def __init__(self,
in_channels,
mid_channels,
out_channels,
kernel_size=3,
norm_cfg=dict(type='BN'),
act_cfg=dict(type='ReLU'),
init_cfg=None,
**kwargs):
super(CorrelationHead, self).__init__(init_cfg)
self.kernel_convs = ConvModule(in_channels=in_channels,
out_channels=mid_channels,
kernel_size=kernel_size,
norm_cfg=norm_cfg,
act_cfg=act_cfg)
self.search_convs = ConvModule(in_channels=in_channels,
out_channels=mid_channels,
kernel_size=kernel_size,
norm_cfg=norm_cfg,
act_cfg=act_cfg)
self.head_convs = nn.Sequential(
ConvModule(in_channels=mid_channels,
out_channels=mid_channels,
kernel_size=1,
norm_cfg=norm_cfg,
act_cfg=act_cfg),
ConvModule(in_channels=mid_channels,
out_channels=out_channels,
kernel_size=1,
act_cfg=None))
def test_revert_mmsyncbn():
if 'SLURM_NTASKS' not in os.environ or int(os.environ['SLURM_NTASKS']) < 2:
print('Must run on slurm with more than 1 process!\n'
'srun -p test --gres=gpu:2 -n2')
return
rank = int(os.environ['SLURM_PROCID'])
world_size = int(os.environ['SLURM_NTASKS'])
local_rank = int(os.environ['SLURM_LOCALID'])
node_list = str(os.environ['SLURM_NODELIST'])
node_parts = re.findall('[0-9]+', node_list)
os.environ['MASTER_ADDR'] = (f'{node_parts[1]}.{node_parts[2]}' +
f'.{node_parts[3]}.{node_parts[4]}')
os.environ['MASTER_PORT'] = '12341'
os.environ['WORLD_SIZE'] = str(world_size)
os.environ['RANK'] = str(rank)
dist.init_process_group('nccl')
torch.cuda.set_device(local_rank)
x = torch.randn(1, 3, 10, 10).cuda()
dist.broadcast(x, src=0)
conv = ConvModule(3, 8, 2, norm_cfg=dict(type='MMSyncBN')).cuda()
conv.eval()
y_mmsyncbn = conv(x).detach().cpu().numpy()
conv = revert_sync_batchnorm(conv)
y_bn = conv(x).detach().cpu().numpy()
assert np.all(np.isclose(y_bn, y_mmsyncbn, 1e-3))
conv, x = conv.to('cpu'), x.to('cpu')
y_bn_cpu = conv(x).detach().numpy()
assert np.all(np.isclose(y_bn, y_bn_cpu, 1e-3))
def __init__(self,
in_channels,
out_channels=2048,
norm_cfg=dict(type='BN', momentum=0.1),
init_cfg=dict(type='Normal', layer='Linear', std=0.01)):
super(HRFuseScales, self).__init__(init_cfg=init_cfg)
self.in_channels = in_channels
self.out_channels = out_channels
self.norm_cfg = norm_cfg
block_type = Bottleneck
out_channels = [128, 256, 512, 1024]
# Increase the channels on each resolution
# from C, 2C, 4C, 8C to 128, 256, 512, 1024
increase_layers = []
for i in range(len(in_channels)):
increase_layers.append(
ResLayer(
block_type,
in_channels=in_channels[i],
out_channels=out_channels[i],
num_blocks=1,
stride=1,
))
self.increase_layers = nn.ModuleList(increase_layers)
# Downsample feature maps in each scale.
downsample_layers = []
for i in range(len(in_channels) - 1):
downsample_layers.append(
ConvModule(
in_channels=out_channels[i],
out_channels=out_channels[i + 1],
kernel_size=3,
stride=2,
padding=1,
norm_cfg=self.norm_cfg,
bias=False,
))
self.downsample_layers = nn.ModuleList(downsample_layers)
