def test_compute_ignore_class(self):
class_config = ClassConfig(names=['one', 'two'])
class_config.update()
class_config.ensure_null_class()
null_class_id = class_config.get_null_class_id()
gt_array = np.zeros((4, 4, 1), dtype=np.uint8)
gt_array[0, 0, 0] = 2
gt_raster = MockRasterSource([0], 1)
gt_raster.set_raster(gt_array)
gt_label_source = SemanticSegmentationLabelSource(
gt_raster, null_class_id)
pred_array = np.zeros((4, 4, 1), dtype=np.uint8)
pred_array[0, 0, 0] = 1
pred_raster = MockRasterSource([0], 1)
pred_raster.set_raster(pred_array)
pred_label_source = SemanticSegmentationLabelSource(
pred_raster, null_class_id)
eval = SemanticSegmentationEvaluation(class_config)
eval.compute(gt_label_source.get_labels(),
pred_label_source.get_labels())
self.assertAlmostEqual(1.0, eval.class_to_eval_item[0].f1)
self.assertAlmostEqual(1.0, eval.avg_item.f1)
def test_compute(self):
class_config = ClassConfig(names=['one', 'two'])
class_config.update()
class_config.ensure_null_class()
null_class_id = class_config.get_null_class_id()
gt_array = np.zeros((4, 4, 1), dtype=np.uint8)
gt_array[2, 2, 0] = 1
gt_array[0, 0, 0] = 2
gt_raster = MockRasterSource([0], 1)
gt_raster.set_raster(gt_array)
gt_label_source = SemanticSegmentationLabelSource(
gt_raster, null_class_id)
p_array = np.zeros((4, 4, 1), dtype=np.uint8)
p_array[1, 1, 0] = 1
p_raster = MockRasterSource([0], 1)
p_raster.set_raster(p_array)
p_label_source = SemanticSegmentationLabelSource(
p_raster, null_class_id)
eval = SemanticSegmentationEvaluation(class_config)
eval.compute(gt_label_source.get_labels(), p_label_source.get_labels())
tp0 = 16 - 3 # 4*4 - 3 true positives for class 0
fp0 = 1 # 1 false positive (2,2) and one don't care at (0,0)
fn0 = 1 # one false negative (1,1)
precision0 = float(tp0) / (tp0 + fp0)
recall0 = float(tp0) / (tp0 + fn0)
f10 = 2 * float(precision0 * recall0) / (precision0 + recall0)
tp1 = 0 # 0 true positives for class 1
fn1 = 1 # one false negative (2,2)
precision1 = 0 # float(tp1) / (tp1 + fp1) where fp1 == 1
recall1 = float(tp1) / (tp1 + fn1)
f11 = None
self.assertAlmostEqual(precision0,
eval.class_to_eval_item[0].precision)
self.assertAlmostEqual(recall0, eval.class_to_eval_item[0].recall)
self.assertAlmostEqual(f10, eval.class_to_eval_item[0].f1)
self.assertEqual(precision1, eval.class_to_eval_item[1].precision)
self.assertAlmostEqual(recall1, eval.class_to_eval_item[1].recall)
self.assertAlmostEqual(f11, eval.class_to_eval_item[1].f1)
avg_conf_mat = np.array([[0, 0, 0], [13., 1, 0], [1, 0, 0]])
avg_recall = (14 / 15) * recall0 + (1 / 15) * recall1
self.assertTrue(np.array_equal(avg_conf_mat, eval.avg_item.conf_mat))
self.assertEqual(avg_recall, eval.avg_item.recall)
def test_get_labels_rgb(self):
data = np.zeros((10, 10, 3), dtype=np.uint8)
data[7:, 7:, :] = [1, 1, 1]
null_class_id = 2
raster_source = MockRasterSource([0, 1, 2], 3)
raster_source.set_raster(data)
rgb_class_config = ClassConfig(names=['a'], colors=['#010101'])
rgb_class_config.ensure_null_class()
label_source = SemanticSegmentationLabelSource(
raster_source, null_class_id, rgb_class_config=rgb_class_config)
with label_source.activate():
window = Box.make_square(7, 7, 3)
labels = label_source.get_labels(window=window)
label_arr = labels.get_label_arr(window)
expected_label_arr = np.zeros((3, 3))
np.testing.assert_array_equal(label_arr, expected_label_arr)
def test_vector_compute(self):
class_config = ClassConfig(names=['one', 'two'])
class_config.update()
class_config.ensure_null_class()
gt_uri = data_file_path('2-gt-polygons.geojson')
pred_uri = data_file_path('2-pred-polygons.geojson')
eval = SemanticSegmentationEvaluation(class_config)
eval.compute_vector(gt_uri, pred_uri, 'polygons', 0)
