Spaces:
Runtime error
Runtime error
File size: 11,450 Bytes
51f6859 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 |
# Copyright (c) OpenMMLab. All rights reserved.
import torch
import torch.nn as nn
import torch.nn.functional as F
from mmcv.cnn import (PLUGIN_LAYERS, Conv2d, ConvModule, caffe2_xavier_init,
normal_init, xavier_init)
from mmcv.cnn.bricks.transformer import (build_positional_encoding,
build_transformer_layer_sequence)
from mmcv.runner import BaseModule, ModuleList
from mmdet.core.anchor import MlvlPointGenerator
from mmdet.models.utils.transformer import MultiScaleDeformableAttention
@PLUGIN_LAYERS.register_module()
class MSDeformAttnPixelDecoder(BaseModule):
"""Pixel decoder with multi-scale deformable attention.
Args:
in_channels (list[int] | tuple[int]): Number of channels in the
input feature maps.
strides (list[int] | tuple[int]): Output strides of feature from
backbone.
feat_channels (int): Number of channels for feature.
out_channels (int): Number of channels for output.
num_outs (int): Number of output scales.
norm_cfg (:obj:`mmcv.ConfigDict` | dict): Config for normalization.
Defaults to dict(type='GN', num_groups=32).
act_cfg (:obj:`mmcv.ConfigDict` | dict): Config for activation.
Defaults to dict(type='ReLU').
encoder (:obj:`mmcv.ConfigDict` | dict): Config for transformer
encoder. Defaults to `DetrTransformerEncoder`.
positional_encoding (:obj:`mmcv.ConfigDict` | dict): Config for
transformer encoder position encoding. Defaults to
dict(type='SinePositionalEncoding', num_feats=128,
normalize=True).
init_cfg (:obj:`mmcv.ConfigDict` | dict): Initialization config dict.
"""
def __init__(self,
in_channels=[256, 512, 1024, 2048],
strides=[4, 8, 16, 32],
feat_channels=256,
out_channels=256,
num_outs=3,
norm_cfg=dict(type='GN', num_groups=32),
act_cfg=dict(type='ReLU'),
encoder=dict(
type='DetrTransformerEncoder',
num_layers=6,
transformerlayers=dict(
type='BaseTransformerLayer',
attn_cfgs=dict(
type='MultiScaleDeformableAttention',
embed_dims=256,
num_heads=8,
num_levels=3,
num_points=4,
im2col_step=64,
dropout=0.0,
batch_first=False,
norm_cfg=None,
init_cfg=None),
feedforward_channels=1024,
ffn_dropout=0.0,
operation_order=('self_attn', 'norm', 'ffn', 'norm')),
init_cfg=None),
positional_encoding=dict(
type='SinePositionalEncoding',
num_feats=128,
normalize=True),
init_cfg=None):
super().__init__(init_cfg=init_cfg)
self.strides = strides
self.num_input_levels = len(in_channels)
self.num_encoder_levels = \
encoder.transformerlayers.attn_cfgs.num_levels
assert self.num_encoder_levels >= 1, \
'num_levels in attn_cfgs must be at least one'
input_conv_list = []
# from top to down (low to high resolution)
for i in range(self.num_input_levels - 1,
self.num_input_levels - self.num_encoder_levels - 1,
-1):
input_conv = ConvModule(
in_channels[i],
feat_channels,
kernel_size=1,
norm_cfg=norm_cfg,
act_cfg=None,
bias=True)
input_conv_list.append(input_conv)
self.input_convs = ModuleList(input_conv_list)
self.encoder = build_transformer_layer_sequence(encoder)
self.postional_encoding = build_positional_encoding(
positional_encoding)
# high resolution to low resolution
self.level_encoding = nn.Embedding(self.num_encoder_levels,
feat_channels)
# fpn-like structure
self.lateral_convs = ModuleList()
self.output_convs = ModuleList()
self.use_bias = norm_cfg is None
# from top to down (low to high resolution)
# fpn for the rest features that didn't pass in encoder
for i in range(self.num_input_levels - self.num_encoder_levels - 1, -1,
-1):
lateral_conv = ConvModule(
in_channels[i],
feat_channels,
kernel_size=1,
bias=self.use_bias,
norm_cfg=norm_cfg,
act_cfg=None)
output_conv = ConvModule(
feat_channels,
feat_channels,
kernel_size=3,
stride=1,
padding=1,
bias=self.use_bias,
norm_cfg=norm_cfg,
act_cfg=act_cfg)
self.lateral_convs.append(lateral_conv)
self.output_convs.append(output_conv)
self.mask_feature = Conv2d(
feat_channels, out_channels, kernel_size=1, stride=1, padding=0)
self.num_outs = num_outs
self.point_generator = MlvlPointGenerator(strides)
def init_weights(self):
"""Initialize weights."""