# The final conv block before final classifier linear layer.
self.final_layer = ConvModule(
in_channels=out_channels[3],
out_channels=self.out_channels,
kernel_size=1,
norm_cfg=self.norm_cfg,
bias=False,
)
def __init__(self,
in_channels,
out_channels,
mid_channels,
kernel_size=3,
stride=1,
se_cfg=None,
with_residual=True,
conv_cfg=None,
norm_cfg=dict(type='BN'),
act_cfg=dict(type='ReLU'),
drop_path_rate=0.,
with_cp=False,
init_cfg=None,
**kwargs):
super(EdgeResidual, self).__init__(init_cfg=init_cfg)
assert stride in [1, 2]
self.with_cp = with_cp
self.drop_path = DropPath(
drop_path_rate) if drop_path_rate > 0 else nn.Identity()
self.with_se = se_cfg is not None
self.with_residual = (stride == 1 and in_channels == out_channels
and with_residual)
if self.with_se:
assert isinstance(se_cfg, dict)
self.conv1 = ConvModule(in_channels=in_channels,
out_channels=mid_channels,
kernel_size=kernel_size,
stride=1,
padding=kernel_size // 2,
conv_cfg=conv_cfg,
norm_cfg=norm_cfg,
act_cfg=act_cfg)
if self.with_se:
self.se = SELayer(**se_cfg)
self.conv2 = ConvModule(in_channels=mid_channels,
out_channels=out_channels,
kernel_size=1,
stride=stride,
padding=0,
conv_cfg=conv_cfg,
norm_cfg=norm_cfg,
act_cfg=None)
def test_revert_syncbn():
conv = ConvModule(3, 8, 2, norm_cfg=dict(type='SyncBN'))
x = torch.randn(1, 3, 10, 10)
# Expect a ValueError prompting that SyncBN is not supported on CPU
with pytest.raises(ValueError):
y = conv(x)
conv = revert_sync_batchnorm(conv)
y = conv(x)
assert y.shape == (1, 8, 9, 9)
def test_bias():
# bias: auto, without norm
conv = ConvModule(3, 8, 2)
assert conv.conv.bias is not None
# bias: auto, with norm
conv = ConvModule(3, 8, 2, norm_cfg=dict(type='BN'))
assert conv.conv.bias is None
# bias: False, without norm
conv = ConvModule(3, 8, 2, bias=False)
assert conv.conv.bias is None
# bias: True, with norm
with pytest.warns(UserWarning) as record:
ConvModule(3, 8, 2, bias=True, norm_cfg=dict(type='BN'))
assert len(record) == 1
assert record[0].message.args[
0] == 'ConvModule has norm and bias at the same time'
def conv_module(in_planes, out_planes, kernel_size=3, padding=1):
# The module's pipeline: Conv -> BN -> ReLU.
return ConvModule(
in_channels=in_planes,
out_channels=out_planes,
kernel_size=kernel_size,
padding=padding,
bias=True,
norm_cfg=dict(type='BN', requires_grad=True),
act_cfg=dict(type='ReLU'),
inplace=True)
def test_bias():
# bias: auto, without norm
conv = ConvModule(3, 8, 2)
assert conv.conv.bias is not None
# bias: auto, with norm
conv = ConvModule(3, 8, 2, norm_cfg=dict(type='BN'))
assert conv.conv.bias is None
# bias: False, without norm
conv = ConvModule(3, 8, 2, bias=False)
assert conv.conv.bias is None
# bias: True, with batch norm
with pytest.warns(UserWarning) as record:
ConvModule(3, 8, 2, bias=True, norm_cfg=dict(type='BN'))
assert len(record) == 1
assert record[0].message.args[
0] == 'Unnecessary conv bias before batch/instance norm'
# bias: True, with instance norm
with pytest.warns(UserWarning) as record:
ConvModule(3, 8, 2, bias=True, norm_cfg=dict(type='IN'))
assert len(record) == 1
assert record[0].message.args[
0] == 'Unnecessary conv bias before batch/instance norm'
# bias: True, with other norm
with pytest.warns(UserWarning) as record:
