deformable_transformer.py 19 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517
  1. # Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
  2. #
  3. # Licensed under the Apache License, Version 2.0 (the "License");
  4. # you may not use this file except in compliance with the License.
  5. # You may obtain a copy of the License at
  6. #
  7. # http://www.apache.org/licenses/LICENSE-2.0
  8. #
  9. # Unless required by applicable law or agreed to in writing, software
  10. # distributed under the License is distributed on an "AS IS" BASIS,
  11. # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
  12. # See the License for the specific language governing permissions and
  13. # limitations under the License.
  14. #
  15. # Modified from Deformable-DETR (https://github.com/fundamentalvision/Deformable-DETR)
  16. # Copyright (c) 2020 SenseTime. All Rights Reserved.
  17. from __future__ import absolute_import
  18. from __future__ import division
  19. from __future__ import print_function
  20. import math
  21. import paddle
  22. import paddle.nn as nn
  23. import paddle.nn.functional as F
  24. from paddle import ParamAttr
  25. from paddlex.ppdet.core.workspace import register
  26. from ..layers import MultiHeadAttention
  27. from .position_encoding import PositionEmbedding
  28. from .utils import _get_clones, deformable_attention_core_func
  29. from ..initializer import linear_init_, constant_, xavier_uniform_, normal_
  30. __all__ = ['DeformableTransformer']
  31. class MSDeformableAttention(nn.Layer):
  32. def __init__(self,
  33. embed_dim=256,
  34. num_heads=8,
  35. num_levels=4,
  36. num_points=4,
  37. lr_mult=0.1):
  38. """
  39. Multi-Scale Deformable Attention Module
  40. """
  41. super(MSDeformableAttention, self).__init__()
  42. self.embed_dim = embed_dim
  43. self.num_heads = num_heads
  44. self.num_levels = num_levels
  45. self.num_points = num_points
  46. self.total_points = num_heads * num_levels * num_points
  47. self.head_dim = embed_dim // num_heads
  48. assert self.head_dim * num_heads == self.embed_dim, "embed_dim must be divisible by num_heads"
  49. self.sampling_offsets = nn.Linear(
  50. embed_dim,
  51. self.total_points * 2,
  52. weight_attr=ParamAttr(learning_rate=lr_mult),
  53. bias_attr=ParamAttr(learning_rate=lr_mult))
  54. self.attention_weights = nn.Linear(embed_dim, self.total_points)
  55. self.value_proj = nn.Linear(embed_dim, embed_dim)
  56. self.output_proj = nn.Linear(embed_dim, embed_dim)
  57. self._reset_parameters()
  58. def _reset_parameters(self):
  59. # sampling_offsets
  60. constant_(self.sampling_offsets.weight)
  61. thetas = paddle.arange(
  62. self.num_heads,
  63. dtype=paddle.float32) * (2.0 * math.pi / self.num_heads)
  64. grid_init = paddle.stack([thetas.cos(), thetas.sin()], -1)
  65. grid_init = grid_init / grid_init.abs().max(-1, keepdim=True)
  66. grid_init = grid_init.reshape([self.num_heads, 1, 1, 2]).tile(
  67. [1, self.num_levels, self.num_points, 1])
  68. scaling = paddle.arange(
  69. 1, self.num_points + 1,
  70. dtype=paddle.float32).reshape([1, 1, -1, 1])
  71. grid_init *= scaling
  72. self.sampling_offsets.bias.set_value(grid_init.flatten())
  73. # attention_weights
  74. constant_(self.attention_weights.weight)
  75. constant_(self.attention_weights.bias)
  76. # proj
  77. xavier_uniform_(self.value_proj.weight)
  78. constant_(self.value_proj.bias)
  79. xavier_uniform_(self.output_proj.weight)
  80. constant_(self.output_proj.bias)
  81. def forward(self,
  82. query,
  83. reference_points,
  84. value,
  85. value_spatial_shapes,
