def test_config(self):
m_obj = metrics.MeanIoU(num_classes=2, name='mean_iou')
assert m_obj.name == 'mean_iou'
assert m_obj.num_classes == 2
m_obj2 = metrics.MeanIoU.from_config(m_obj.get_config())
assert m_obj2.name == 'mean_iou'
assert m_obj2.num_classes == 2
def test_mean_iou():
import tensorflow as tf
if not tf.__version__.startswith('2.'):
return
model = Sequential([Dense(1, input_shape=(3, ))])
model.compile('rmsprop', 'mse', metrics=[metrics.MeanIoU(2)])
x = np.random.random((10, 3))
y = np.random.random((10, ))
model.fit(x, y)
model.evaluate(x, y)
def test_zero_and_non_zero_entries(self):
y_pred = K.constant([1], dtype='float32')
y_true = K.constant([1])
m_obj = metrics.MeanIoU(num_classes=2)
result = m_obj(y_true, y_pred)
# cm = [[0, 0],
# [0, 1]]
# sum_row = [0, 1], sum_col = [0, 1], true_positives = [0, 1]
# iou = true_positives / (sum_row + sum_col - true_positives))
expected_result = (0. + 1. / (1 + 1 - 1)) / 1
assert np.allclose(K.eval(result), expected_result, atol=1e-3)
def test_unweighted(self):
y_pred = K.constant([0, 1, 0, 1], shape=(1, 4))
y_true = K.constant([0, 0, 1, 1], shape=(1, 4))
m_obj = metrics.MeanIoU(num_classes=2)
result = m_obj(y_true, y_pred)
# cm = [[1, 1],
# [1, 1]]
# sum_row = [2, 2], sum_col = [2, 2], true_positives = [1, 1]
# iou = true_positives / (sum_row + sum_col - true_positives))
expected_result = (1. / (2 + 2 - 1) + 1. / (2 + 2 - 1)) / 2
assert np.allclose(K.eval(result), expected_result, atol=1e-3)
def test_multi_dim_input(self):
y_pred = K.constant([[0, 1], [0, 1]], dtype='float32')
y_true = K.constant([[0, 0], [1, 1]])
sample_weight = K.constant([[0.2, 0.3], [0.4, 0.1]])
m_obj = metrics.MeanIoU(num_classes=2)
result = m_obj(y_true, y_pred, sample_weight=sample_weight)
# cm = [[0.2, 0.3],
# [0.4, 0.1]]
# sum_row = [0.6, 0.4], sum_col = [0.5, 0.5], true_positives = [0.2, 0.1]
# iou = true_positives / (sum_row + sum_col - true_positives))
expected_result = (0.2 / (0.6 + 0.5 - 0.2) + 0.1 /
(0.4 + 0.5 - 0.1)) / 2
assert np.allclose(K.eval(result), expected_result, atol=1e-3)
def test_zero_valid_entries(self):
m_obj = metrics.MeanIoU(num_classes=2)
assert np.allclose(K.eval(m_obj.result()), 0, atol=1e-3)
model.train(train_images=train_images,
train_annotations=train_annotations,
val_images=init + "sunrgb/val/rgb/",
val_annotations=init + "sunrgb/val/seg/",
checkpoints_path=init,
batch_size=2,
steps_per_epoch=512,
val_batch_size=2,
n_classes=38,
validate=True,
verify_dataset=False,
optimizer_name=opt,
loss_type=mild_categorical_crossentropy,
metrics_used=[
masked_categorical_accuracy,
metrics.MeanIoU(name='mean_iou', num_classes=n_classes)
],
do_augment=False,
gen_use_multiprocessing=False,
epochs=5)
# print(model.summary())
# print(model.evaluate_segmentation(
# inp_images_dir = "/Users/salvatorecapuozzo/Desktop/sunrgb/test/rgb/",
# annotations_dir = "/Users/salvatorecapuozzo/Desktop/sunrgb/test/seg/"
# ))
elif mode == 1:
import sys
if writeOnNotes:
f = open("C:/Users/UC/Desktop/test_models.out", 'w')
sys.stdout = f