Spaces:
Running
Running
File size: 16,249 Bytes
b6396ac |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 |
# Copyright (c) Facebook, Inc. and its affiliates.
import logging
import numpy as np
from typing import Callable, Dict, List, Optional, Tuple, Union
import fvcore.nn.weight_init as weight_init
import torch
from torch import nn
from torch.nn import functional as F
from torch.nn.init import xavier_uniform_, constant_, uniform_, normal_
from torch.cuda.amp import autocast
from detectron2.config import configurable
from detectron2.layers import Conv2d, ShapeSpec, get_norm
from detectron2.modeling import SEM_SEG_HEADS_REGISTRY
from ..transformer_decoder.position_encoding import PositionEmbeddingSine
from .ops.modules import MSDeformAttn
import copy
def _get_clones(module, N):
return nn.ModuleList([copy.deepcopy(module) for i in range(N)])
def _get_activation_fn(activation):
"""Return an activation function given a string"""
if activation == "relu":
return F.relu
if activation == "gelu":
return F.gelu
if activation == "glu":
return F.glu
raise RuntimeError(f"activation should be relu/gelu, not {activation}.")
def build_pixel_decoder(cfg, input_shape):
"""
Build a pixel decoder from `cfg.MODEL.ONE_FORMER.PIXEL_DECODER_NAME`.
"""
name = cfg.MODEL.SEM_SEG_HEAD.PIXEL_DECODER_NAME
model = SEM_SEG_HEADS_REGISTRY.get(name)(cfg, input_shape)
forward_features = getattr(model, "forward_features", None)
if not callable(forward_features):
raise ValueError(
"Only SEM_SEG_HEADS with forward_features method can be used as pixel decoder. "
f"Please implement forward_features for {name} to only return mask features."
)
return model
# MSDeformAttn Transformer encoder in deformable detr
class MSDeformAttnTransformerEncoderOnly(nn.Module):
def __init__(self, d_model=256, nhead=8,
num_encoder_layers=6, dim_feedforward=1024, dropout=0.1,
activation="relu",
num_feature_levels=4, enc_n_points=4,
):
super().__init__()
self.d_model = d_model
self.nhead = nhead
encoder_layer = MSDeformAttnTransformerEncoderLayer(d_model, dim_feedforward,
dropout, activation,
num_feature_levels, nhead, enc_n_points)
self.encoder = MSDeformAttnTransformerEncoder(encoder_layer, num_encoder_layers)
self.level_embed = nn.Parameter(torch.Tensor(num_feature_levels, d_model))
self._reset_parameters()
def _reset_parameters(self):
for p in self.parameters():
if p.dim() > 1:
nn.init.xavier_uniform_(p)
for m in self.modules():
if isinstance(m, MSDeformAttn):
m._reset_parameters()
normal_(self.level_embed)
def get_valid_ratio(self, mask):
_, H, W = mask.shape
valid_H = torch.sum(~mask[:, :, 0], 1)
valid_W = torch.sum(~mask[:, 0, :], 1)
valid_ratio_h = valid_H.float() / H
valid_ratio_w = valid_W.float() / W
valid_ratio = torch.stack([valid_ratio_w, valid_ratio_h], -1)
return valid_ratio
def forward(self, srcs, pos_embeds):
masks = [torch.zeros((x.size(0), x.size(2), x.size(3)), device=x.device, dtype=torch.bool) for x in srcs]
# prepare input for encoder
src_flatten = []
mask_flatten = []
lvl_pos_embed_flatten = []
spatial_shapes = []
for lvl, (src, mask, pos_embed) in enumerate(zip(srcs, masks, pos_embeds)):
bs, c, h, w = src.shape
spatial_shape = (h, w)
spatial_shapes.append(spatial_shape)
src = src.flatten(2).transpose(1, 2)
mask = mask.flatten(1)
pos_embed = pos_embed.flatten(2).transpose(1, 2)
lvl_pos_embed = pos_embed + self.level_embed[lvl].view(1, 1, -1)
lvl_pos_embed_flatten.append(lvl_pos_embed)
src_flatten.append(src)
mask_flatten.append(mask)
