def test_variable_kernel_size(self):
width = 32
cv = networks.ConvChain(self.c, ksize=[3, 5, 3], depth=3, width=width)
# Check sizes and children length is correct
self.assertListEqual(list(cv.conv0.conv.weight.shape),
[width, self.c, 3, 3])
self.assertListEqual(list(cv.conv1.conv.weight.shape),
[width, width, 5, 5])
self.assertListEqual(list(cv.conv2.conv.weight.shape),
[width, width, 3, 3])
def __init__(self, depth=2):
super(BasicModel, self).__init__()
assert depth > 1, "depth should be > 1"
self.net = networks.ConvChain(3,
width=16,
ksize=3,
depth=depth,
activation='leaky_relu')
self.regressor = networks.ConvModule(16, 3, ksize=1)
def test_even_no_padding(self):
w = 32
k = 4
depth = 2
cv = networks.ConvChain(self.c,
depth=depth,
width=w,
ksize=k,
pad=False)
out_ = cv(self.in_data)
self.assertEqual(out_.shape[2], self.h - depth * (k - 1))
self.assertEqual(out_.shape[3], self.w - depth * (k - 1))
def test_normalization(self):
w = 32
k = 3
cv = networks.ConvChain(self.c,
depth=4,
width=w,
ksize=k,
norm_layer="batch")
self.assertEqual(len(list(cv.conv0.children())), 3)
self.assertEqual(len(list(cv.conv1.children())), 3)
self.assertEqual(len(list(cv.conv2.children())), 3)
self.assertEqual(len(list(cv.conv3.children())), 3)
self.assertIsNotNone(cv.conv0.norm)
self.assertIsNotNone(cv.conv1.norm)
self.assertIsNotNone(cv.conv2.norm)
self.assertIsNotNone(cv.conv3.norm)
def test_variable_width(self):
# Width should have 3-1 = 2 values
self.assertRaises(AssertionError,
networks.ConvChain,
self.c,
depth=3,
width=[12])
k = 3
cv = networks.ConvChain(self.c, ksize=k, depth=3, width=[12, 24, 21])
# Check sizes and children length is correct
self.assertListEqual(list(cv.conv0.conv.weight.shape),
[12, self.c, k, k])
self.assertListEqual(list(cv.conv1.conv.weight.shape), [24, 12, k, k])
self.assertListEqual(list(cv.conv2.conv.weight.shape), [21, 24, k, k])
self.assertRaises(AttributeError, getattr, cv, "conv3")
self.assertEqual(len(list(cv.children())), 3)
def test_strided(self):
k = 3
# 3 strides should be passed
self.assertRaises(AssertionError,
networks.ConvChain,
self.c,
ksize=k,
depth=3,
width=32,
strides=[1, 2])
cv = networks.ConvChain(self.c,
ksize=k,
depth=3,
width=32,
strides=[1, 2, 2])
# Check sizes and children length is correct
self.assertEqual(cv.conv0.conv.stride[0], 1)
self.assertEqual(cv.conv1.conv.stride[0], 2)
self.assertEqual(cv.conv2.conv.stride[0], 2)
def test_default(self):
w = 32
k = 3
cv = networks.ConvChain(self.c, depth=5, width=w, ksize=k)
self.assertListEqual(list(cv.conv0.conv.weight.shape),
[w, self.c, k, k])
self.assertListEqual(list(cv.conv1.conv.weight.shape), [w, w, k, k])
self.assertListEqual(list(cv.conv2.conv.weight.shape), [w, w, k, k])
self.assertListEqual(list(cv.conv3.conv.weight.shape), [w, w, k, k])
self.assertListEqual(list(cv.conv4.conv.weight.shape), [w, w, k, k])
self.assertRaises(AttributeError, getattr, cv, "conv5")
self.assertEqual(len(list(cv.children())), 5)
self.assertEqual(len(list(cv.conv0.children())), 2)
self.assertEqual(len(list(cv.conv1.children())), 2)
self.assertEqual(len(list(cv.conv2.children())), 2)
self.assertEqual(len(list(cv.conv3.children())), 2)
self.assertEqual(len(list(cv.conv4.children())), 2)
out_ = cv(self.in_data)
self.assertListEqual(list(out_.shape), [self.bs, w, self.h, self.w])
def test_output_activation(self):
w = 32
k = 3
cv = networks.ConvChain(self.c, depth=5, width=w, ksize=k)
self.assertEqual(len(list(cv.conv4.children())), 2)
def test_even_padding(self):
w = 32
k = 4
cv = networks.ConvChain(self.c, depth=3, width=w, ksize=k)
out_ = cv(self.in_data)
self.assertEqual(out_.shape[1], w)