norm_cfg = dict(type='GN', num_groups=1)
ConvModule(3, 8, 2, bias=True, norm_cfg=norm_cfg)
warnings.warn('No warnings')
assert len(record) == 1
assert record[0].message.args[0] == 'No warnings'
def test_order():
with pytest.raises(AssertionError):
# order must be a tuple
order = ['conv', 'norm', 'act']
ConvModule(3, 8, 2, order=order)
with pytest.raises(AssertionError):
# length of order must be 3
order = ('conv', 'norm')
ConvModule(3, 8, 2, order=order)
with pytest.raises(AssertionError):
# order must be an order of 'conv', 'norm', 'act'
order = ('conv', 'norm', 'norm')
ConvModule(3, 8, 2, order=order)
with pytest.raises(AssertionError):
# order must be an order of 'conv', 'norm', 'act'
order = ('conv', 'norm', 'something')
ConvModule(3, 8, 2, order=order)
# ('conv', 'norm', 'act')
conv = ConvModule(3, 8, 2, norm_cfg=dict(type='BN'))
out = conv('input')
assert out == 'input_conv_bn_relu'
# ('norm', 'conv', 'act')
conv = ConvModule(3,
8,
2,
norm_cfg=dict(type='BN'),
order=('norm', 'conv', 'act'))
out = conv('input')
assert out == 'input_bn_conv_relu'
# ('conv', 'norm', 'act'), activate=False
conv = ConvModule(3, 8, 2, norm_cfg=dict(type='BN'))
out = conv('input', activate=False)
assert out == 'input_conv_bn'
# ('conv', 'norm', 'act'), activate=False
conv = ConvModule(3, 8, 2, norm_cfg=dict(type='BN'))
out = conv('input', norm=False)
assert out == 'input_conv_relu'
def __init__(self,
fp_channels=((768, 256, 256), (384, 256, 256),
(320, 256, 128), (128 + 6, 128, 128, 128)),
fp_norm_cfg=dict(type='BN2d'),
**kwargs):
super(PAConvHead, self).__init__(fp_channels, fp_norm_cfg, **kwargs)
# https://github.com/CVMI-Lab/PAConv/blob/main/scene_seg/model/pointnet2/pointnet2_paconv_seg.py#L53
# PointNet++'s decoder conv has bias while PAConv's doesn't have
# so we need to rebuild it here
self.pre_seg_conv = ConvModule(fp_channels[-1][-1],
self.channels,
kernel_size=1,
bias=False,
conv_cfg=self.conv_cfg,
norm_cfg=self.norm_cfg,
act_cfg=self.act_cfg)
def __init__(self,
fp_channels=((768, 256, 256), (384, 256, 256),
(320, 256, 128), (128, 128, 128, 128)),
**kwargs):
super(PointNet2Head, self).__init__(**kwargs)
self.num_fp = len(fp_channels)
self.FP_modules = nn.ModuleList()
for cur_fp_mlps in fp_channels:
self.FP_modules.append(PointFPModule(mlp_channels=cur_fp_mlps))
# https://github.com/charlesq34/pointnet2/blob/master/models/pointnet2_sem_seg.py#L40
self.pre_seg_conv = ConvModule(fp_channels[-1][-1],
self.channels,
kernel_size=1,
bias=True,
conv_cfg=self.conv_cfg,
norm_cfg=self.norm_cfg,
act_cfg=self.act_cfg)
def test_revert_sync_batchnorm():
conv_syncbn = ConvModule(3, 8, 2, norm_cfg=dict(type='SyncBN')).to('cpu')
conv_syncbn.train()
x = torch.randn(1, 3, 10, 10)
# Will raise an ValueError saying SyncBN does not run on CPU
with pytest.raises(ValueError):
y = conv_syncbn(x)
conv_bn = revert_sync_batchnorm(conv_syncbn)
y = conv_bn(x)
assert y.shape == (1, 8, 9, 9)
assert conv_bn.training == conv_syncbn.training
conv_syncbn.eval()
conv_bn = revert_sync_batchnorm(conv_syncbn)
assert conv_bn.training == conv_syncbn.training
def __init__(self,
num_convs=1,
channels=256,
kernel_size=3,
norm_cfg=None,
act_cfg=dict(type='ReLU')):
super(EmbedAggregator, self).__init__()
assert num_convs > 0, 'The number of convs must be bigger than 1.'
