File size: 11,772 Bytes
da2e2ac |
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 |
import torch
from mmcv.cnn.bricks.registry import (ATTENTION,
TRANSFORMER_LAYER,
POSITIONAL_ENCODING,
TRANSFORMER_LAYER_SEQUENCE)
from mmdet.models.utils.transformer import inverse_sigmoid
from mmcv.cnn.bricks.transformer import TransformerLayerSequence, BaseTransformerLayer
import copy
import warnings
@TRANSFORMER_LAYER_SEQUENCE.register_module()
class MapTRDecoder(TransformerLayerSequence):
"""Implements the decoder in DETR3D transformer.
Args:
return_intermediate (bool): Whether to return intermediate outputs.
coder_norm_cfg (dict): Config of last normalization layer. Default：
`LN`.
"""
def __init__(self, *args, return_intermediate=False, **kwargs):
super(MapTRDecoder, self).__init__(*args, **kwargs)
self.return_intermediate = return_intermediate
self.fp16_enabled = False
def forward(self,
query,
*args,
reference_points=None,
reg_branches=None,
key_padding_mask=None,
**kwargs):
"""Forward function for `Detr3DTransformerDecoder`.
Args:
query (Tensor): Input query with shape
`(num_query, bs, embed_dims)`.
reference_points (Tensor): The reference
points of offset. has shape
(bs, num_query, 4) when as_two_stage,
otherwise has shape ((bs, num_query, 2).
reg_branch: (obj:`nn.ModuleList`): Used for
refining the regression results. Only would
be passed when with_box_refine is True,
otherwise would be passed a `None`.
Returns:
Tensor: Results with shape [1, num_query, bs, embed_dims] when
return_intermediate is `False`, otherwise it has shape
[num_layers, num_query, bs, embed_dims].
"""
output = query
intermediate = []
intermediate_reference_points = []
for lid, layer in enumerate(self.layers):
reference_points_input = reference_points[..., :2].unsqueeze(
2) # BS NUM_QUERY NUM_LEVEL 2
output = layer(
output,
*args,
reference_points=reference_points_input,
key_padding_mask=key_padding_mask,
**kwargs)
output = output.permute(1, 0, 2)
if reg_branches is not None:
tmp = reg_branches[lid](output)
# assert reference_points.shape[-1] == 2
new_reference_points = torch.zeros_like(reference_points)
new_reference_points = tmp + inverse_sigmoid(reference_points)
# new_reference_points[..., 2:3] = tmp[
# ..., 4:5] + inverse_sigmoid(reference_points[..., 2:3])
new_reference_points = new_reference_points.sigmoid()
reference_points = new_reference_points.detach()
output = output.permute(1, 0, 2)
if self.return_intermediate:
intermediate.append(output)
intermediate_reference_points.append(reference_points)
if self.return_intermediate:
return torch.stack(intermediate), torch.stack(
intermediate_reference_points)
return output, reference_points
@TRANSFORMER_LAYER.register_module()
class DecoupledDetrTransformerDecoderLayer(BaseTransformerLayer):
"""Implements decoder layer in DETR transformer.
Args:
attn_cfgs (list[`mmcv.ConfigDict`] | list[dict] | dict )):
Configs for self_attention or cross_attention, the order
should be consistent with it in `operation_order`. If it is
a dict, it would be expand to the number of attention in
`operation_order`.
feedforward_channels (int): The hidden dimension for FFNs.
ffn_dropout (float): Probability of an element to be zeroed
in ffn. Default 0.0.
operation_order (tuple[str]): The execution order of operation
in transformer. Such as ('self_attn', 'norm', 'ffn', 'norm').
Default：None
act_cfg (dict): The activation config for FFNs. Default: `LN`
norm_cfg (dict): Config dict for normalization layer.
Default: `LN`.
ffn_num_fcs (int): The number of fully-connected layers in FFNs.
Default：2.
"""
def __init__(self,
attn_cfgs,
feedforward_channels,
num_vec=50,
num_pts_per_vec=20,
ffn_dropout=0.0,
operation_order=None,
act_cfg=dict(type='ReLU', inplace=True),
norm_cfg=dict(type='LN'),
ffn_num_fcs=2,
**kwargs):
super(DecoupledDetrTransformerDecoderLayer, self).__init__(
attn_cfgs=attn_cfgs,
feedforward_channels=feedforward_channels,
ffn_dropout=ffn_dropout,
operation_order=operation_order,
act_cfg=act_cfg,
norm_cfg=norm_cfg,
ffn_num_fcs=ffn_num_fcs,
**kwargs)
assert len(operation_order) == 8
assert set(operation_order) == set(
['self_attn', 'norm', 'cross_attn', 'ffn'])
self.num_vec = num_vec
self.num_pts_per_vec = num_pts_per_vec
def forward(self,
query,
key=None,
value=None,
query_pos=None,
key_pos=None,
attn_masks=None,
query_key_padding_mask=None,
key_padding_mask=None,
**kwargs):
"""Forward function for `TransformerDecoderLayer`.
