def test_Mixup():
numbers1 = [1, 2, 3]
numbers2 = [4, 5, 6]
samples = list(zip(numbers1, numbers2))
build_batch = (nb.BuildBatch(3, prefetch=0).input(0, 'number',
float).output(
1, 'number', float))
# no mixup, return original batch
mixup = nb.Mixup(0.0)
batches = samples >> build_batch >> mixup >> Collect()
inputs, outputs = batches[0]
assert list(inputs[0]) == numbers1
assert list(outputs[0]) == numbers2
# mixup with alpaha=1.0
mixup = nb.Mixup(1.0)
batches = samples >> build_batch >> mixup >> Collect()
for input, output in batches:
input, output = input[0], output[0]
assert min(input) >= 1 and max(input) <= 3
assert min(output) >= 4 and max(output) <= 6
ri, ro = input[0] - samples[0][0], output[0] - samples[0][1]
assert approx(ri, 1e-3) == ro
def test_Network():
model = FakeModel()
weightspath = 'dummy_filepath'
network = FakeNetwork(model, weightspath)
assert network.weightspath == weightspath
batches = [((1, 2), (3, 4)), ((5, 6), (7, 8))]
train_err = batches >> network.train() >> Collect()
assert train_err == [(4, 8), (12, 16)]
val_err = batches >> network.validate() >> Collect()
assert val_err == [(5, 9), (13, 17)]
prediction = batches >> GetCols(0) >> network.predict() >> Collect()
assert prediction == [(1, 2), (5, 6)]
prediction = batches >> GetCols(0) >> network.predict(False) >> Collect()
assert prediction == [((1, 2),), ((5, 6),)]
batches = [((1, 2), (1, 2)), ((5, 6), (5, 6))]
acc = lambda X, y: np.sum(X == y)
assert batches >> network.evaluate([acc]) == 4
batches = [(((0, 1), (0, 2)), (0, 2)), (((5, 5), (6, 6)), (6, 6))]
assert batches >> network.evaluate([acc], predcol=1) == 4
assert batches >> network.evaluate([acc], predcol=0) == 1
def test_ImageWriter():
samples = [('nut_color', 1), ('nut_grayscale', 2)]
inpath = 'tests/data/img_formats/*.bmp'
img_samples = samples >> ReadImage(0, inpath) >> Collect()
imagepath = 'tests/data/test_*.bmp'
names = samples >> Get(0) >> Collect()
img_samples >> WriteImage(0, imagepath, names) >> Consume()
for sample, name in zip(img_samples, names):
filepath = 'tests/data/test_{}.bmp'.format(name)
arr = load_image(filepath)
assert np.array_equal(arr, sample[0])
os.remove(filepath)
pathfunc = lambda sample, name: 'tests/data/test_{}.jpg'.format(name)
img_samples >> WriteImage(0, pathfunc) >> Consume()
for i, sample in enumerate(img_samples):
filepath = 'tests/data/test_{}.jpg'.format(i)
os.path.exists(filepath)
os.remove(filepath)
with pytest.raises(ValueError) as ex:
img_samples >> WriteImage(0, ()) >> Consume()
assert str(ex.value).startswith('Expect path or function')
def test_Boost():
negatives = [(0, 0), (1, 0)]
positives = [(2, 1), (3, 1), (4, 1)]
samples = negatives + positives
build_batch = (BuildBatch(3,
prefetch=0).input(0, 'number', 'uint8').output(
1, 'one_hot', 'uint8', 2))
network = FakeNetwork(predict_all_positive)
boost = Boost(build_batch, network)
boosted = samples >> boost >> Collect()
assert boosted == negatives, 'Expect negatives boosted'
network = FakeNetwork(predict_all_negative)
boost = Boost(build_batch, network)
boosted = samples >> boost >> Collect()
assert boosted == positives, 'Expect positives boosted'
network = FakeNetwork(predict_all_perfect)
boost = Boost(build_batch, network)
boosted = samples >> boost >> Collect()
assert boosted == [], 'Expect no samples left for boosting'
network = FakeNetwork(predict_all_wrong)
boost = Boost(build_batch, network)
boosted = samples >> boost >> Collect()
assert boosted == samples, 'Expect all samples boosted'
def test_ImagePatchesByMask():
img = np.reshape(np.arange(25), (5, 5))
mask = np.eye(5, dtype='uint8') * 255
