def __init__(self,
inplanes,
planes,
stride=1,
downsample=None,
conv_cfg=None,
norm_cfg=None):
spconv.SparseModule.__init__(self)
Bottleneck.__init__(self,
inplanes,
planes,
stride=stride,
downsample=downsample,
conv_cfg=conv_cfg,
norm_cfg=norm_cfg)
def _add_conv_branch(self):
"""Add the fc branch which consists of a sequential of conv layers."""
branch_convs = ModuleList()
for i in range(self.num_convs):
branch_convs.append(
Bottleneck(inplanes=self.conv_out_channels,
planes=self.conv_out_channels // 4,
conv_cfg=self.conv_cfg,
norm_cfg=self.norm_cfg))
return branch_convs
def test_resnet_bottleneck():
with pytest.raises(AssertionError):
# Style must be in ['pytorch', 'caffe']
Bottleneck(64, 64, style='tensorflow')
with pytest.raises(AssertionError):
# Allowed positions are 'after_conv1', 'after_conv2', 'after_conv3'
plugins = [
dict(
cfg=dict(type='ContextBlock', ratio=1. / 16),
position='after_conv4')
]
Bottleneck(64, 16, plugins=plugins)
with pytest.raises(AssertionError):
# Need to specify different postfix to avoid duplicate plugin name
plugins = [
dict(
cfg=dict(type='ContextBlock', ratio=1. / 16),
position='after_conv3'),
dict(
cfg=dict(type='ContextBlock', ratio=1. / 16),
position='after_conv3')
]
Bottleneck(64, 16, plugins=plugins)
with pytest.raises(KeyError):
# Plugin type is not supported
plugins = [dict(cfg=dict(type='WrongPlugin'), position='after_conv3')]
Bottleneck(64, 16, plugins=plugins)
# Test Bottleneck with checkpoint forward
block = Bottleneck(64, 16, with_cp=True)
assert block.with_cp
x = torch.randn(1, 64, 56, 56)
x_out = block(x)
assert x_out.shape == torch.Size([1, 64, 56, 56])
# Test Bottleneck style
block = Bottleneck(64, 64, stride=2, style='pytorch')
assert block.conv1.stride == (1, 1)
assert block.conv2.stride == (2, 2)
block = Bottleneck(64, 64, stride=2, style='caffe')
assert block.conv1.stride == (2, 2)
assert block.conv2.stride == (1, 1)
# Test Bottleneck DCN
dcn = dict(type='DCN', deformable_groups=1, fallback_on_stride=False)
with pytest.raises(AssertionError):
Bottleneck(64, 64, dcn=dcn, conv_cfg=dict(type='Conv'))
block = Bottleneck(64, 64, dcn=dcn)
assert isinstance(block.conv2, DeformConvPack)
# Test Bottleneck forward
block = Bottleneck(64, 16)
x = torch.randn(1, 64, 56, 56)
x_out = block(x)
assert x_out.shape == torch.Size([1, 64, 56, 56])
# Test Bottleneck with 1 ContextBlock after conv3
plugins = [
dict(
cfg=dict(type='ContextBlock', ratio=1. / 16),
position='after_conv3')
]
block = Bottleneck(64, 16, plugins=plugins)
assert block.context_block.in_channels == 64
x = torch.randn(1, 64, 56, 56)
x_out = block(x)
assert x_out.shape == torch.Size([1, 64, 56, 56])
# Test Bottleneck with 1 GeneralizedAttention after conv2
plugins = [
dict(
cfg=dict(
type='GeneralizedAttention',
spatial_range=-1,
num_heads=8,
attention_type='0010',
kv_stride=2),
position='after_conv2')
]
block = Bottleneck(64, 16, plugins=plugins)
assert block.gen_attention_block.in_channels == 16
x = torch.randn(1, 64, 56, 56)
x_out = block(x)
assert x_out.shape == torch.Size([1, 64, 56, 56])
# Test Bottleneck with 1 GeneralizedAttention after conv2, 1 NonLocal2D
# after conv2, 1 ContextBlock after conv3
plugins = [
dict(
cfg=dict(
type='GeneralizedAttention',
spatial_range=-1,
num_heads=8,
attention_type='0010',
kv_stride=2),
position='after_conv2'),
dict(cfg=dict(type='NonLocal2D'), position='after_conv2'),
dict(
cfg=dict(type='ContextBlock', ratio=1. / 16),
position='after_conv3')
]
block = Bottleneck(64, 16, plugins=plugins)
assert block.gen_attention_block.in_channels == 16
assert block.nonlocal_block.in_channels == 16
assert block.context_block.in_channels == 64
x = torch.randn(1, 64, 56, 56)
x_out = block(x)
assert x_out.shape == torch.Size([1, 64, 56, 56])
# Test Bottleneck with 1 ContextBlock after conv2, 2 ContextBlock after
# conv3
plugins = [
dict(
cfg=dict(type='ContextBlock', ratio=1. / 16, postfix=1),
position='after_conv2'),
dict(
cfg=dict(type='ContextBlock', ratio=1. / 16, postfix=2),
position='after_conv3'),
dict(
cfg=dict(type='ContextBlock', ratio=1. / 16, postfix=3),
position='after_conv3')
]
block = Bottleneck(64, 16, plugins=plugins)
assert block.context_block1.in_channels == 16
assert block.context_block2.in_channels == 64
assert block.context_block3.in_channels == 64
x = torch.randn(1, 64, 56, 56)
x_out = block(x)
assert x_out.shape == torch.Size([1, 64, 56, 56])