# NOTE: The two geojson files referenced above contain three
# unique geometries total, each file contains two geometries,
# and there is one geometry shared between the two.
tp = 1.0
fp = 1.0
fn = 1.0
precision = float(tp) / (tp + fp)
recall = float(tp) / (tp + fn)
self.assertAlmostEqual(precision, eval.class_to_eval_item[0].precision)
self.assertAlmostEqual(recall, eval.class_to_eval_item[0].recall)
class TestSemanticSegmentationEvaluator(unittest.TestCase):
def setUp(self):
self.tmp_dir = rv_config.get_tmp_dir()
self.class_config = ClassConfig(names=['one', 'two'])
self.class_config.update()
self.class_config.ensure_null_class()
self.null_class_id = self.class_config.get_null_class_id()
def tearDown(self):
self.tmp_dir.cleanup()
def get_scene(self, class_id):
# Make scene where ground truth is all set to class_id
# and predictions are set to half 0's and half 1's
scene_id = str(class_id)
rs = MockRasterSource(channel_order=[0, 1, 2], num_channels=3)
rs.set_raster(np.zeros((10, 10, 3)))
gt_rs = MockRasterSource(channel_order=[0], num_channels=1)
gt_arr = np.full((10, 10, 1), class_id)
gt_rs.set_raster(gt_arr)
gt_ls = SemanticSegmentationLabelSource(gt_rs, self.null_class_id)
pred_rs = MockRasterSource(channel_order=[0], num_channels=1)
pred_arr = np.zeros((10, 10, 1))
pred_arr[5:10, :, :] = 1
pred_rs.set_raster(pred_arr)
pred_ls = SemanticSegmentationLabelSource(pred_rs, self.null_class_id)
return Scene(scene_id, rs, gt_ls, pred_ls)
def test_evaluator(self):
output_uri = join(self.tmp_dir.name, 'out.json')
scenes = [self.get_scene(0), self.get_scene(1)]
evaluator = SemanticSegmentationEvaluator(self.class_config,
output_uri, None)
evaluator.process(scenes, self.tmp_dir.name)
eval_json = file_to_json(output_uri)
exp_eval_json = file_to_json(data_file_path('expected-eval.json'))
self.assertDictEqual(eval_json, exp_eval_json)
def get_vector_scene(self, class_id, use_aoi=False):
gt_uri = data_file_path('{}-gt-polygons.geojson'.format(class_id))
pred_uri = data_file_path('{}-pred-polygons.geojson'.format(class_id))
scene_id = str(class_id)
rs = MockRasterSource(channel_order=[0, 1, 3], num_channels=3)
rs.set_raster(np.zeros((10, 10, 3)))
crs_transformer = IdentityCRSTransformer()
extent = Box.make_square(0, 0, 360)
config = RasterizedSourceConfig(
vector_source=GeoJSONVectorSourceConfig(uri=gt_uri,
default_class_id=0),
rasterizer_config=RasterizerConfig(background_class_id=1))
gt_rs = config.build(self.class_config, crs_transformer, extent)
gt_ls = SemanticSegmentationLabelSource(gt_rs, self.null_class_id)
config = RasterizedSourceConfig(
vector_source=GeoJSONVectorSourceConfig(uri=pred_uri,
default_class_id=0),
rasterizer_config=RasterizerConfig(background_class_id=1))
pred_rs = config.build(self.class_config, crs_transformer, extent)
pred_ls = SemanticSegmentationLabelSource(pred_rs, self.null_class_id)
pred_ls.vector_output = [
PolygonVectorOutputConfig(uri=pred_uri,
denoise=0,
class_id=class_id)
]
if use_aoi:
aoi_uri = data_file_path('{}-aoi.geojson'.format(class_id))
aoi_geojson = file_to_json(aoi_uri)
aoi_polygons = [shape(aoi_geojson['features'][0]['geometry'])]
return Scene(scene_id, rs, gt_ls, pred_ls, aoi_polygons)
return Scene(scene_id, rs, gt_ls, pred_ls)
def test_vector_evaluator(self):
output_uri = join(self.tmp_dir.name, 'raster-out.json')
vector_output_uri = join(self.tmp_dir.name, 'vector-out.json')
scenes = [self.get_vector_scene(0), self.get_vector_scene(1)]
evaluator = SemanticSegmentationEvaluator(self.class_config,
output_uri,
vector_output_uri)
evaluator.process(scenes, self.tmp_dir.name)
vector_eval_json = file_to_json(vector_output_uri)
exp_vector_eval_json = file_to_json(
data_file_path('expected-vector-eval.json'))