self.embed_convs = nn.ModuleList()
for i in range(num_convs):
if i == num_convs - 1:
new_norm_cfg = None
new_act_cfg = None
else:
new_norm_cfg = norm_cfg
new_act_cfg = act_cfg
self.embed_convs.append(
ConvModule(in_channels=channels,
out_channels=channels,
kernel_size=kernel_size,
padding=(kernel_size - 1) // 2,
norm_cfg=new_norm_cfg,
act_cfg=new_act_cfg))
def test_conv_module():
with pytest.raises(AssertionError):
# conv_cfg must be a dict or None
conv_cfg = 'conv'
ConvModule(3, 8, 2, conv_cfg=conv_cfg)
with pytest.raises(AssertionError):
# norm_cfg must be a dict or None
norm_cfg = 'norm'
ConvModule(3, 8, 2, norm_cfg=norm_cfg)
with pytest.raises(KeyError):
# softmax is not supported
act_cfg = dict(type='softmax')
ConvModule(3, 8, 2, act_cfg=act_cfg)
# conv + norm + act
conv = ConvModule(3, 8, 2, norm_cfg=dict(type='BN'))
assert conv.with_activation
assert hasattr(conv, 'activate')
assert conv.with_norm
assert hasattr(conv, 'norm')
x = torch.rand(1, 3, 256, 256)
output = conv(x)
assert output.shape == (1, 8, 255, 255)
# conv + act
conv = ConvModule(3, 8, 2)
assert conv.with_activation
assert hasattr(conv, 'activate')
assert not conv.with_norm
assert not hasattr(conv, 'norm')
x = torch.rand(1, 3, 256, 256)
output = conv(x)
assert output.shape == (1, 8, 255, 255)
# conv
conv = ConvModule(3, 8, 2, act_cfg=None)
assert not conv.with_norm
assert not hasattr(conv, 'norm')
assert not conv.with_activation
assert not hasattr(conv, 'activate')
x = torch.rand(1, 3, 256, 256)
output = conv(x)
assert output.shape == (1, 8, 255, 255)
# with_spectral_norm=True
conv = ConvModule(3, 8, 3, padding=1, with_spectral_norm=True)
assert hasattr(conv.conv, 'weight_orig')
output = conv(x)
assert output.shape == (1, 8, 256, 256)
# padding_mode='reflect'
conv = ConvModule(3, 8, 3, padding=1, padding_mode='reflect')
assert isinstance(conv.padding_layer, nn.ReflectionPad2d)
output = conv(x)
assert output.shape == (1, 8, 256, 256)
# non-existing padding mode
with pytest.raises(KeyError):
conv = ConvModule(3, 8, 3, padding=1, padding_mode='non_exists')
# leaky relu
conv = ConvModule(3, 8, 3, padding=1, act_cfg=dict(type='LeakyReLU'))
assert isinstance(conv.activate, nn.LeakyReLU)
output = conv(x)
assert output.shape == (1, 8, 256, 256)
def __init__(self,
arch='b0',
drop_path_rate=0.,
out_indices=(6, ),
frozen_stages=0,
conv_cfg=dict(type='Conv2dAdaptivePadding'),
norm_cfg=dict(type='BN', eps=1e-3),
act_cfg=dict(type='Swish'),
norm_eval=False,
with_cp=False,
init_cfg=[
dict(type='Kaiming', layer='Conv2d'),
dict(type='Constant',
layer=['_BatchNorm', 'GroupNorm'],
val=1)
]):
super(EfficientNet, self).__init__(init_cfg)
assert arch in self.arch_settings, \
f'"{arch}" is not one of the arch_settings ' \
f'({", ".join(self.arch_settings.keys())})'
self.arch_setting = self.arch_settings[arch]
self.layer_setting = self.layer_settings[arch[:1]]
for index in out_indices:
if index not in range(0, len(self.layer_setting)):
raise ValueError('the item in out_indices must in '
f'range(0, {len(self.layer_setting)}). '
f'But received {index}')
if frozen_stages not in range(len(self.layer_setting) + 1):
raise ValueError('frozen_stages must be in range(0, '
f'{len(self.layer_setting) + 1}). '
f'But received {frozen_stages}')
self.drop_path_rate = drop_path_rate
self.out_indices = out_indices
self.frozen_stages = frozen_stages
self.conv_cfg = conv_cfg
self.norm_cfg = norm_cfg
self.act_cfg = act_cfg
self.norm_eval = norm_eval
self.with_cp = with_cp
self.layer_setting = model_scaling(self.layer_setting,
self.arch_setting)
block_cfg_0 = self.layer_setting[0][0]
block_cfg_last = self.layer_setting[-1][0]
self.in_channels = make_divisible(block_cfg_0[1], 8)
self.out_channels = block_cfg_last[1]
self.layers = nn.ModuleList()
self.layers.append(
ConvModule(in_channels=3,
out_channels=self.in_channels,
kernel_size=block_cfg_0[0],
stride=block_cfg_0[3],
padding=block_cfg_0[0] // 2,
conv_cfg=self.conv_cfg,
norm_cfg=self.norm_cfg,
act_cfg=self.act_cfg))
self.make_layer()