samples = [(img, mask)]
np.random.seed(0)
get_patches = ImagePatchesByMask(0, 1, (3, 3), 1, 1, retlabel=False)
patches = samples >> get_patches >> Collect()
assert len(patches) == 2
p, m = patches[0]
img_patch0 = np.array([[12, 13, 14], [17, 18, 19], [22, 23, 24]])
mask_patch0 = np.array([[255, 0, 0], [0, 255, 0], [0, 0, 255]])
nt.assert_allclose(p, img_patch0)
nt.assert_allclose(m, mask_patch0)
p, m = patches[1]
img_patch1 = np.array([[10, 11, 12], [15, 16, 17], [20, 21, 22]])
mask_patch1 = np.array([[0, 0, 255], [0, 0, 0], [0, 0, 0]])
nt.assert_allclose(p, img_patch1)
nt.assert_allclose(m, mask_patch1)
with pytest.raises(ValueError) as ex:
mask = np.eye(3, dtype='uint8') * 255
samples = [(img, mask)]
get_patches = ImagePatchesByMask(0, 1, (3, 3), 1, 1)
samples >> get_patches >> Collect()
assert str(ex.value).startswith('Image and mask size don' 't match!')
def test_ImagePatchesByAnnotation():
img = np.reshape(np.arange(25), (5, 5))
anno = ('point', (
(3, 2),
(2, 3),
))
samples = [(img, anno)]
np.random.seed(0)
get_patches = ImagePatchesByAnnotation(0, 1, (3, 3), 1, 1)
patches = samples >> get_patches >> Collect()
assert len(patches) == 3
p, l = patches[0]
img_patch0 = np.array([[12, 13, 14], [17, 18, 19], [22, 23, 24]])
assert l == 0
nt.assert_allclose(p, img_patch0)
p, l = patches[1]
img_patch1 = np.array([[11, 12, 13], [16, 17, 18], [21, 22, 23]])
assert l == 1
nt.assert_allclose(p, img_patch1)
p, l = patches[2]
img_patch1 = np.array([[7, 8, 9], [12, 13, 14], [17, 18, 19]])
assert l == 1
nt.assert_allclose(p, img_patch1)
np.random.seed(0)
get_patches = ImagePatchesByAnnotation(0, 1, (3, 3), 1, 1, retlabel=False)
patches = samples >> get_patches >> Collect()
p, m = patches[0]
img_patch0 = np.array([[12, 13, 14], [17, 18, 19], [22, 23, 24]])
img_mask0 = np.array([[0, 255, 0], [255, 0, 0], [0, 0, 0]])
nt.assert_allclose(m, img_mask0)
nt.assert_allclose(p, img_patch0)
def evaluate(cfg: CFG, net: Network, testdata: Samples) -> (IVec, IVec, float):
net.eval()
with to.no_grad():
tars, preds, probs = (testdata >> LoadGraph() >> MakeBatch(
cfg.batchsize) >> PredBatch(net) >> Unzip())
tars = tars >> Flatten() >> Collect()
preds = preds >> Flatten() >> Collect()
acc = 100.0 * [t == p
for t, p in zip(tars, preds)].count(True) / len(tars)
return tars, preds, acc
def test_TransformImage():
TransformImage.register('fake_trans1', lambda e: e + 1)
TransformImage.register('fake_trans2', lambda e, x: e + x)
samples = [(1, 2), (3, 4)]
transform = TransformImage(0).by('fake_trans1')
assert samples >> transform >> Collect() == [(2, 2), (4, 4)]
transform = TransformImage((0, 1)).by('fake_trans1').by('fake_trans2', 3)
assert samples >> transform >> Collect() == [(5, 6), (7, 8)]
def test_AugmentImage():
TransformImage.register('fake_trans1', lambda e: e + 1)
TransformImage.register('fake_trans2', lambda e: e + 2)
samples = [(1, 2), (3, 4)]
augment = AugmentImage(0).by('fake_trans1', 1.0).by('fake_trans2', 1.0)
assert samples >> augment >> Collect() == [(2, 2), (3, 2), (4, 4), (5, 4)]
augment = AugmentImage(0).by('fake_trans1', 1.0).by('fake_trans2', 0.0)
assert samples >> augment >> Collect() == [(2, 2), (4, 4)]
def test_DplyToList():
empty_dplyframe = dp.DplyFrame(pd.DataFrame())
assert empty_dplyframe >> DplyToList() >> Collect() == []
pandasframe = pd.DataFrame(data={'c1': [1, 2, 3], 'c2': [4, 5, 6]})
dplyframe = dp.DplyFrame(pandasframe)
assert dplyframe >> DplyToList() >> Collect() == [[1, 4], [2, 5], [3, 6]]
with pytest.raises(ValueError) as ex:
[1] >> DplyToList() >> Collect()
assert str(ex.value) == 'Expect Dplyr dataframe!'