  86. value_mask=None):
  87. """
  88. Args:
  89. query (Tensor): [bs, query_length, C]
  90. reference_points (Tensor): [bs, query_length, n_levels, 2], range in [0, 1], top-left (0,0),
  91. bottom-right (1, 1), including padding area
  92. value (Tensor): [bs, value_length, C]
  93. value_spatial_shapes (Tensor): [n_levels, 2], [(H_0, W_0), (H_1, W_1), ..., (H_{L-1}, W_{L-1})]
  94. value_mask (Tensor): [bs, value_length], True for non-padding elements, False for padding elements
  95. Returns:
  96. output (Tensor): [bs, Length_{query}, C]
  97. """
  98. bs, Len_q = query.shape[:2]
  99. Len_v = value.shape[1]
  100. assert int(value_spatial_shapes.prod(1).sum()) == Len_v
  101. value = self.value_proj(value)
  102. if value_mask is not None:
  103. value_mask = value_mask.astype(value.dtype).unsqueeze(-1)
  104. value *= value_mask
  105. value = value.reshape([bs, Len_v, self.num_heads, self.head_dim])
  106. sampling_offsets = self.sampling_offsets(query).reshape(
  107. [bs, Len_q, self.num_heads, self.num_levels, self.num_points, 2])
  108. attention_weights = self.attention_weights(query).reshape(
  109. [bs, Len_q, self.num_heads, self.num_levels * self.num_points])
  110. attention_weights = F.softmax(attention_weights, -1).reshape(
  111. [bs, Len_q, self.num_heads, self.num_levels, self.num_points])
  112. offset_normalizer = value_spatial_shapes.flip([1]).reshape(
  113. [1, 1, 1, self.num_levels, 1, 2])
  114. sampling_locations = reference_points.reshape([
  115. bs, Len_q, 1, self.num_levels, 1, 2
  116. ]) + sampling_offsets / offset_normalizer
  117. output = deformable_attention_core_func(
  118. value, value_spatial_shapes, sampling_locations, attention_weights)
  119. output = self.output_proj(output)
  120. return output
  121. class DeformableTransformerEncoderLayer(nn.Layer):
  122. def __init__(self,
  123. d_model=256,
  124. n_head=8,
  125. dim_feedforward=1024,
  126. dropout=0.1,
  127. activation="relu",
  128. n_levels=4,
  129. n_points=4,
  130. weight_attr=None,
  131. bias_attr=None):
  132. super(DeformableTransformerEncoderLayer, self).__init__()
  133. # self attention
  134. self.self_attn = MSDeformableAttention(d_model, n_head, n_levels,
  135. n_points)
  136. self.dropout1 = nn.Dropout(dropout)
  137. self.norm1 = nn.LayerNorm(d_model)
  138. # ffn
  139. self.linear1 = nn.Linear(d_model, dim_feedforward, weight_attr,
  140. bias_attr)
  141. self.activation = getattr(F, activation)
  142. self.dropout2 = nn.Dropout(dropout)
  143. self.linear2 = nn.Linear(dim_feedforward, d_model, weight_attr,
  144. bias_attr)
  145. self.dropout3 = nn.Dropout(dropout)
  146. self.norm2 = nn.LayerNorm(d_model)
  147. self._reset_parameters()
  148. def _reset_parameters(self):
  149. linear_init_(self.linear1)
  150. linear_init_(self.linear2)
  151. xavier_uniform_(self.linear1.weight)
  152. xavier_uniform_(self.linear2.weight)
  153. def with_pos_embed(self, tensor, pos):
  154. return tensor if pos is None else tensor + pos
  155. def forward_ffn(self, src):
  156. src2 = self.linear2(self.dropout2(self.activation(self.linear1(src))))
  157. src = src + self.dropout3(src2)
  158. src = self.norm2(src)
  159. return src
  160. def forward(self,
  161. src,
  162. reference_points,
  163. spatial_shapes,
  164. src_mask=None,
  165. pos_embed=None):
  166. # self attention
  167. src2 = self.self_attn(
  168. self.with_pos_embed(src, pos_embed), reference_points, src,
  169. spatial_shapes, src_mask)
  170. src = src + self.dropout1(src2)
  171. src = self.norm1(src)
  172. # ffn
