_segmentation_mask(hypes,
gt_image)[:, :, :-1]
def get_class_distribution(hypes, labeled_dataset):
"""
Get the distribution of classes in a labeled dataset.
Parameters
----------
hypes : dict
The hyperparameters have to specify 'classes'.
labeled_dataset : list of dicts
Each dict has to have the keys 'raw' and 'mask' which have the absolute
path to image files.
Returns
-------
dict
Mapping class indices according to hypes['classes'] to pixel counts.
"""
classes = {}
for item in labeled_dataset:
im = utils.load_segmentation_mask(hypes, item['mask'])
for y in range(im.shape[0]):
for x in range(im.shape[1]):
cl = im[y][x]
if cl in classes:
classes[cl] += 1
else:
classes[cl] = 1
return classes
def main(hypes_file, output_dir, override):
"""
Orchestrate.
Parameters
----------
hypes_file : str
Path to a JSON file
output_dir : str
Path where the output should be stored
override : bool
If True, then override the model if it exists.
"""
# Load hyperparameters
with open(hypes_file, 'r') as f:
hypes = json.load(f)
# Set serialization path
base = os.path.dirname(hypes_file)
model_file_path = os.path.join(base, '%s.json' % hypes['model']['name'])
model_file_path = os.path.abspath(model_file_path)
if not os.path.isfile(model_file_path) or override:
if not os.path.isfile(model_file_path):
logging.info("Did not find '%s'. Start training...",
model_file_path)
else:
logging.info("Override '%s'. Start training...",
model_file_path)
# Get training data
x_files, y_files = get_file_list(hypes, 'train')
# "Train" "classifier" (it just counts the classes)
model = {}
for i in range(len(hypes['classes'])):
model[i] = 0
for y_file in y_files:
logging.info("Read '%s'...", y_file)
mask = load_segmentation_mask(hypes, y_file)
for row in mask:
for pixel in row:
model[pixel] += 1
# save model as json file
with open(model_file_path, 'w') as f:
json.dump(model, f)
else:
# load model from json file
with open(model_file_path) as f:
model = json.load(f)
# Evaluate
data = get_file_list(hypes, 'test')
analyze.evaluate(hypes,
data,
output_dir,
model,
elements=[0, 1],
get_segmentation=get_segmentation)
def eval_image(hypes, gt_image, cnn_image):
"""."""
thresh = np.array(range(0, 256))/255.0
valid_gt = utils.load_segmentation_mask(hypes, gt_image)[:, :, :-1]
return seg.evalExp(hypes, valid_gt, cnn_image)
def main(hypes_file, output_dir, override):
"""
Orchestrate.
Parameters
----------
hypes_file : str
Path to a JSON file
output_dir : str
Path where the output should be stored
override : bool
If True, then override the model if it exists.
"""
# Load hyperparameters
with open(hypes_file, 'r') as f:
hypes = json.load(f)
# Set serialization path
base = os.path.dirname(hypes_file)
model_file_path = os.path.join(base, '%s.pickle' % hypes['model']['name'])
model_file_path = os.path.abspath(model_file_path)
if not os.path.isfile(model_file_path) or override:
if not os.path.isfile(model_file_path):
logging.info("Did not find '%s'. Start training...",
model_file_path)
else:
logging.info("Override '%s'. Start training...", model_file_path)
# Get training data
x_files, y_files = get_file_list(hypes, 'train')
# "Train" "classifier" (it just counts the classes)
model = {'positions': None, 'files': 0}
for y_file in y_files:
logging.info("Read '%s'...", y_file)
mask = load_segmentation_mask(hypes, y_file)
if model['positions'] is None:
model['positions'] = mask
else:
model['positions'] += mask
model['files'] += 1
# save model as pickle file
scipy.misc.imsave("instruments.png", model['positions'])
with open(model_file_path, 'wb') as handle:
pickle.dump(model, handle, protocol=pickle.HIGHEST_PROTOCOL)
else:
# load model from pickle file
with open(model_file_path, 'rb') as handle:
model = pickle.load(handle)
# Evaluate
data = get_file_list(hypes, 'test')
analyze.evaluate(hypes,
data,
output_dir,
model,
elements=[0, 1],
get_segmentation=get_segmentation)
def get_traindata_single_file(hypes, x, y):
"""Get trainingdata for a single file x with segmentation file y."""
xs, ys = [], []
logging.info("Read '%s' for data...", x)
label = load_segmentation_mask(hypes, y)
im = Image.open(x, 'RGB')
width, height, _ = im.size
for x in range(width):
for y in range(height):
image_val = get_features(hypes, im, 'data')
label_val = label[y][x]
xs.append(image_val)
ys.append(label_val)
return numpy.array(xs), numpy.array(ys, dtype=int)
def get_traindata_single_file(hypes, x, y):
"""Get trainingdata for a single file x with segmentation file y."""