src_flatten = torch.cat(src_flatten, 1)
mask_flatten = torch.cat(mask_flatten, 1)
lvl_pos_embed_flatten = torch.cat(lvl_pos_embed_flatten, 1)
spatial_shapes = torch.as_tensor(spatial_shapes, dtype=torch.long, device=src_flatten.device)
level_start_index = torch.cat((spatial_shapes.new_zeros((1, )), spatial_shapes.prod(1).cumsum(0)[:-1]))
valid_ratios = torch.stack([self.get_valid_ratio(m) for m in masks], 1)
# encoder
memory = self.encoder(src_flatten, spatial_shapes, level_start_index, valid_ratios, lvl_pos_embed_flatten, mask_flatten)
return memory, spatial_shapes, level_start_index
class MSDeformAttnTransformerEncoderLayer(nn.Module):
def __init__(self,
d_model=256, d_ffn=1024,
dropout=0.1, activation="relu",
n_levels=4, n_heads=8, n_points=4):
super().__init__()
# self attention
self.self_attn = MSDeformAttn(d_model, n_levels, n_heads, n_points)
self.dropout1 = nn.Dropout(dropout)
self.norm1 = nn.LayerNorm(d_model)
# ffn
self.linear1 = nn.Linear(d_model, d_ffn)
self.activation = _get_activation_fn(activation)
self.dropout2 = nn.Dropout(dropout)
self.linear2 = nn.Linear(d_ffn, d_model)
self.dropout3 = nn.Dropout(dropout)
self.norm2 = nn.LayerNorm(d_model)
@staticmethod
def with_pos_embed(tensor, pos):
return tensor if pos is None else tensor + pos
def forward_ffn(self, src):
src2 = self.linear2(self.dropout2(self.activation(self.linear1(src))))
src = src + self.dropout3(src2)
src = self.norm2(src)
return src
def forward(self, src, pos, reference_points, spatial_shapes, level_start_index, padding_mask=None):
# self attention
src2 = self.self_attn(self.with_pos_embed(src, pos), reference_points, src, spatial_shapes, level_start_index, padding_mask)
src = src + self.dropout1(src2)
src = self.norm1(src)
# ffn
src = self.forward_ffn(src)
return src
class MSDeformAttnTransformerEncoder(nn.Module):
def __init__(self, encoder_layer, num_layers):
super().__init__()
self.layers = _get_clones(encoder_layer, num_layers)
self.num_layers = num_layers
@staticmethod
def get_reference_points(spatial_shapes, valid_ratios, device):
reference_points_list = []
for lvl, (H_, W_) in enumerate(spatial_shapes):
ref_y, ref_x = torch.meshgrid(torch.linspace(0.5, H_ - 0.5, H_, dtype=torch.float32, device=device),
torch.linspace(0.5, W_ - 0.5, W_, dtype=torch.float32, device=device))
ref_y = ref_y.reshape(-1)[None] / (valid_ratios[:, None, lvl, 1] * H_)
ref_x = ref_x.reshape(-1)[None] / (valid_ratios[:, None, lvl, 0] * W_)
ref = torch.stack((ref_x, ref_y), -1)
reference_points_list.append(ref)
reference_points = torch.cat(reference_points_list, 1)
reference_points = reference_points[:, :, None] * valid_ratios[:, None]
return reference_points
def forward(self, src, spatial_shapes, level_start_index, valid_ratios, pos=None, padding_mask=None):
output = src
reference_points = self.get_reference_points(spatial_shapes, valid_ratios, device=src.device)
for _, layer in enumerate(self.layers):
output = layer(output, pos, reference_points, spatial_shapes, level_start_index, padding_mask)
return output
@SEM_SEG_HEADS_REGISTRY.register()
class MSDeformAttnPixelDecoder(nn.Module):
@configurable
def __init__(
self,
input_shape: Dict[str, ShapeSpec],
*,
transformer_dropout: float,
transformer_nheads: int,
transformer_dim_feedforward: int,
transformer_enc_layers: int,
conv_dim: int,
mask_dim: int,
norm: Optional[Union[str, Callable]] = None,
# deformable transformer encoder args
transformer_in_features: List[str],
common_stride: int,
):
"""
NOTE: this interface is experimental.
