def test_dense_layer(self):
in_channels = 512
out_channels = 8
x = Variable(FloatTensor(1, in_channels, 32, 64))
print('x:', x.shape)
layer_params = [
{},
{
'dropout': 0.2
},
{
'bottleneck_ratio': 4
},
{
'bottleneck_ratio': 4,
'dropout': 0.2
},
]
for params in layer_params:
with self.subTest(**params):
dense_layer = DenseLayer(in_channels, out_channels, **params)
print(dense_layer)
print('Parameters:', count_parameters(dense_layer))
y = dense_layer(x)
print('y:', y.shape)
self.assertEqual(y.shape[1], out_channels)
def test_dense_block(self):
in_channels = 64
x = Variable(FloatTensor(1, in_channels, 32, 64))
print('x:', x.shape)
block_params = [
{'growth_rate': 4, 'num_layers': 2, 'concat_input': False},
{'growth_rate': 4, 'num_layers': 2, 'concat_input': True},
{'growth_rate': 4, 'num_layers': 3, 'concat_input': False},
{'growth_rate': 4, 'num_layers': 3, 'concat_input': True},
]
dense_layer_params = {'bottleneck_ratio': 4, 'dropout': 0.2}
for params in block_params:
with self.subTest(**params):
dense_block = DenseBlock(in_channels, **params, dense_layer_params=dense_layer_params)
print(dense_block)
print('Parameters:', count_parameters(dense_block))
y = dense_block(x)
print('y:', y.shape)
expected_out_channels = params['growth_rate'] * params['num_layers']
if params['concat_input']:
expected_out_channels += in_channels
self.assertEqual(y.shape[1], expected_out_channels)
def test_fc_densenet(self):
densenet = FCDenseNet(in_channels=self.rgb_channels,
out_channels=self.num_classes,
initial_num_features=24,
dropout=0.2,
down_dense_growth_rates=8,
down_dense_bottleneck_ratios=None,
down_dense_num_layers=(4, 5, 7),
down_transition_compression_factors=1.0,
middle_dense_growth_rate=8,
middle_dense_bottleneck=None,
middle_dense_num_layers=10,
up_dense_growth_rates=8,
up_dense_bottleneck_ratios=None,
up_dense_num_layers=(7, 5, 4))
print(densenet)
print('Layers:', count_conv2d(densenet))
print('Parameters:', count_parameters(densenet))
logits = densenet(self.images)
print('Logits:', logits.shape)
self.assertEqual(
logits.shape,
Size((self.batch_size, self.num_classes, self.H, self.W)))
def test_fc_densenet_103(self):
densenet = FCDenseNet103()
layers = count_conv2d(densenet)
print('Layers:', layers)
print('Parameters:', count_parameters(densenet))
self.assertEqual(103, layers)
def test_bottleneck(self):
in_channels = 512
out_channels = 8
x = torch.empty(1, in_channels, 32, 64)
print('x:', x.shape)
bottleneck = Bottleneck(in_channels, out_channels)
print(bottleneck)
print('Parameters:', count_parameters(bottleneck))
y = bottleneck(x)
print('y:', y.shape)
self.assertEqual(y.shape[1], out_channels)
def test_bottleneck(self):
in_channels = 512
out_channels = 8
x = Variable(FloatTensor(1, in_channels, 32, 64))
print('x:', x.shape)
bottleneck = Bottleneck(in_channels, out_channels)
print(bottleneck)
print('Parameters:', count_parameters(bottleneck))
y = bottleneck(x)
print('y:', y.shape)
self.assertEqual(y.shape[1], out_channels)
def test_transition_up(self):
# Transposed convolution upsamples this to (:, :, 13, 17)
upsample = Variable(FloatTensor(1, 3, 6, 8).zero_())
skip = Variable(FloatTensor(1, 7, 15, 14).zero_())
print('upsample:', upsample.shape)
print('skip:', skip.shape)
transition = TransitionUp(upsample.size(1))
print(transition)
print('Parameters:', count_parameters(transition))
res = transition(upsample, skip)
print('res:', res.shape)
self.assertEqual(res.size(1), upsample.size(1) + skip.size(1))
self.assertEqual(res.size(2), min(13, skip.size(2)))
self.assertEqual(res.size(3), min(17, skip.size(3)))
def test_transition(self):
in_channels = 512
H = 32
W = 64
x = Variable(FloatTensor(1, in_channels, H, W))
print('x:', x.shape)
for c in [1.0, 0.9, 0.001]:
with self.subTest(compression=c):
transition = Transition(in_channels, compression=c)
print(transition)
print('Parameters:', count_parameters(transition))
y = transition(x)
print('y:', y.shape)
self.assertEqual(y.shape[1], int(ceil(c * in_channels)))
self.assertEqual(y.shape[2], H // 2)
self.assertEqual(y.shape[3], W // 2)
def test_densenet(self):
densenets = [
DenseNet(self.rgb_channels, self.imagenet_classes),
DenseNet121(),
DenseNet161(),
DenseNet169(),
DenseNet201(),
]
for densenet in densenets:
with self.subTest(klass=type(densenet).__name__):
print(densenet)
layers = count_conv2d(densenet)
print('Layers:', layers)
print('Parameters:', count_parameters(densenet))
logits = densenet(self.images)
print('Logits:', logits.shape)
self.assertEqual(logits.shape, Size((self.batch_size, self.imagenet_classes)))
def test_transition_down(self):
in_channels = 512
H = 32
W = 64
x = Variable(FloatTensor(1, in_channels, H, W))
print('x:', x.shape)
for d in [0.0, 0.2]:
with self.subTest(dropout=d):
transition = TransitionDown(in_channels, dropout=d)
print(transition)
print('Parameters:', count_parameters(transition))
y = transition(x)
print('y:', y.shape)
self.assertEqual(y.shape[1], in_channels)
self.assertEqual(y.shape[2], H // 2)
self.assertEqual(y.shape[3], W // 2)