# Avoid building unused layers in mmdetection.
if len(self.layers) < max(self.out_indices) + 1:
self.layers.append(
ConvModule(in_channels=self.in_channels,
out_channels=self.out_channels,
kernel_size=block_cfg_last[0],
stride=block_cfg_last[3],
padding=block_cfg_last[0] // 2,
conv_cfg=self.conv_cfg,
norm_cfg=self.norm_cfg,
act_cfg=self.act_cfg))
def __init__(self,
img_scale_factor,
out_indices=[2, 3, 4, 5, 6],
flow_scale_factor=5.0,
flow_img_norm_std=[255.0, 255.0, 255.0],
flow_img_norm_mean=[0.411, 0.432, 0.450]):
super(FlowNetSimple, self).__init__()
self.img_scale_factor = img_scale_factor
self.out_indices = out_indices
self.flow_scale_factor = flow_scale_factor
self.flow_img_norm_mean = flow_img_norm_mean
self.flow_img_norm_std = flow_img_norm_std
self.conv_layers = []
conv_layers_setting = self.arch_setting['conv_layers']
for i in range(len(conv_layers_setting['inplanes'])):
num_convs = conv_layers_setting['num_convs'][i]
kernel_size = conv_layers_setting['kernel_size'][i]
inplanes = conv_layers_setting['inplanes'][i]
if i == len(conv_layers_setting['inplanes']) - 1:
planes = 2 * inplanes
else:
planes = conv_layers_setting['inplanes'][i + 1]
conv_layer = nn.ModuleList()
conv_layer.append(
ConvModule(in_channels=inplanes,
out_channels=planes,
kernel_size=kernel_size,
stride=2,
padding=(kernel_size - 1) // 2,
bias=True,
conv_cfg=dict(type='Conv'),
act_cfg=dict(type='LeakyReLU', negative_slope=0.1)))
for j in range(1, num_convs):
kernel_size = 3 if i == 2 else kernel_size
conv_layer.append(
ConvModule(in_channels=planes,
out_channels=planes,
kernel_size=kernel_size,
stride=1,
padding=(kernel_size - 1) // 2,
bias=True,
conv_cfg=dict(type='Conv'),
act_cfg=dict(type='LeakyReLU',
negative_slope=0.1)))
self.add_module(f'conv{i+1}', conv_layer)
self.conv_layers.append(f'conv{i+1}')
self.deconv_layers = []
self.flow_layers = []
self.upflow_layers = []
deconv_layers_setting = self.arch_setting['deconv_layers']
planes = deconv_layers_setting['inplanes'][-1] // 2
for i in range(len(deconv_layers_setting['inplanes']) - 1, -1, -1):
inplanes = deconv_layers_setting['inplanes'][i]
deconv_layer = ConvModule(in_channels=inplanes,
out_channels=planes,
kernel_size=4,
stride=2,
padding=1,
bias=False,
conv_cfg=dict(type='deconv'),
act_cfg=dict(type='LeakyReLU',
negative_slope=0.1))
self.add_module(f'deconv{i+2}', deconv_layer)
self.deconv_layers.insert(0, f'deconv{i+2}')
flow_layer = ConvModule(in_channels=inplanes,
out_channels=2,
kernel_size=3,
stride=1,
padding=1,
bias=False,
conv_cfg=dict(type='Conv'),
act_cfg=None)
self.add_module(f'predict_flow{i+3}', flow_layer)
self.flow_layers.insert(0, f'predict_flow{i+3}')
upflow_layer = ConvModule(in_channels=2,
out_channels=2,
kernel_size=4,
stride=2,
padding=1,
bias=False,
conv_cfg=dict(type='deconv'),
act_cfg=None)
self.add_module(f'upsample_flow{i+2}', upflow_layer)
self.upflow_layers.insert(0, f'upsample_flow{i+2}')
planes = planes // 2
self.predict_flow = ConvModule(in_channels=planes * (2 + 4) + 2,
out_channels=2,
kernel_size=3,