def test_LogToFile_reset(filepath):
data = [[1, 2], [3, 4]]
with LogToFile(filepath, cols=0, reset=True) as logtofile:
assert data >> logtofile >> Collect() == data
with open(filepath) as f:
assert f.read() == '1\n3\n'
with LogToFile(filepath, cols=1, reset=False) as logtofile:
assert data >> logtofile >> Collect() == data
with open(filepath) as f:
assert f.read() == '1\n3\n2\n4\n'
def test_LogToFile_numpy(filepath):
data = [np.array([1, 2]), np.array([3, 4])]
with LogToFile(filepath) as logtofile:
assert data >> logtofile >> Collect() == data
with open(filepath) as f:
assert f.read() == '1,2\n3,4\n'
data = [np.array(1), np.array(2)]
with LogToFile(filepath) as logtofile:
assert data >> logtofile >> Collect() == data
with open(filepath) as f:
assert f.read() == '1\n2\n'
def test_ReadLabelDirs():
read = ReadLabelDirs('tests/data/labeldirs', '*.txt')
samples = read >> Collect()
assert samples == [('tests/data/labeldirs/0/test0.txt', '0'),
('tests/data/labeldirs/1/test1.txt', '1'),
('tests/data/labeldirs/1/test11.txt', '1')]
read = ReadLabelDirs('tests/data/labeldirs', '*.txt', '')
samples = read >> Collect()
assert samples == [('tests/data/labeldirs/0/test0.txt', '0'),
('tests/data/labeldirs/1/test1.txt', '1'),
('tests/data/labeldirs/1/test11.txt', '1'),
('tests/data/labeldirs/_2/test2.txt', '_2')]
def test_ImageAnnotationToMask():
img = np.zeros((3, 3), dtype='uint8')
anno = ('point', ((0, 1), (2, 0)))
samples = [(img, anno)]
masks = samples >> ImageAnnotationToMask(0, 1) >> Collect()
expected = np.array([[0, 0, 255], [255, 0, 0], [0, 0, 0]], dtype='uint8')
assert str(masks[0][1]) == str(expected) # nt.assert_allclose fails!
def test_BuildBatch():
numbers = [4.1, 3.2, 1.1]
vectors = [np.array([1, 2, 3]), np.array([2, 3, 4]), np.array([3, 4, 5])]
images = [np.zeros((5, 3)), np.ones((5, 3)), np.ones((5, 3))]
class_ids = [1, 2, 1]
samples = zip(numbers, vectors, images, class_ids)
build_batch = (nb.BuildBatch(2, prefetch=0).input(
0, 'number',
float).input(1, 'vector',
np.uint8).input(2, 'image', np.uint8,
False).output(3, 'one_hot', 'uint8', 3))
batches = samples >> build_batch >> Collect()
assert len(batches) == 2
batch = batches[0]
assert len(batch) == 2, 'Expect inputs and outputs'
ins, outs = batch
assert len(ins) == 3, 'Expect three input columns in batch'
assert len(outs) == 1, 'Expect one output column in batch'
assert np.array_equal(ins[0], nb.build_number_batch(numbers[:2], float))
assert np.array_equal(ins[1], nb.build_vector_batch(vectors[:2], 'uint8'))
assert np.array_equal(ins[2], nb.build_image_batch(images[:2], 'uint8'))
assert np.array_equal(outs[0],
nb.build_one_hot_batch(class_ids[:2], 'uint8', 3))
def Boost(iterable, batcher, network, targetcol=-1):
"""
iterable >> Boost(batcher, network, targetcol=-1)
Boost samples with high softmax probability for incorrect class.