xs, ys = [], []
logging.info("Read '%s' for data...", x)
label = load_segmentation_mask(hypes, y)
im = Image.open(x, 'RGB')
width, height, _ = im.size
for x in range(width):
for y in range(height):
image_val = get_features(hypes, im, 'data')
label_val = label[y][x]
xs.append(image_val)
ys.append(label_val)
return numpy.array(xs), numpy.array(ys, dtype=int)
def test_load_segmentation_mask():
"""Test load_segmentation_mask."""
from tensorvision.utils import load_segmentation_mask
import json
import scipy.misc
import numpy
import os
hypes_path = os.path.abspath('tensorvision/tests/croco.json')
seg_image = os.path.abspath('tensorvision/tests/croco-mask.png')
with open(hypes_path) as f:
hypes = json.load(f)
gt = load_segmentation_mask(hypes, seg_image)
# scipy.misc.imsave("tensorvision/tests/test-generated-mask.png", gt)
seg_image = 'tensorvision/tests/Crocodylus-johnsoni-3-mask.png'
true_gt = (scipy.misc.imread(seg_image) / 255.).astype(int)
numpy.testing.assert_array_equal(true_gt, gt)
assert gt.shape == (480, 640)
def test_load_segmentation_mask():
"""Test load_segmentation_mask."""
from tensorvision.utils import load_segmentation_mask
import json
import scipy.misc
import numpy
import os
hypes_path = os.path.abspath('tensorvision/tests/croco.json')
seg_image = os.path.abspath('tensorvision/tests/croco-mask.png')
with open(hypes_path) as f:
hypes = json.load(f)
gt = load_segmentation_mask(hypes, seg_image)
# scipy.misc.imsave("tensorvision/tests/test-generated-mask.png", gt)
seg_image = 'tensorvision/tests/Crocodylus-johnsoni-3-mask.png'
true_gt = (scipy.misc.imread(seg_image) / 255.).astype(int)
numpy.testing.assert_array_equal(true_gt, gt)
assert gt.shape == (480, 640)
def generate_batch(hypes, phase):
"""
Generate patches.
Parameters
----------
hypes : dict
phase : 'train' or 'test'
"""
x_files, y_files = get_file_list(hypes, phase)
x_files, y_files = sklearn.utils.shuffle(x_files,
y_files,
random_state=0)
batch_x, batch_y = [], []
while True:
for x, y in zip(x_files, y_files):
logging.info("Read '%s' for data...", x)
image = get_image(x, 'RGB')
label = load_segmentation_mask(hypes, y)
im = Image.open(x, 'r')
width, height = im.size
image_vals = get_features(hypes, image, 'data')
label_vals = get_features(hypes, label, 'label')
# print("image_vals = %s" % str(list(image_vals)))
for patch, label_ in zip(image_vals, label_vals):
patch = img_to_array(patch)
label_ = img_to_array(label_)
_, w, h = label_.shape
label_ = label_.reshape((w, h))
if phase == 'val' and 1.0 not in label_:
print("continue")
continue
# scipy.misc.imshow(patch)
# scipy.misc.imshow(label_)
batch_x.append(patch)
batch_y.append(label_) # .flatten()
if len(batch_x) == hypes['solver']['batch_size']:
yield (np.array(batch_x), np.array(batch_y))
batch_x, batch_y = [], []
def inputs(hypes, _, phase, data_dir):
"""
Get data.
Parameters
----------
hypes : dict
_ : ignore this
phase : {'train', 'val'}
data_dir : str
Returns
-------
tuple
(xs, ys), where xs and ys are lists of the same length.
xs are paths to the input images and ys are paths to the expected
output
"""
x_files, y_files = get_file_list(hypes, 'train')
x_files, y_files = sklearn.utils.shuffle(x_files,
y_files,
random_state=0)
xs, ys = [], []
for x, y in zip(x_files, y_files):
logging.info("Read '%s' for data...", x)
image = get_image(x, 'RGB')
label = load_segmentation_mask(hypes, y)
im = Image.open(x, 'r')
width, height = im.size
for x in range(width):
for y in range(height):
image_val = get_features(x, y, image, hypes['model_nr'])
label_val = label[y][x]
xs.append(image_val)
ys.append(label_val)
return xs, np.array(ys, dtype=int)
def generate_batch(hypes, phase):
"""
Generate patches.
