def __init__(self,
in_channels=728,
out_channels=1024,
dilation=1,
atrous=False):
# If atrous mode increase dilation, padding and reset stride
out_dilation = 2 * dilation if atrous else 1 * dilation
last_padding = 2 * dilation if atrous else 1 * dilation
stride = 1 if atrous else 2
super().__init__(main_path=nn.Sequential(
nn_ext.SeparableConv2dLayer(in_channels,
in_channels,
kernel_size=3,
dilation=dilation,
padding=1 * dilation),
nn_ext.SeparableConv2dLayer(in_channels,
out_channels,
kernel_size=3,
dilation=dilation,
padding=1 * dilation),
nn_ext.SeparableConv2dLayer(out_channels,
out_channels,
kernel_size=3,
dilation=out_dilation,
stride=stride,
padding=last_padding)),
skip_path=nn_ext.Conv2dLayer(
in_channels,
out_channels,
kernel_size=1,
dilation=out_dilation,
stride=stride),
aggregator=torch.add)
def __init__(self, low_in_channels, num_classes):
super().__init__()
self.conv_low = nn_ext.Conv2dLayer(low_in_channels,
48,
kernel_size=1)
self.conv_logit = nn.Conv2d(48 + 256,
num_classes,
kernel_size=3,
padding=1)
def __init__(self, in_channels, output_stride=16):
super().__init__()
if output_stride not in {8, 16}:
raise ValueError(
'Invalid output_stride; Supported values: {8, 16}')
dilation_factor = 1 if output_stride == 16 else 2
self.aspp = nn.ModuleList([
nn_ext.Conv2dLayer(in_channels, 256, kernel_size=1,
dilation=1),
nn_ext.Conv2dLayer(in_channels,
256,
kernel_size=3,
dilation=6 * dilation_factor,
padding=6 * dilation_factor),
nn_ext.Conv2dLayer(in_channels,
256,
kernel_size=3,
dilation=12 * dilation_factor,
padding=12 * dilation_factor),
nn_ext.Conv2dLayer(in_channels,
256,
kernel_size=3,
dilation=18 * dilation_factor,
padding=18 * dilation_factor)
])
self.global_avg_pool = nn.Sequential(
nn.AdaptiveAvgPool2d(output_size=(1, 1)),
nn.Conv2d(in_channels, 256, kernel_size=1),
nn.ReLU(inplace=True))
self.output_conv = nn_ext.Conv2dLayer(256 * 4 + 256,
256,
kernel_size=1)
def __init__(self, output_stride=16):
if output_stride not in {8, 16}:
raise ValueError(
'Invalid output_stride; Supported values: {8, 16}')
low_out_channels = 128 if output_stride == 16 else 64
super().__init__(output_stride=16,
out_channels=2048,
low_out_channels=low_out_channels)
# Adjust dilation rates to control output_stride
opts = defaultdict(lambda: {})
if output_stride == 16:
opts['exit_flow_block_0'] = {'atrous': True}
opts['exit_flow_block_1'] = {'dilation': 2, 'padding': 2}
elif output_stride == 8:
opts['entry_flow_block_3'] = {'atrous': True}
opts['middle_flow_block'] = {'dilation': 2}
opts['exit_flow_block_0'] = {'dilation': 2, 'atrous': True}
opts['exit_flow_block_1'] = {'dilation': 4, 'padding': 4}
else:
raise ValueError(
'Invalid output_stride; Supported values: {8, 16}')
entry_flow = nn.Sequential(
OrderedDict([
('block_0',
nn.Sequential(
nn_ext.Conv2dLayer(3,
32,
kernel_size=3,
stride=2,
padding=1),
nn_ext.Conv2dLayer(32, 64, kernel_size=3, padding=1))),
('block_1',
Xception.EntryFlowBlock(64, 128,
**opts['entry_flow_block_1'])),
('block_2',
Xception.EntryFlowBlock(128, 256,
**opts['entry_flow_block_2'])),
('block_3',
Xception.EntryFlowBlock(256, 728,
**opts['entry_flow_block_3'])),
]))
# [DeepLabV3+ specific] Split-entry flow sequence to extract
# low-level features needed by decoder (at output_stride // 4)
if output_stride == 16:
self.entry_flow = nn.ModuleList([entry_flow[0:2], entry_flow[2:4]])
elif output_stride == 8:
self.entry_flow = nn.ModuleList([entry_flow[0:1], entry_flow[1:4]])
else:
raise ValueError(
'Invalid output_stride; Supported values: {8, 16}')
self.middle_flow = nn.Sequential(
OrderedDict([
('block_0',
Xception.MiddleFlowBlock(**opts['middle_flow_block'])),
('block_1',
Xception.MiddleFlowBlock(**opts['middle_flow_block'])),
('block_2',
Xception.MiddleFlowBlock(**opts['middle_flow_block'])),
('block_3',
Xception.MiddleFlowBlock(**opts['middle_flow_block'])),
('block_4',
Xception.MiddleFlowBlock(**opts['middle_flow_block'])),
('block_5',
Xception.MiddleFlowBlock(**opts['middle_flow_block'])),
('block_6',
Xception.MiddleFlowBlock(**opts['middle_flow_block'])),
('block_7',
Xception.MiddleFlowBlock(**opts['middle_flow_block'])),
('block_8',
Xception.MiddleFlowBlock(**opts['middle_flow_block'])),
('block_9',
Xception.MiddleFlowBlock(**opts['middle_flow_block'])),
('block_10',
Xception.MiddleFlowBlock(**opts['middle_flow_block'])),
('block_11',
Xception.MiddleFlowBlock(**opts['middle_flow_block'])),
('block_12',
Xception.MiddleFlowBlock(**opts['middle_flow_block'])),
('block_13',
Xception.MiddleFlowBlock(**opts['middle_flow_block'])),
('block_14',
Xception.MiddleFlowBlock(**opts['middle_flow_block'])),
('block_15',
Xception.MiddleFlowBlock(**opts['middle_flow_block'])),
]))
self.exit_flow = nn.Sequential(
OrderedDict([
('block_0',
Xception.ExitFlowBlock(**opts['exit_flow_block_0'])),
('block_1',
nn.Sequential(
nn_ext.SeparableConv2dLayer(1024,
1536,
kernel_size=3,
**opts['exit_flow_block_1']),
nn_ext.SeparableConv2dLayer(1536,
1536,
kernel_size=3,
**opts['exit_flow_block_1']),
nn_ext.SeparableConv2dLayer(1536,
2048,
kernel_size=3,
**opts['exit_flow_block_1'])))
]))