Expects one-hot encoded targets and softmax predictions for output.
| network = Network()
| build_batch = BuildBatch(BATCHSIZE, colspec)
| boost = Boost(build_batch, network)
| samples >> boost >> build_batch >> network.train() >> Consume()
:param iterable iterable: Iterable with samples.
:param nutsml.BuildBatch batcher: Batcher used for network training.
:param nutsml.Network network: Network used for prediction
:param int targetcol: Column in sample that contains target values.
:return: Generator over samples to boost
:rtype: generator
"""
def do_boost(probs, target):
assert len(target) > 1, 'Expect one-hot encoded target: ' + str(target)
assert len(target) == len(probs), 'Expect softmax probs: ' + str(probs)
return random() > probs[np.argmax(target)]
samples1, samples2 = iterable >> Tee(2)
for batch in samples1 >> batcher:
p_batch, target = batch[:targetcol], batch[targetcol]
pred = [p_batch] >> network.predict() >> Collect()
for p, t, s in zip(pred, target, samples2):
if do_boost(p, t):
yield s
def test_LogCols(filepath):
data = [[1, 2], [3, 4]]
with LogCols(filepath) as logcols:
assert data >> logcols >> Collect() == data
with open(filepath) as f:
assert f.read() == '1,2\n3,4\n'
def test_LogToFile_delete(filepath):
data = [[1, 2], [3, 4]]
logtofile = LogToFile(filepath)
assert data >> logtofile >> Collect() == data
assert os.path.exists(filepath)
logtofile.delete()
assert not os.path.exists(filepath)
def test_SplitRandom_constraint():
same_letter = lambda t: t[0]
data = zip('aabbccddee', range(10))
train, val = data >> SplitRandom(rand=StableRandom(0), ratio=0.6,
constraint=same_letter) >> Collect()
print(train)
print(val)
assert train == [('a', 1), ('a', 0), ('d', 7), ('b', 2), ('d', 6), ('b', 3)]
assert val == [('c', 5), ('e', 8), ('e', 9), ('c', 4)]
def test_AugmentImage():
TransformImage.register('fake_trans1', lambda e: e + 1)
TransformImage.register('fake_trans2', lambda e: e + 2)
samples = [(1, 2), (3, 4)]
augment = AugmentImage(0).by('fake_trans1', 1.0).by('fake_trans2', 1.0)
assert samples >> augment >> Collect() == [(2, 2), (3, 2), (4, 4), (5, 4)]
augment = AugmentImage(0).by('fake_trans1', 1.0).by('fake_trans2', 0.0)
assert samples >> augment >> Collect() == [(2, 2), (4, 4)]
samples = [(x, ) for x in range(1000)]
augment = AugmentImage(0).by('identical', 0.5)
n = samples >> augment >> Count()
assert 450 < n < 550
augment = AugmentImage(0).by('identical', 10)
assert [(1, )] >> augment >> Count() == 10
def test_Stratify():
samples = [('pos', 1)] * 1000 + [('neg', 0)] * 100
dist = samples >> CountValues(1)
stratify = Stratify(1, dist, rand=StableRandom(0))
stratified1 = samples >> stratify >> Collect()
stratified2 = samples >> stratify >> Collect()
assert stratified1 != stratified2
dist1 = stratified1 >> Get(1) >> CountValues()
print(dist1)
assert dist1[0] == 100
assert 90 < dist1[1] < 110
dist2 = stratified2 >> Get(1) >> CountValues()
print(dist2)
assert dist1[0] == 100
assert 90 < dist1[1] < 110
def test_BuildBatch_exceptions():
class_ids = [1, 2]
numbers = [4.1, 3.2]