Parameters
----------
hypes : dict
phase : 'train' or 'test'
"""
x_files, y_files = get_file_list(hypes, phase)
x_files, y_files = sklearn.utils.shuffle(x_files, y_files, random_state=0)
batch_x, batch_y = [], []
while True:
for x, y in zip(x_files, y_files):
logging.info("Read '%s' for data...", x)
image = get_image(x, 'RGB')
label = load_segmentation_mask(hypes, y)
im = Image.open(x, 'r')
width, height = im.size
image_vals = get_features(hypes, image, 'data')
label_vals = get_features(hypes, label, 'label')
# print("image_vals = %s" % str(list(image_vals)))
for patch, label_ in zip(image_vals, label_vals):
patch = img_to_array(patch)
label_ = img_to_array(label_)
_, w, h = label_.shape
label_ = label_.reshape((w, h))
if phase == 'val' and 1.0 not in label_:
print("continue")
continue
# scipy.misc.imshow(patch)
# scipy.misc.imshow(label_)
batch_x.append(patch)
batch_y.append(label_) # .flatten()
if len(batch_x) == hypes['solver']['batch_size']:
yield (np.array(batch_x), np.array(batch_y))
batch_x, batch_y = [], []
def main(hypes_file, output_dir, override):
"""
Orchestrate.
Parameters
----------
hypes_file : str
Path to a JSON file
output_dir : str
Path where the output should be stored
override : bool
If True, then override the model if it exists.
"""
# Load hyperparameters
with open(hypes_file, 'r') as f:
hypes = json.load(f)
# Set serialization path
base = os.path.dirname(hypes_file)
model_file_path = os.path.join(base, '%s.pickle' % hypes['model']['name'])
model_file_path = os.path.abspath(model_file_path)
if not os.path.isfile(model_file_path) or override:
if not os.path.isfile(model_file_path):
logging.info("Did not find '%s'. Start training...",
model_file_path)
else:
logging.info("Override '%s'. Start training...",
model_file_path)
# Get training data
x_files, y_files = get_file_list(hypes, 'train')
# "Train" "classifier" (it just counts the classes)
model = {'positions': None, 'files': 0}
for y_file in y_files:
logging.info("Read '%s'...", y_file)
mask = load_segmentation_mask(hypes, y_file)
if model['positions'] is None:
model['positions'] = mask
else:
model['positions'] += mask
model['files'] += 1
# save model as pickle file
scipy.misc.imsave("instruments.png", model['positions'])
with open(model_file_path, 'wb') as handle:
pickle.dump(model, handle, protocol=pickle.HIGHEST_PROTOCOL)
else:
# load model from pickle file
with open(model_file_path, 'rb') as handle:
model = pickle.load(handle)
# Evaluate
data = get_file_list(hypes, 'test')
analyze.evaluate(hypes,
data,
output_dir,
model,
elements=[0, 1],
get_segmentation=get_segmentation)
def evaluate(hypes,
data,
out_dir,
model,
elements,
get_segmentation,
verbose=True):
"""
Analyze how well a model does on given data.
Parameters
----------
hypes : dict
Hyperparameters (model specific information)
data : tuple
(x_files, y_files) where x_files and y_files are lists of strings
out_dir : str
Path to the directory where the output gets stored
model : object
elements : iterable
A list / set or another iterable which contains the possible
segmentation classes (commonly 0 and 1)
get_segmentation : function
Takes a string and a model the model as input and returns a numpy
array. The string is the path to an image file.
load_label_seg : function
Takes a path as a string and returns a segmentation mask as a numpy
array.
verbose : bool
Print messages when in verbose mode.
"""
# Initialize confusion matrix
cm = {}
for i in elements:
cm[i] = {}
for j in elements:
cm[i][j] = 0
# Initialize timings for segmentation
times = []
i = 1
for xfile, yfile in zip(*data):
if verbose:
logging.info("Segmentation of '%s'...", xfile)
logging.info(" label='%s'...", yfile)
t0 = time.time()
segmentation = get_segmentation(hypes, xfile, model)
t1 = time.time()
times.append(t1 - t0)
correct_seg = load_segmentation_mask(hypes, yfile)
cm_tmp = get_confusion_matrix(correct_seg, segmentation, elements)
color_changes = get_color_changes_dict(hypes)
gen_img_and_overlay(xfile, segmentation, out_dir, color_changes, i)
cm = merge_cms(cm, cm_tmp)
if verbose:
show_metrics(cm, indentation=4)
print(" time: %0.4fs" % (t1 - t0))
i += 1
if verbose:
show_metrics(cm)
print("Average time: %0.4f" % (sum(times) / len(times)))
return {'cm': cm, 'times': times}
