File size: 6,221 Bytes
03da825 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 |
from abc import ABCMeta, abstractmethod
import torch
import torch.nn as nn
from mmcv.cnn import normal_init
from mmcv.runner import auto_fp16
from additional_modules.segformer.ops import resize
class BaseDecodeHead(nn.Module, metaclass=ABCMeta):
"""Base class for BaseDecodeHead.
Args:
in_channels (int|Sequence[int]): Input channels.
channels (int): Channels after modules, before conv_seg.
num_classes (int): Number of classes.
dropout_ratio (float): Ratio of dropout layer. Default: 0.1.
conv_cfg (dict|None): Config of conv layers. Default: None.
norm_cfg (dict|None): Config of norm layers. Default: None.
act_cfg (dict): Config of activation layers.
Default: dict(type='ReLU')
in_index (int|Sequence[int]): Input feature index. Default: -1
input_transform (str|None): Transformation type of input features.
Options: 'resize_concat', 'multiple_select', None.
'resize_concat': Multiple feature maps will be resize to the
same size as first one and than concat together.
Usually used in FCN head of HRNet.
'multiple_select': Multiple feature maps will be bundle into
a list and passed into decode head.
None: Only one select feature map is allowed.
Default: None.
ignore_index (int | None): The label index to be ignored. When using
masked BCE loss, ignore_index should be set to None. Default: 255
sampler (dict|None): The config of segmentation map sampler.
Default: None.
align_corners (bool): align_corners argument of F.interpolate.
Default: False.
"""
def __init__(self,
in_channels,
channels,
*,
num_classes,
dropout_ratio=0.1,
conv_cfg=None,
norm_cfg=None,
act_cfg=dict(type='ReLU'),
in_index=-1,
input_transform=None,
decoder_params=None,
ignore_index=255,
align_corners=False):
super(BaseDecodeHead, self).__init__()
self._init_inputs(in_channels, in_index, input_transform)
self.channels = channels
self.num_classes = num_classes
self.dropout_ratio = dropout_ratio
self.conv_cfg = conv_cfg
self.norm_cfg = norm_cfg
self.act_cfg = act_cfg
self.in_index = in_index
self.ignore_index = ignore_index
self.align_corners = align_corners
self.sampler = None
self.conv_seg = nn.Conv2d(channels, num_classes, kernel_size=1)
if dropout_ratio > 0:
self.dropout = nn.Dropout2d(dropout_ratio)
else:
self.dropout = None
self.fp16_enabled = False
def extra_repr(self):
"""Extra repr."""
s = f'input_transform={self.input_transform}, ' \
f'ignore_index={self.ignore_index}, ' \
f'align_corners={self.align_corners}'
return s
def _init_inputs(self, in_channels, in_index, input_transform):
"""Check and initialize input transforms.
The in_channels, in_index and input_transform must match.
Specifically, when input_transform is None, only single feature map
will be selected. So in_channels and in_index must be of type int.
When input_transform
Args:
in_channels (int|Sequence[int]): Input channels.
in_index (int|Sequence[int]): Input feature index.
input_transform (str|None): Transformation type of input features.
Options: 'resize_concat', 'multiple_select', None.
'resize_concat': Multiple feature maps will be resize to the
same size as first one and than concat together.
Usually used in FCN head of HRNet.
'multiple_select': Multiple feature maps will be bundle into
a list and passed into decode head.
None: Only one select feature map is allowed.
"""
if input_transform is not None:
assert input_transform in ['resize_concat', 'multiple_select']
self.input_transform = input_transform
self.in_index = in_index
if input_transform is not None:
assert isinstance(in_channels, (list, tuple))
assert isinstance(in_index, (list, tuple))
assert len(in_channels) == len(in_index)
if input_transform == 'resize_concat':
self.in_channels = sum(in_channels)
else:
self.in_channels = in_channels
else:
assert isinstance(in_channels, int)
assert isinstance(in_index, int)
self.in_channels = in_channels
def init_weights(self):
"""Initialize weights of classification layer."""
normal_init(self.conv_seg, mean=0, std=0.01)
def _transform_inputs(self, inputs):
"""Transform inputs for decoder.
Args:
inputs (list[Tensor]): List of multi-level img features.
Returns:
Tensor: The transformed inputs
"""
if self.input_transform == 'resize_concat':
inputs = [inputs[i] for i in self.in_index]
upsampled_inputs = [
resize(
input=x,
size=inputs[0].shape[2:],
mode='bilinear',
align_corners=self.align_corners) for x in inputs
]
inputs = torch.cat(upsampled_inputs, dim=1)
elif self.input_transform == 'multiple_select':
inputs = [inputs[i] for i in self.in_index]
else:
inputs = inputs[self.in_index]
return inputs
@auto_fp16()
@abstractmethod
def forward(self, inputs):
"""Placeholder of forward function."""
pass
def cls_seg(self, feat):
"""Classify each pixel."""
if self.dropout is not None:
feat = self.dropout(feat)
output = self.conv_seg(feat)
return output
|