samples = zip(numbers, class_ids)
with pytest.raises(ValueError) as ex:
build_batch = (nb.BuildBatch(2, prefetch=0).by(0, 'number', float).by(
1, 'invalid', 'uint8', 3))
samples >> build_batch >> Collect()
assert str(ex.value).startswith('Invalid builder')
def test_CheckNaN():
assert [1, 2] >> CheckNaN() >> Collect() == [1, 2]
with pytest.raises(RuntimeError) as ex:
[1, np.NaN, 3] >> CheckNaN() >> Consume()
assert str(ex.value).startswith('NaN encountered')
with pytest.raises(RuntimeError) as ex:
[(1, np.NaN), (2, 4)] >> CheckNaN() >> Consume()
assert str(ex.value).startswith('NaN encountered')
def test_RegularImagePatches():
img1 = np.reshape(np.arange(12), (3, 4))
samples = [(img1, 0)]
get_patches = RegularImagePatches(0, (2, 2), 2)
expected = [(np.array([[0, 1], [4, 5]]), 0), (np.array([[2, 3], [6,
7]]), 0)]
patches = samples >> get_patches >> Collect()
for (p, ps), (e, es) in zip(patches, expected):
nt.assert_allclose(p, e)
assert ps == es
samples = [(img1, img1 + 1)]
get_patches = RegularImagePatches((0, 1), (1, 1), 3)
expected = [(np.array([[0]]), np.array([[1]])),
(np.array([[3]]), np.array([[4]]))]
patches = samples >> get_patches >> Collect()
for p, e in zip(patches, expected):
nt.assert_allclose(p, e)
def test_SplitRandom_constraint():
same_letter = lambda t: t[0]
data = zip('aabbccddee', range(10))
train, val = data >> SplitRandom(
rand=None, ratio=0.6, constraint=same_letter) >> Collect()
train.sort()
val.sort()
assert train == [('a', 0), ('a', 1), ('b', 2), ('b', 3), ('d', 6),
('d', 7)]
assert val == [('c', 4), ('c', 5), ('e', 8), ('e', 9)]
def test_ImageChannelMean():
meansfile = 'tests/data/temp_image_channel_mean.npy'
img1 = np.dstack([np.ones((3, 3)), np.ones((3, 3))])
img2 = np.dstack([np.ones((3, 3)), np.ones((3, 3)) * 3])
samples = [(img1, ), (img2, )]
# provide means directly
img_mean = ImageChannelMean(0, means=[1, 2])
expected = np.array([[[1., 2.]]])
nt.assert_allclose(img_mean.means, expected)
# compute means
img_mean = ImageChannelMean(0, filepath=meansfile)
results = samples >> img_mean.train() >> Collect()
expected = np.array([[[1., 2.]]])
nt.assert_allclose(img_mean.means, expected)
assert results == samples
assert os.path.exists(meansfile)
# re-loading means from file
img_mean = ImageChannelMean(0, filepath=meansfile)
nt.assert_allclose(img_mean.means, expected)
# subtract means
results = samples >> img_mean >> Collect()
expected0 = np.array([[0., 0., 0.], [0., 0., 0.], [0., 0., 0.]])
expected1 = np.array([[1., 1., 1.], [1., 1., 1.], [1., 1., 1.]])
nt.assert_allclose(results[1][0][:, :, 0], expected0)
nt.assert_allclose(results[1][0][:, :, 1], expected1)
with pytest.raises(ValueError) as ex:
other_samples = [(np.eye(5), )]
img_mean = ImageChannelMean(0, filepath=meansfile)
other_samples >> img_mean >> Collect()
assert str(ex.value).startswith('Mean loaded was computed on different')
with pytest.raises(ValueError) as ex:
img_mean = ImageChannelMean(0, 'file does not exist')
samples >> img_mean >> Collect()
assert str(ex.value).startswith('Mean has not yet been computed!')