  173. src = self.forward_ffn(src)
  174. return src
  175. class DeformableTransformerEncoder(nn.Layer):
  176. def __init__(self, encoder_layer, num_layers):
  177. super(DeformableTransformerEncoder, self).__init__()
  178. self.layers = _get_clones(encoder_layer, num_layers)
  179. self.num_layers = num_layers
  180. @staticmethod
  181. def get_reference_points(spatial_shapes, valid_ratios):
  182. valid_ratios = valid_ratios.unsqueeze(1)
  183. reference_points = []
  184. for i, (H, W) in enumerate(spatial_shapes.tolist()):
  185. ref_y, ref_x = paddle.meshgrid(
  186. paddle.linspace(0.5, H - 0.5, H),
  187. paddle.linspace(0.5, W - 0.5, W))
  188. ref_y = ref_y.flatten().unsqueeze(0) / (valid_ratios[:, :, i, 1] *
  189. H)
  190. ref_x = ref_x.flatten().unsqueeze(0) / (valid_ratios[:, :, i, 0] *
  191. W)
  192. reference_points.append(paddle.stack((ref_x, ref_y), axis=-1))
  193. reference_points = paddle.concat(reference_points, 1).unsqueeze(2)
  194. reference_points = reference_points * valid_ratios
  195. return reference_points
  196. def forward(self,
  197. src,
  198. spatial_shapes,
  199. src_mask=None,
  200. pos_embed=None,
  201. valid_ratios=None):
  202. output = src
  203. if valid_ratios is None:
  204. valid_ratios = paddle.ones(
  205. [src.shape[0], spatial_shapes.shape[0], 2])
  206. reference_points = self.get_reference_points(spatial_shapes,
  207. valid_ratios)
  208. for layer in self.layers:
  209. output = layer(output, reference_points, spatial_shapes, src_mask,
  210. pos_embed)
  211. return output
  212. class DeformableTransformerDecoderLayer(nn.Layer):
  213. def __init__(self,
  214. d_model=256,
  215. n_head=8,
  216. dim_feedforward=1024,
  217. dropout=0.1,
  218. activation="relu",
  219. n_levels=4,
  220. n_points=4,
  221. weight_attr=None,
  222. bias_attr=None):
  223. super(DeformableTransformerDecoderLayer, self).__init__()
  224. # self attention
  225. self.self_attn = MultiHeadAttention(d_model, n_head, dropout=dropout)
  226. self.dropout1 = nn.Dropout(dropout)
  227. self.norm1 = nn.LayerNorm(d_model)
  228. # cross attention
  229. self.cross_attn = MSDeformableAttention(d_model, n_head, n_levels,
  230. n_points)
  231. self.dropout2 = nn.Dropout(dropout)
  232. self.norm2 = nn.LayerNorm(d_model)
  233. # ffn
  234. self.linear1 = nn.Linear(d_model, dim_feedforward, weight_attr,
  235. bias_attr)
  236. self.activation = getattr(F, activation)
  237. self.dropout3 = nn.Dropout(dropout)
  238. self.linear2 = nn.Linear(dim_feedforward, d_model, weight_attr,
  239. bias_attr)
  240. self.dropout4 = nn.Dropout(dropout)
  241. self.norm3 = nn.LayerNorm(d_model)
  242. self._reset_parameters()
  243. def _reset_parameters(self):
  244. linear_init_(self.linear1)
  245. linear_init_(self.linear2)
  246. xavier_uniform_(self.linear1.weight)
  247. xavier_uniform_(self.linear2.weight)
  248. def with_pos_embed(self, tensor, pos):
  249. return tensor if pos is None else tensor + pos
  250. def forward_ffn(self, tgt):
  251. tgt2 = self.linear2(self.dropout3(self.activation(self.linear1(tgt))))
  252. tgt = tgt + self.dropout4(tgt2)
  253. tgt = self.norm3(tgt)
  254. return tgt
  255. def forward(self,
  256. tgt,
  257. reference_points,
  258. memory,
  259. memory_spatial_shapes,
  260. memory_mask=None,
  261. query_pos_embed=None):
  262. # self attention
  263. q = k = self.with_pos_embed(tgt, query_pos_embed)
  264. tgt2 = self.self_attn(q, k, value=tgt)
  265. tgt = tgt + self.dropout1(tgt2)
  266. tgt = self.norm1(tgt)
  267. # cross attention
  268. tgt2 = self.cross_attn(