**kwargs contains some specific arguments of attentions.
Args:
query (Tensor): The input query with shape
[num_queries, bs, embed_dims] if
self.batch_first is False, else
[bs, num_queries embed_dims].
key (Tensor): The key tensor with shape [num_keys, bs,
embed_dims] if self.batch_first is False, else
[bs, num_keys, embed_dims] .
value (Tensor): The value tensor with same shape as `key`.
query_pos (Tensor): The positional encoding for `query`.
Default: None.
key_pos (Tensor): The positional encoding for `key`.
Default: None.
attn_masks (List[Tensor] | None): 2D Tensor used in
calculation of corresponding attention. The length of
it should equal to the number of `attention` in
`operation_order`. Default: None.
query_key_padding_mask (Tensor): ByteTensor for `query`, with
shape [bs, num_queries]. Only used in `self_attn` layer.
Defaults to None.
key_padding_mask (Tensor): ByteTensor for `query`, with
shape [bs, num_keys]. Default: None.
Returns:
Tensor: forwarded results with shape [num_queries, bs, embed_dims].
"""
norm_index = 0
attn_index = 0
ffn_index = 0
identity = query
if attn_masks is None:
attn_masks = [None for _ in range(self.num_attn)]
elif isinstance(attn_masks, torch.Tensor):
attn_masks = [
copy.deepcopy(attn_masks) for _ in range(self.num_attn)
]
warnings.warn(f'Use same attn_mask in all attentions in '
f'{self.__class__.__name__} ')
else:
assert len(attn_masks) == self.num_attn, f'The length of ' \
f'attn_masks {len(attn_masks)} must be equal ' \
f'to the number of attention in ' \
f'operation_order {self.num_attn}'
#
num_vec = kwargs['num_vec']
num_pts_per_vec = kwargs['num_pts_per_vec']
for layer in self.operation_order:
if layer == 'self_attn':
# import ipdb;ipdb.set_trace()
if attn_index == 0:
n_pts, n_batch, n_dim = query.shape
query = query.view(num_vec, num_pts_per_vec,n_batch,n_dim).flatten(1,2)
query_pos = query_pos.view(num_vec, num_pts_per_vec,n_batch,n_dim).flatten(1,2)
temp_key = temp_value = query
query = self.attentions[attn_index](
query,
temp_key,
temp_value,
identity if self.pre_norm else None,
query_pos=query_pos,
key_pos=query_pos,
attn_mask=kwargs['self_attn_mask'],
key_padding_mask=query_key_padding_mask,
**kwargs)
# import ipdb;ipdb.set_trace()
query = query.view(num_vec, num_pts_per_vec, n_batch, n_dim).flatten(0,1)
query_pos = query_pos.view(num_vec, num_pts_per_vec, n_batch, n_dim).flatten(0,1)
attn_index += 1
identity = query
else:
# import ipdb;ipdb.set_trace()
n_pts, n_batch, n_dim = query.shape
query = query.view(num_vec, num_pts_per_vec,n_batch,n_dim).permute(1,0,2,3).contiguous().flatten(1,2)
query_pos = query_pos.view(num_vec, num_pts_per_vec,n_batch,n_dim).permute(1,0,2,3).contiguous().flatten(1,2)
temp_key = temp_value = query
query = self.attentions[attn_index](
query,
temp_key,
temp_value,
identity if self.pre_norm else None,
query_pos=query_pos,
key_pos=query_pos,
attn_mask=attn_masks[attn_index],
key_padding_mask=query_key_padding_mask,
**kwargs)
# import ipdb;ipdb.set_trace()
query = query.view(num_pts_per_vec, num_vec, n_batch, n_dim).permute(1,0,2,3).contiguous().flatten(0,1)
query_pos = query_pos.view(num_pts_per_vec, num_vec, n_batch, n_dim).permute(1,0,2,3).contiguous().flatten(0,1)
attn_index += 1
identity = query
elif layer == 'norm':
query = self.norms[norm_index](query)
norm_index += 1
elif layer == 'cross_attn':
query = self.attentions[attn_index](
query,
key,
value,
identity if self.pre_norm else None,
query_pos=query_pos,
key_pos=key_pos,
attn_mask=attn_masks[attn_index],
key_padding_mask=key_padding_mask,
**kwargs)
attn_index += 1
identity = query
elif layer == 'ffn':
query = self.ffns[ffn_index](
query, identity if self.pre_norm else None)
ffn_index += 1
return query
|