Args:
input_shape: shapes (channels and stride) of the input features
transformer_dropout: dropout probability in transformer
transformer_nheads: number of heads in transformer
transformer_dim_feedforward: dimension of feedforward network
transformer_enc_layers: number of transformer encoder layers
conv_dims: number of output channels for the intermediate conv layers.
mask_dim: number of output channels for the final conv layer.
norm (str or callable): normalization for all conv layers
"""
super().__init__()
transformer_input_shape = {
k: v for k, v in input_shape.items() if k in transformer_in_features
}
# this is the input shape of pixel decoder
input_shape = sorted(input_shape.items(), key=lambda x: x[1].stride)
self.in_features = [k for k, v in input_shape] # starting from "res2" to "res5"
self.feature_strides = [v.stride for k, v in input_shape]
self.feature_channels = [v.channels for k, v in input_shape]
# this is the input shape of transformer encoder (could use less features than pixel decoder
transformer_input_shape = sorted(transformer_input_shape.items(), key=lambda x: x[1].stride)
self.transformer_in_features = [k for k, v in transformer_input_shape] # starting from "res2" to "res5"
transformer_in_channels = [v.channels for k, v in transformer_input_shape]
self.transformer_feature_strides = [v.stride for k, v in transformer_input_shape] # to decide extra FPN layers
self.transformer_num_feature_levels = len(self.transformer_in_features)
if self.transformer_num_feature_levels > 1:
input_proj_list = []
# from low resolution to high resolution (res5 -> res2)
for in_channels in transformer_in_channels[::-1]:
input_proj_list.append(nn.Sequential(
nn.Conv2d(in_channels, conv_dim, kernel_size=1),
nn.GroupNorm(32, conv_dim),
))
self.input_proj = nn.ModuleList(input_proj_list)
else:
self.input_proj = nn.ModuleList([
nn.Sequential(
nn.Conv2d(transformer_in_channels[-1], conv_dim, kernel_size=1),
nn.GroupNorm(32, conv_dim),
)])
for proj in self.input_proj:
nn.init.xavier_uniform_(proj[0].weight, gain=1)
nn.init.constant_(proj[0].bias, 0)
self.transformer = MSDeformAttnTransformerEncoderOnly(
d_model=conv_dim,
dropout=transformer_dropout,
nhead=transformer_nheads,
dim_feedforward=transformer_dim_feedforward,
num_encoder_layers=transformer_enc_layers,
num_feature_levels=self.transformer_num_feature_levels,
)
N_steps = conv_dim // 2
self.pe_layer = PositionEmbeddingSine(N_steps, normalize=True)
self.mask_dim = mask_dim
# use 1x1 conv instead
self.mask_features = Conv2d(
conv_dim,
mask_dim,
kernel_size=1,
stride=1,
padding=0,
)
weight_init.c2_xavier_fill(self.mask_features)
self.maskformer_num_feature_levels = 3 # always use 3 scales
self.common_stride = common_stride
# extra fpn levels
stride = min(self.transformer_feature_strides)
self.num_fpn_levels = int(np.log2(stride) - np.log2(self.common_stride))
lateral_convs = []
output_convs = []
use_bias = norm == ""
for idx, in_channels in enumerate(self.feature_channels[:self.num_fpn_levels]):
lateral_norm = get_norm(norm, conv_dim)
output_norm = get_norm(norm, conv_dim)
lateral_conv = Conv2d(
in_channels, conv_dim, kernel_size=1, bias=use_bias, norm=lateral_norm
)
output_conv = Conv2d(
conv_dim,
conv_dim,
kernel_size=3,
stride=1,
padding=1,
bias=use_bias,
norm=output_norm,
activation=F.relu,
)
weight_init.c2_xavier_fill(lateral_conv)
weight_init.c2_xavier_fill(output_conv)
self.add_module("adapter_{}".format(idx + 1), lateral_conv)
self.add_module("layer_{}".format(idx + 1), output_conv)
lateral_convs.append(lateral_conv)
output_convs.append(output_conv)