  269. self.with_pos_embed(tgt, query_pos_embed), reference_points,
  270. memory, memory_spatial_shapes, memory_mask)
  271. tgt = tgt + self.dropout2(tgt2)
  272. tgt = self.norm2(tgt)
  273. # ffn
  274. tgt = self.forward_ffn(tgt)
  275. return tgt
  276. class DeformableTransformerDecoder(nn.Layer):
  277. def __init__(self, decoder_layer, num_layers, return_intermediate=False):
  278. super(DeformableTransformerDecoder, self).__init__()
  279. self.layers = _get_clones(decoder_layer, num_layers)
  280. self.num_layers = num_layers
  281. self.return_intermediate = return_intermediate
  282. def forward(self,
  283. tgt,
  284. reference_points,
  285. memory,
  286. memory_spatial_shapes,
  287. memory_mask=None,
  288. query_pos_embed=None):
  289. output = tgt
  290. intermediate = []
  291. for lid, layer in enumerate(self.layers):
  292. output = layer(output, reference_points, memory,
  293. memory_spatial_shapes, memory_mask, query_pos_embed)
  294. if self.return_intermediate:
  295. intermediate.append(output)
  296. if self.return_intermediate:
  297. return paddle.stack(intermediate)
  298. return output.unsqueeze(0)
  299. @register
  300. class DeformableTransformer(nn.Layer):
  301. __shared__ = ['hidden_dim']
  302. def __init__(self,
  303. num_queries=300,
  304. position_embed_type='sine',
  305. return_intermediate_dec=True,
  306. backbone_num_channels=[512, 1024, 2048],
  307. num_feature_levels=4,
  308. num_encoder_points=4,
  309. num_decoder_points=4,
  310. hidden_dim=256,
  311. nhead=8,
  312. num_encoder_layers=6,
  313. num_decoder_layers=6,
  314. dim_feedforward=1024,
  315. dropout=0.1,
  316. activation="relu",
  317. lr_mult=0.1,
  318. weight_attr=None,
  319. bias_attr=None):
  320. super(DeformableTransformer, self).__init__()
  321. assert position_embed_type in ['sine', 'learned'], \
  322. f'ValueError: position_embed_type not supported {position_embed_type}!'
  323. assert len(backbone_num_channels) <= num_feature_levels
  324. self.hidden_dim = hidden_dim
  325. self.nhead = nhead
  326. self.num_feature_levels = num_feature_levels
  327. encoder_layer = DeformableTransformerEncoderLayer(
  328. hidden_dim, nhead, dim_feedforward, dropout, activation,
  329. num_feature_levels, num_encoder_points, weight_attr, bias_attr)
  330. self.encoder = DeformableTransformerEncoder(encoder_layer,
  331. num_encoder_layers)
  332. decoder_layer = DeformableTransformerDecoderLayer(
  333. hidden_dim, nhead, dim_feedforward, dropout, activation,
  334. num_feature_levels, num_decoder_points, weight_attr, bias_attr)
  335. self.decoder = DeformableTransformerDecoder(
  336. decoder_layer, num_decoder_layers, return_intermediate_dec)
  337. self.level_embed = nn.Embedding(num_feature_levels, hidden_dim)
  338. self.tgt_embed = nn.Embedding(num_queries, hidden_dim)
  339. self.query_pos_embed = nn.Embedding(num_queries, hidden_dim)
  340. self.reference_points = nn.Linear(
  341. hidden_dim,
  342. 2,
  343. weight_attr=ParamAttr(learning_rate=lr_mult),
  344. bias_attr=ParamAttr(learning_rate=lr_mult))
  345. self.input_proj = nn.LayerList()
  346. for in_channels in backbone_num_channels:
  347. self.input_proj.append(
  348. nn.Sequential(
  349. nn.Conv2D(
  350. in_channels,
  351. hidden_dim,
  352. kernel_size=1,
  353. weight_attr=weight_attr,
  354. bias_attr=bias_attr),
  355. nn.GroupNorm(32, hidden_dim)))
  356. in_channels = backbone_num_channels[-1]
  357. for _ in range(num_feature_levels - len(backbone_num_channels)):
  358. self.input_proj.append(
  359. nn.Sequential(
  360. nn.Conv2D(
  361. in_channels,
  362. hidden_dim,