os.remove(meansfile)
def test_ReadImage():
arr0 = np.load('tests/data/img_arrays/nut_color.jpg.npy')
arr1 = np.load('tests/data/img_arrays/nut_grayscale.jpg.npy')
samples = [('nut_color', 1), ('nut_grayscale', 2)]
imagepath = 'tests/data/img_formats/*.jpg'
img_samples = samples >> ReadImage(0, imagepath) >> Collect()
nt.assert_equal(img_samples[0][0], arr0)
nt.assert_equal(img_samples[1][0], arr1)
assert img_samples[0][1] == 1
assert img_samples[1][1] == 2
pathfunc = lambda sample: 'tests/data/img_formats/{0}.jpg'.format(*sample)
img_samples = samples >> ReadImage(0, pathfunc) >> Collect()
nt.assert_equal(img_samples[0][0], arr0)
nt.assert_equal(img_samples[1][0], arr1)
samples = [('label', 'tests/data/img_formats/nut_color.jpg')]
img_samples = samples >> ReadImage(1, as_grey=False) >> Collect()
assert img_samples[0][1].shape == (213, 320, 3)
img_samples = samples >> ReadImage(1, as_grey=True) >> Collect()
assert img_samples[0][1].shape == (213, 320)
samples = ['tests/data/img_formats/nut_color.jpg']
img_samples = samples >> ReadImage(None, as_grey=False) >> Collect()
assert img_samples[0][0].shape == (213, 320, 3)
samples = ['tests/data/img_formats/nut_color.jpg']
img_samples = samples >> ReadImage(None, dtype=float) >> Collect()
assert img_samples[0][0].dtype == float
def DISABLED_test_plotlines():
filepath = 'tests/data/temp_plotlines.png'
xs = np.arange(0, 6.3, 1.2)
ysin, ycos = np.sin(xs), np.cos(xs)
data = zip(xs, ysin, ycos)
out = data >> pl.PlotLines(1, 0, filepath=filepath) >> Collect()
assert out == data
expected = 'tests/data/img/plotlines.png'
image = ni.load_image(filepath)
os.remove(filepath)
assert_equal_image(expected, image)
def test_Stratify():
samples = [('pos', 1), ('pos', 1), ('neg', 0)]
stratify = Stratify(1, mode='up', rand=StableRandom(0))
stratified = samples >> stratify >> Sort()
assert stratified == [('neg', 0), ('neg', 0), ('pos', 1), ('pos', 1)]
samples = [('pos', 1), ('pos', 1), ('pos', 1), ('neg1', 0), ('neg2', 0)]
stratify = Stratify(1, mode='downrnd', rand=StableRandom(0))
stratified = samples >> stratify >> Sort()
assert stratified == [('neg1', 0), ('neg2', 0), ('pos', 1), ('pos', 1)]
with pytest.raises(ValueError) as ex:
samples >> Stratify(1, mode='invalid') >> Collect()
assert str(ex.value).startswith('Unknown mode')
def test_LogToFile(filepath):
data = [[1, 2], [3, 4]]
with LogToFile(filepath) as logtofile:
assert data >> logtofile >> Collect() == data
with open(filepath) as f:
assert f.read() == '1,2\n3,4\n'
with LogToFile(filepath, delimiter='; ') as logtofile:
assert data >> logtofile >> Collect() == data
with open(filepath) as f:
assert f.read() == '1; 2\n3; 4\n'
with LogToFile(filepath, cols=0, reset=True) as logtofile:
assert data >> logtofile >> Collect() == data
assert data >> logtofile >> Collect() == data
with open(filepath) as f:
assert f.read() == '1\n3\n1\n3\n'
with LogToFile(filepath, cols=(1, 0), colnames=('a', 'b')) as logtofile:
assert data >> logtofile >> Collect() == data
with open(filepath) as f:
assert f.read() == 'a,b\n2,1\n4,3\n'