for i in range(0, self.num_encoder_levels):
xavier_init(
self.input_convs[i].conv,
gain=1,
bias=0,
distribution='uniform')
for i in range(0, self.num_input_levels - self.num_encoder_levels):
caffe2_xavier_init(self.lateral_convs[i].conv, bias=0)
caffe2_xavier_init(self.output_convs[i].conv, bias=0)
caffe2_xavier_init(self.mask_feature, bias=0)
normal_init(self.level_encoding, mean=0, std=1)
for p in self.encoder.parameters():
if p.dim() > 1:
nn.init.xavier_normal_(p)
# init_weights defined in MultiScaleDeformableAttention
for layer in self.encoder.layers:
for attn in layer.attentions:
if isinstance(attn, MultiScaleDeformableAttention):
attn.init_weights()
def forward(self, feats):
"""
Args:
feats (list[Tensor]): Feature maps of each level. Each has
shape of (batch_size, c, h, w).
Returns:
tuple: A tuple containing the following:
- mask_feature (Tensor): shape (batch_size, c, h, w).
- multi_scale_features (list[Tensor]): Multi scale \
features, each in shape (batch_size, c, h, w).
"""
# generate padding mask for each level, for each image
batch_size = feats[0].shape[0]
encoder_input_list = []
padding_mask_list = []
level_positional_encoding_list = []
spatial_shapes = []
reference_points_list = []
for i in range(self.num_encoder_levels):
level_idx = self.num_input_levels - i - 1
feat = feats[level_idx]
feat_projected = self.input_convs[i](feat)
h, w = feat.shape[-2:]
# no padding
padding_mask_resized = feat.new_zeros(
(batch_size, ) + feat.shape[-2:], dtype=torch.bool)
pos_embed = self.postional_encoding(padding_mask_resized)
level_embed = self.level_encoding.weight[i]
level_pos_embed = level_embed.view(1, -1, 1, 1) + pos_embed
# (h_i * w_i, 2)
reference_points = self.point_generator.single_level_grid_priors(
feat.shape[-2:], level_idx, device=feat.device)
# normalize
factor = feat.new_tensor([[w, h]]) * self.strides[level_idx]
reference_points = reference_points / factor
# shape (batch_size, c, h_i, w_i) -> (h_i * w_i, batch_size, c)
feat_projected = feat_projected.flatten(2).permute(2, 0, 1)
level_pos_embed = level_pos_embed.flatten(2).permute(2, 0, 1)
padding_mask_resized = padding_mask_resized.flatten(1)
encoder_input_list.append(feat_projected)
padding_mask_list.append(padding_mask_resized)
level_positional_encoding_list.append(level_pos_embed)
spatial_shapes.append(feat.shape[-2:])
reference_points_list.append(reference_points)
# shape (batch_size, total_num_query),
# total_num_query=sum([., h_i * w_i,.])
padding_masks = torch.cat(padding_mask_list, dim=1)
# shape (total_num_query, batch_size, c)
encoder_inputs = torch.cat(encoder_input_list, dim=0)
level_positional_encodings = torch.cat(
level_positional_encoding_list, dim=0)
device = encoder_inputs.device
# shape (num_encoder_levels, 2), from low
# resolution to high resolution
spatial_shapes = torch.as_tensor(
spatial_shapes, dtype=torch.long, device=device)
# shape (0, h_0*w_0, h_0*w_0+h_1*w_1, ...)
level_start_index = torch.cat((spatial_shapes.new_zeros(
(1, )), spatial_shapes.prod(1).cumsum(0)[:-1]))
reference_points = torch.cat(reference_points_list, dim=0)
reference_points = reference_points[None, :, None].repeat(
batch_size, 1, self.num_encoder_levels, 1)
valid_radios = reference_points.new_ones(
(batch_size, self.num_encoder_levels, 2))
# shape (num_total_query, batch_size, c)
memory = self.encoder(
query=encoder_inputs,
key=None,
value=None,
query_pos=level_positional_encodings,
key_pos=None,
attn_masks=None,
key_padding_mask=None,
query_key_padding_mask=padding_masks,
spatial_shapes=spatial_shapes,
reference_points=reference_points,
level_start_index=level_start_index,
valid_radios=valid_radios)
# (num_total_query, batch_size, c) -> (batch_size, c, num_total_query)
memory = memory.permute(1, 2, 0)
# from low resolution to high resolution
num_query_per_level = [e[0] * e[1] for e in spatial_shapes]
outs = torch.split(memory, num_query_per_level, dim=-1)
outs = [
x.reshape(batch_size, -1, spatial_shapes[i][0],
spatial_shapes[i][1]) for i, x in enumerate(outs)
]
for i in range(self.num_input_levels - self.num_encoder_levels - 1, -1,
-1):
x = feats[i]
cur_feat = self.lateral_convs[i](x)
y = cur_feat + F.interpolate(
outs[-1],
size=cur_feat.shape[-2:],
mode='bilinear',
align_corners=False)
y = self.output_convs[i](y)
outs.append(y)
multi_scale_features = outs[:self.num_outs]
mask_feature = self.mask_feature(outs[-1])
return mask_feature, multi_scale_features
|