def test_compute_strided_conv():
conv = Conv2D(conv_format=TensorFlowNHWC, stride=2)
features = np.arange(16, dtype=np.float32).reshape(1, 4, 4, 1)
filters = np.arange(4, dtype=np.float32).reshape(2, 2, 1, 1)
biases = np.zeros((1, 1))
conv.check_incoming_shapes(StaticShape.of_tensor(features),
StaticShape.of_tensor(filters),
StaticShape.of_tensor(biases))
conv.compute(features, filters, biases)
actual = conv.output
assert actual.shape == (1, 2, 2, 1)
np.testing.assert_allclose(
actual, np.array([[[[24.], [36.]], [[72.], [84.]]]], dtype=np.float32))
dx, df, db = conv.partials(np.ones(actual.shape))
assert dx.shape == features.shape
assert df.shape == filters.shape
assert db.shape == biases.shape
np.testing.assert_allclose(
dx,
np.array([[[[0.], [1.], [0.], [1.]], [[2.], [3.], [2.], [3.]],
[[0.], [1.], [0.], [1.]], [[2.], [3.], [2.], [3.]]]],
dtype=np.float32))
np.testing.assert_allclose(
df, np.array([[[[20.]], [[24.]]], [[[36.]], [[40.]]]],
dtype=np.float32))
def test_valid_shape_without_padding_and_stride():
conv = Conv2D(padding=0, stride=1)
m = Dim.unknown()
features = StaticShape(m, 3, 40, 30)
filters = StaticShape(16, 3, 9, 9)
biases = StaticShape(16, 1)
conv.check_incoming_shapes(features, filters, biases)
out_shape = conv.compute_out_shape(features, filters, biases)
assert out_shape == StaticShape(m, 16, 32, 22)
def test_invalid_bias():
conv = Conv2D()
m = Dim.unknown()
features = StaticShape(m, 3, 40, 30)
filters = StaticShape(16, 3, 2, 2)
biases = StaticShape(17, 1)
with pytest.raises(ValueError) as ex:
conv.check_incoming_shapes(features, filters, biases)
assert str(ex.value) == f"Number of bias should match number of filters " \
f"but got 17 and 16"
def test_invalid_number_of_channels():
conv = Conv2D()
m = Dim.unknown()
features = StaticShape(m, 3, 40, 30)
filters = StaticShape(16, 4, 2, 2)
biases = StaticShape(16, 1)
with pytest.raises(ValueError) as ex:
conv.check_incoming_shapes(features, filters, biases)
assert str(ex.value) == f"Number of channels should be the same " \
f"in features(3) and filters(4)"
def test_compute_valid_conv():
conv = Conv2D(conv_format=TensorFlowNHWC)
features = np.arange(72, dtype=np.float32).reshape(2, 3, 4, 3)
filters = np.arange(24, dtype=np.float32).reshape(2, 2, 3, 2)
biases = np.zeros((2, 1))
conv.check_incoming_shapes(StaticShape.of_tensor(features),
StaticShape.of_tensor(filters),
StaticShape.of_tensor(biases))
conv.compute(features, filters, biases)
actual = conv.output
assert actual.shape == (2, 2, 3, 2)
np.testing.assert_allclose(
actual,
np.array([[[[1624., 1726.], [2020., 2158.], [2416., 2590.]],
[[3208., 3454.], [3604., 3886.], [4000., 4318.]]],
[[[6376., 6910.], [6772., 7342.], [7168., 7774.]],
[[7960., 8638.], [8356., 9070.], [8752., 9502.]]]],
dtype=np.float32))
dfeatures, dfilters, dbias = conv.partials(np.ones(actual.shape))
assert dfeatures.shape == features.shape
assert dfilters.shape == filters.shape
assert dbias.shape == biases.shape
np.testing.assert_allclose(
dfeatures,
np.array([[[[1., 5., 9.], [14., 22., 30.], [14., 22., 30.],
[13., 17., 21.]],
[[26., 34., 42.], [76., 92., 108.], [76., 92., 108.],
[50., 58., 66.]],
[[25., 29., 33.], [62., 70., 78.], [62., 70., 78.],
[37., 41., 45.]]],
[[[1., 5., 9.], [14., 22., 30.], [14., 22., 30.],
[13., 17., 21.]],
[[26., 34., 42.], [76., 92., 108.], [76., 92., 108.],
[50., 58., 66.]],
[[25., 29., 33.], [62., 70., 78.], [62., 70., 78.],
[37., 41., 45.]]]],
dtype=np.float32))
np.testing.assert_allclose(
dfilters,
np.array([[[[324., 324.], [336., 336.], [348., 348.]],
[[360., 360.], [372., 372.], [384., 384.]]],
[[[468., 468.], [480., 480.], [492., 492.]],
[[504., 504.], [516., 516.], [528., 528.]]]],
dtype=np.float32))
def test_compute_same_conv():
conv = Conv2D(conv_format=TensorFlowNHWC, padding=1)
features = np.arange(16, dtype=np.float32).reshape(1, 4, 4, 1)
filters = np.arange(9, dtype=np.float32).reshape(3, 3, 1, 1)
biases = np.zeros((1, 1))
conv.check_incoming_shapes(StaticShape.of_tensor(features),
StaticShape.of_tensor(filters),
StaticShape.of_tensor(biases))
conv.compute(features, filters, biases)
actual = conv.output
assert actual.shape == (1, 4, 4, 1)
np.testing.assert_allclose(
actual,
np.array([[[[73.], [121.], [154.], [103.]],
[[171.], [258.], [294.], [186.]],
[[279.], [402.], [438.], [270.]],
[[139.], [187.], [202.], [113.]]]],
dtype=np.float32))
dx, df, db = conv.partials(np.ones(actual.shape))
assert dx.shape == features.shape
assert df.shape == filters.shape
assert db.shape == biases.shape
np.testing.assert_allclose(
dx,
np.array(
[[[[8.], [15.], [15.], [12.]], [[21.], [36.], [36.], [27.]],
[[21.], [36.], [36.], [27.]], [[20.], [33.], [33.], [24.]]]],
dtype=np.float32))
np.testing.assert_allclose(
df,
np.array([[[[45.]], [[66.]], [[54.]]], [[[84.]], [[120.]], [[96.]]],
[[[81.]], [[114.]], [[90.]]]],
dtype=np.float32))