  363. kernel_size=3,
  364. stride=2,
  365. padding=1,
  366. weight_attr=weight_attr,
  367. bias_attr=bias_attr),
  368. nn.GroupNorm(32, hidden_dim)))
  369. in_channels = hidden_dim
  370. self.position_embedding = PositionEmbedding(
  371. hidden_dim // 2,
  372. normalize=True if position_embed_type == 'sine' else False,
  373. embed_type=position_embed_type,
  374. offset=-0.5)
  375. self._reset_parameters()
  376. def _reset_parameters(self):
  377. normal_(self.level_embed.weight)
  378. normal_(self.tgt_embed.weight)
  379. normal_(self.query_pos_embed.weight)
  380. xavier_uniform_(self.reference_points.weight)
  381. constant_(self.reference_points.bias)
  382. for l in self.input_proj:
  383. xavier_uniform_(l[0].weight)
  384. constant_(l[0].bias)
  385. @classmethod
  386. def from_config(cls, cfg, input_shape):
  387. return {'backbone_num_channels': [i.channels for i in input_shape], }
  388. def _get_valid_ratio(self, mask):
  389. mask = mask.astype(paddle.float32)
  390. _, H, W = mask.shape
  391. valid_ratio_h = paddle.sum(mask[:, :, 0], 1) / H
  392. valid_ratio_w = paddle.sum(mask[:, 0, :], 1) / W
  393. valid_ratio = paddle.stack([valid_ratio_w, valid_ratio_h], -1)
  394. return valid_ratio
  395. def forward(self, src_feats, src_mask=None):
  396. srcs = []
  397. for i in range(len(src_feats)):
  398. srcs.append(self.input_proj[i](src_feats[i]))
  399. if self.num_feature_levels > len(srcs):
  400. len_srcs = len(srcs)
  401. for i in range(len_srcs, self.num_feature_levels):
  402. if i == len_srcs:
  403. srcs.append(self.input_proj[i](src_feats[-1]))
  404. else:
  405. srcs.append(self.input_proj[i](srcs[-1]))
  406. src_flatten = []
  407. mask_flatten = []
  408. lvl_pos_embed_flatten = []
  409. spatial_shapes = []
  410. valid_ratios = []
  411. for level, src in enumerate(srcs):
  412. bs, c, h, w = src.shape
  413. spatial_shapes.append([h, w])
  414. src = src.flatten(2).transpose([0, 2, 1])
  415. src_flatten.append(src)
  416. if src_mask is not None:
  417. mask = F.interpolate(
  418. src_mask.unsqueeze(0).astype(src.dtype),
  419. size=(h, w))[0].astype('bool')
  420. else:
  421. mask = paddle.ones([bs, h, w], dtype='bool')
  422. valid_ratios.append(self._get_valid_ratio(mask))
  423. pos_embed = self.position_embedding(mask).flatten(2).transpose(
  424. [0, 2, 1])
  425. lvl_pos_embed = pos_embed + self.level_embed.weight[level].reshape(
  426. [1, 1, -1])
  427. lvl_pos_embed_flatten.append(lvl_pos_embed)
  428. mask = mask.astype(src.dtype).flatten(1)
  429. mask_flatten.append(mask)
  430. src_flatten = paddle.concat(src_flatten, 1)
  431. mask_flatten = paddle.concat(mask_flatten, 1)
  432. lvl_pos_embed_flatten = paddle.concat(lvl_pos_embed_flatten, 1)
  433. # [l, 2]
  434. spatial_shapes = paddle.to_tensor(spatial_shapes, dtype='int64')
  435. # [b, l, 2]
  436. valid_ratios = paddle.stack(valid_ratios, 1)
  437. # encoder
  438. memory = self.encoder(src_flatten, spatial_shapes, mask_flatten,
  439. lvl_pos_embed_flatten, valid_ratios)
  440. # prepare input for decoder
  441. bs, _, c = memory.shape
  442. query_embed = self.query_pos_embed.weight.unsqueeze(0).tile([bs, 1, 1])
  443. tgt = self.tgt_embed.weight.unsqueeze(0).tile([bs, 1, 1])
  444. reference_points = F.sigmoid(self.reference_points(query_embed))
  445. reference_points_input = reference_points.unsqueeze(
  446. 2) * valid_ratios.unsqueeze(1)
  447. # decoder
  448. hs = self.decoder(tgt, reference_points_input, memory, spatial_shapes,
  449. mask_flatten, query_embed)
  450. return (hs, memory, reference_points)