def test_should_build_train_graph_with_sample_class_weights(
self,
parse_color_map_from_file_mock,
parse_json_file_mock,
read_examples_mock):
parse_color_map_from_file_mock.return_value = SOME_COLOR_MAP
parse_json_file_mock.return_value = SOME_CLASS_WEIGHTS
read_examples_mock.return_value = EXAMPLE_PROPS_1
args = create_args(
DEFAULT_ARGS,
base_loss=BaseLoss.SAMPLE_WEIGHTED_CROSS_ENTROPY,
color_map=COLOR_MAP_FILENAME,
channels=SOME_LABELS,
use_separate_channels=True,
use_unknown_class=True
)
model = Model(args)
tensors = model.build_train_graph(DATA_PATH, BATCH_SIZE)
n_output_channels = len(SOME_LABELS) + 1
assert (
tensors.separate_channel_annotation_tensor.shape.as_list() ==
[BATCH_SIZE, model.image_height, model.image_width, n_output_channels]
)
assert tensors.pos_weight.shape.as_list(
) == [BATCH_SIZE, 1, 1, n_output_channels]
def test_should_build_train_graph_with_defaults(
self, parse_color_map_from_file_mock, read_examples_mock):
parse_color_map_from_file_mock.return_value = SOME_COLOR_MAP
read_examples_mock.return_value = EXAMPLE_PROPS_1
args = create_args(DEFAULT_ARGS)
model = Model(args)
model.build_train_graph(DATA_PATH, BATCH_SIZE)
def test_should_build_predict_graph_with_defaults(
self, parse_color_map_from_file_mock, read_examples_mock):
parse_color_map_from_file_mock.return_value = SOME_COLOR_MAP
read_examples_mock.return_value = EXAMPLE_PROPS_1
args = create_args(DEFAULT_ARGS)
model = Model(args)
tensors = model.build_predict_graph()
n_output_channels = 3
assert (tensors.pred.shape.as_list() == [
None, model.image_height, model.image_width, n_output_channels
])
def test_should_only_include_labels_with_non_zero_class_labels_by_default(
self,
parse_color_map_from_file_mock,
parse_json_file_mock):
parse_color_map_from_file_mock.return_value = {
'a': some_color(1),
'b': some_color(2),
'c': some_color(3)
}
parse_json_file_mock.return_value = {
'a': 0.1,
'b': 0.0,
'c': 0.3
}
args = create_args(
DEFAULT_ARGS,
base_loss=BaseLoss.WEIGHTED_CROSS_ENTROPY,
color_map=COLOR_MAP_FILENAME,
class_weights=CLASS_WEIGHTS_FILENAME,
use_separate_channels=True,
use_unknown_class=True
)
model = Model(args)
assert model.dimension_labels == ['a', 'c']
assert model.dimension_colors == [some_color(1), some_color(3)]
assert model.pos_weight == [0.1, 0.3, DEFAULT_UNKNOWN_CLASS_WEIGHT]
def test_parse_separate_channels_with_color_map_and_class_weights(
self,
parse_color_map_from_file_mock,
parse_json_file_mock):
parse_color_map_from_file_mock.return_value = {
'a': some_color(1),
'b': some_color(2),
'c': some_color(3)
}
parse_json_file_mock.return_value = {
'a': 0.1,
'b': 0.2,
'c': 0.3
}
args = create_args(
DEFAULT_ARGS,
base_loss=BaseLoss.WEIGHTED_CROSS_ENTROPY,
color_map=COLOR_MAP_FILENAME,
class_weights=CLASS_WEIGHTS_FILENAME,
channels=['a', 'b'],
use_separate_channels=True,
use_unknown_class=True
)
model = Model(args)
assert model.dimension_colors == [some_color(1), some_color(2)]
assert model.dimension_labels == ['a', 'b']
assert (
model.dimension_colors_with_unknown == [
some_color(1), some_color(2), UNKNOWN_COLOR]
)
assert model.dimension_labels_with_unknown == ['a', 'b', UNKNOWN_LABEL]
assert model.pos_weight == [0.1, 0.2, DEFAULT_UNKNOWN_CLASS_WEIGHT]
def test_should_build_train_graph_with_class_weights(
self, parse_color_map_from_file_mock, parse_json_file_mock,
read_examples_mock):
parse_color_map_from_file_mock.return_value = SOME_COLOR_MAP
parse_json_file_mock.return_value = SOME_CLASS_WEIGHTS
read_examples_mock.return_value = EXAMPLE_PROPS_1
args = create_args(DEFAULT_ARGS,
base_loss=BaseLoss.WEIGHTED_CROSS_ENTROPY,
color_map=COLOR_MAP_FILENAME,
class_weights=CLASS_WEIGHTS_FILENAME,
channels=['a', 'b'],
use_separate_channels=True,
use_unknown_class=True)
model = Model(args)
tensors = model.build_train_graph(DATA_PATH, BATCH_SIZE)
assert tensors.pos_weight is not None
def test_should_not_load_class_weights_for_sample_weighted_cross_entropy(
self, parse_color_map_from_file_mock):
parse_color_map_from_file_mock.return_value = SOME_COLOR_MAP
args = create_args(DEFAULT_ARGS,
base_loss=BaseLoss.SAMPLE_WEIGHTED_CROSS_ENTROPY,
color_map=COLOR_MAP_FILENAME,
class_weights=CLASS_WEIGHTS_FILENAME,
use_separate_channels=True,
use_unknown_class=True)
model = Model(args)
assert model.pos_weight is None
def test_should_use_unknown_class_weight_from_configuration(
self, parse_color_map_from_file_mock, parse_json_file_mock):
parse_color_map_from_file_mock.return_value = SOME_COLOR_MAP
parse_json_file_mock.return_value = extend_dict(
SOME_CLASS_WEIGHTS, {'unknown': 0.99})
args = create_args(DEFAULT_ARGS,
base_loss=BaseLoss.WEIGHTED_CROSS_ENTROPY,
color_map=COLOR_MAP_FILENAME,
class_weights=CLASS_WEIGHTS_FILENAME,
use_separate_channels=True,
use_unknown_class=True)
model = Model(args)
assert model.pos_weight[-1] == 0.99
def test_parse_separate_channels_with_color_map_without_class_weights(
self, parse_color_map_from_file_mock):
parse_color_map_from_file_mock.return_value = {
'a': some_color(1),
'b': some_color(2),
'c': some_color(3)
}
args = create_args(DEFAULT_ARGS,
color_map=COLOR_MAP_FILENAME,
class_weights=None,
channels=['a', 'b'],
use_separate_channels=True,
use_unknown_class=True)
model = Model(args)
assert model.dimension_colors == [some_color(1), some_color(2)]
assert model.dimension_labels == ['a', 'b']
assert model.dimension_colors_with_unknown == [
some_color(1), some_color(2), UNKNOWN_COLOR
]
assert model.dimension_labels_with_unknown == ['a', 'b', UNKNOWN_LABEL]
assert model.pos_weight is None