# Place convs into top-down order (from low to high resolution)
# to make the top-down computation in forward clearer.
self.lateral_convs = lateral_convs[::-1]
self.output_convs = output_convs[::-1]
@classmethod
def from_config(cls, cfg, input_shape: Dict[str, ShapeSpec]):
ret = {}
ret["input_shape"] = {
k: v for k, v in input_shape.items() if k in cfg.MODEL.SEM_SEG_HEAD.IN_FEATURES
}
ret["conv_dim"] = cfg.MODEL.SEM_SEG_HEAD.CONVS_DIM
ret["mask_dim"] = cfg.MODEL.SEM_SEG_HEAD.MASK_DIM
ret["norm"] = cfg.MODEL.SEM_SEG_HEAD.NORM
ret["transformer_dropout"] = cfg.MODEL.MASK_FORMER.DROPOUT
ret["transformer_nheads"] = cfg.MODEL.MASK_FORMER.NHEADS
# ret["transformer_dim_feedforward"] = cfg.MODEL.MASK_FORMER.DIM_FEEDFORWARD
ret["transformer_dim_feedforward"] = 1024 # use 1024 for deformable transformer encoder
ret[
"transformer_enc_layers"
] = cfg.MODEL.SEM_SEG_HEAD.TRANSFORMER_ENC_LAYERS # a separate config
ret["transformer_in_features"] = cfg.MODEL.SEM_SEG_HEAD.DEFORMABLE_TRANSFORMER_ENCODER_IN_FEATURES
ret["common_stride"] = cfg.MODEL.SEM_SEG_HEAD.COMMON_STRIDE
return ret
@autocast(enabled=False)
def forward_features(self, features):
srcs = []
pos = []
# Reverse feature maps into top-down order (from low to high resolution)
for idx, f in enumerate(self.transformer_in_features[::-1]):
x = features[f].float() # deformable detr does not support half precision
srcs.append(self.input_proj[idx](x))
pos.append(self.pe_layer(x))
y, spatial_shapes, level_start_index = self.transformer(srcs, pos)
bs = y.shape[0]
split_size_or_sections = [None] * self.transformer_num_feature_levels
for i in range(self.transformer_num_feature_levels):
if i < self.transformer_num_feature_levels - 1:
split_size_or_sections[i] = level_start_index[i + 1] - level_start_index[i]
else:
split_size_or_sections[i] = y.shape[1] - level_start_index[i]
y = torch.split(y, split_size_or_sections, dim=1)
out = []
multi_scale_features = []
num_cur_levels = 0
for i, z in enumerate(y):
out.append(z.transpose(1, 2).view(bs, -1, spatial_shapes[i][0], spatial_shapes[i][1]))
# append `out` with extra FPN levels
# Reverse feature maps into top-down order (from low to high resolution)
for idx, f in enumerate(self.in_features[:self.num_fpn_levels][::-1]):
x = features[f].float()
lateral_conv = self.lateral_convs[idx]
output_conv = self.output_convs[idx]
cur_fpn = lateral_conv(x)
# Following FPN implementation, we use nearest upsampling here
y = cur_fpn + F.interpolate(out[-1], size=cur_fpn.shape[-2:], mode="bilinear", align_corners=False)
y = output_conv(y)
out.append(y)
for o in out:
if num_cur_levels < self.maskformer_num_feature_levels:
multi_scale_features.append(o)
num_cur_levels += 1
return self.mask_features(out[-1]), out[0], multi_scale_features
|