# NOTE: The precision and recall values found in the file
# `expected-vector-eval.json` are equal to fractions of the
# form (n-1)/n for n <= 7 which can be seen to be (and have
# been manually verified to be) correct.
self.assertDictEqual(vector_eval_json, exp_vector_eval_json)
def test_vector_evaluator_with_aoi(self):
output_uri = join(self.tmp_dir.name, 'raster-out.json')
vector_output_uri = join(self.tmp_dir.name, 'vector-out.json')
scenes = [self.get_vector_scene(0, use_aoi=True)]
evaluator = SemanticSegmentationEvaluator(self.class_config,
output_uri,
vector_output_uri)
evaluator.process(scenes, self.tmp_dir.name)
vector_eval_json = file_to_json(vector_output_uri)
exp_vector_eval_json = file_to_json(
data_file_path('expected-vector-eval-with-aoi.json'))
# NOTE: The precision and recall values found in the file
# `expected-vector-eval.json` are equal to fractions of the
# form (n-1)/n for n <= 7 which can be seen to be (and have
# been manually verified to be) correct.
self.assertDictEqual(vector_eval_json, exp_vector_eval_json)
def get_config(runner, root_uri, data_uri=None, full_train=False,
nochip=False):
def get_path(part):
if full_train:
return join(data_uri, part)
else:
return join(dirname(__file__), part)
class_config = ClassConfig(names=['red', 'green'], colors=['red', 'green'])
class_config.ensure_null_class()
def make_scene(id, img_path, label_path):
raster_source = RasterioSourceConfig(
channel_order=[0, 1, 2], uris=[img_path])
label_source = SemanticSegmentationLabelSourceConfig(
rgb_class_config=class_config,
raster_source=RasterioSourceConfig(uris=[label_path]))
label_store = SemanticSegmentationLabelStoreConfig(
rgb=True,
vector_output=[
PolygonVectorOutputConfig(class_id=0),
BuildingVectorOutputConfig(class_id=1)
])
return SceneConfig(
id=id,
raster_source=raster_source,
label_source=label_source,
label_store=label_store)
chip_sz = 300
img_sz = chip_sz
scenes = [
make_scene('test-scene', get_path('scene/image.tif'),
get_path('scene/labels.tif')),
make_scene('test-scene2', get_path('scene/image2.tif'),
get_path('scene/labels2.tif'))
]
scene_dataset = DatasetConfig(
class_config=class_config,
train_scenes=scenes,
validation_scenes=scenes)
chip_options = SemanticSegmentationChipOptions(
window_method=SemanticSegmentationWindowMethod.sliding, stride=chip_sz)
if nochip:
window_opts = GeoDataWindowConfig(
method=GeoDataWindowMethod.sliding,
stride=chip_options.stride,
size=chip_sz)
data = SemanticSegmentationGeoDataConfig(
scene_dataset=scene_dataset,
window_opts=window_opts,
img_sz=img_sz,
augmentors=[])
else:
data = SemanticSegmentationImageDataConfig(
img_sz=img_sz, augmentors=[])
if full_train:
model = SemanticSegmentationModelConfig(backbone=Backbone.resnet50)
solver = SolverConfig(
lr=1e-4,
num_epochs=300,
batch_sz=8,
one_cycle=True,
sync_interval=300)
else:
pretrained_uri = (
'https://github.com/azavea/raster-vision-data/releases/download/v0.12/'
'semantic-segmentation.pth')
model = SemanticSegmentationModelConfig(
backbone=Backbone.resnet50, init_weights=pretrained_uri)
solver = SolverConfig(
lr=1e-9,
num_epochs=1,
batch_sz=2,
one_cycle=True,
sync_interval=200)
backend = PyTorchSemanticSegmentationConfig(
data=data,
model=model,
solver=solver,
log_tensorboard=False,
run_tensorboard=False)
return SemanticSegmentationConfig(
root_uri=root_uri,
dataset=scene_dataset,
backend=backend,
train_chip_sz=chip_sz,
predict_chip_sz=chip_sz,
chip_options=chip_options)