stride=1,
padding=1,
bias=False,
conv_cfg=dict(type='Conv'),
act_cfg=None)
def test_conv_module():
with pytest.raises(AssertionError):
# conv_cfg must be a dict or None
conv_cfg = 'conv'
ConvModule(3, 8, 2, conv_cfg=conv_cfg)
with pytest.raises(AssertionError):
# norm_cfg must be a dict or None
norm_cfg = 'norm'
ConvModule(3, 8, 2, norm_cfg=norm_cfg)
with pytest.raises(KeyError):
# softmax is not supported
act_cfg = dict(type='softmax')
ConvModule(3, 8, 2, act_cfg=act_cfg)
# conv + norm + act
conv = ConvModule(3, 8, 2, norm_cfg=dict(type='BN'))
assert conv.with_activation
assert hasattr(conv, 'activate')
assert conv.with_norm
assert hasattr(conv, 'norm')
x = torch.rand(1, 3, 256, 256)
output = conv(x)
assert output.shape == (1, 8, 255, 255)
# conv + act
conv = ConvModule(3, 8, 2)
assert conv.with_activation
assert hasattr(conv, 'activate')
assert not conv.with_norm
assert conv.norm is None
x = torch.rand(1, 3, 256, 256)
output = conv(x)
assert output.shape == (1, 8, 255, 255)
# conv
conv = ConvModule(3, 8, 2, act_cfg=None)
assert not conv.with_norm
assert conv.norm is None
assert not conv.with_activation
assert not hasattr(conv, 'activate')
x = torch.rand(1, 3, 256, 256)
output = conv(x)
assert output.shape == (1, 8, 255, 255)
# conv with its own `init_weights` method
conv_module = ConvModule(3,
8,
2,
conv_cfg=dict(type='ExampleConv'),
act_cfg=None)
assert torch.equal(conv_module.conv.conv0.weight, torch.zeros(8, 3, 2, 2))
# with_spectral_norm=True
conv = ConvModule(3, 8, 3, padding=1, with_spectral_norm=True)
assert hasattr(conv.conv, 'weight_orig')
output = conv(x)
assert output.shape == (1, 8, 256, 256)
# padding_mode='reflect'
conv = ConvModule(3, 8, 3, padding=1, padding_mode='reflect')
assert isinstance(conv.padding_layer, nn.ReflectionPad2d)
output = conv(x)
assert output.shape == (1, 8, 256, 256)
# non-existing padding mode
with pytest.raises(KeyError):
conv = ConvModule(3, 8, 3, padding=1, padding_mode='non_exists')
# leaky relu
conv = ConvModule(3, 8, 3, padding=1, act_cfg=dict(type='LeakyReLU'))
assert isinstance(conv.activate, nn.LeakyReLU)
output = conv(x)
assert output.shape == (1, 8, 256, 256)
# tanh
conv = ConvModule(3, 8, 3, padding=1, act_cfg=dict(type='Tanh'))
assert isinstance(conv.activate, nn.Tanh)
output = conv(x)
assert output.shape == (1, 8, 256, 256)
# Sigmoid
conv = ConvModule(3, 8, 3, padding=1, act_cfg=dict(type='Sigmoid'))
assert isinstance(conv.activate, nn.Sigmoid)
output = conv(x)
assert output.shape == (1, 8, 256, 256)
# PReLU
conv = ConvModule(3, 8, 3, padding=1, act_cfg=dict(type='PReLU'))
assert isinstance(conv.activate, nn.PReLU)
output = conv(x)
assert output.shape == (1, 8, 256, 256)
# HSwish
conv = ConvModule(3, 8, 3, padding=1, act_cfg=dict(type='HSwish'))
assert isinstance(conv.activate, HSwish)
output = conv(x)
assert output.shape == (1, 8, 256, 256)
# HSigmoid
conv = ConvModule(3, 8, 3, padding=1, act_cfg=dict(type='HSigmoid'))
assert isinstance(conv.activate, HSigmoid)
output = conv(x)
assert output.shape == (1, 8, 256, 256)
