deformable_transformer.py 19 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514
  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. from __future__ import absolute_import
  15. from __future__ import division
  16. from __future__ import print_function
  17. import math
  18. import paddle
  19. import paddle.nn as nn
  20. import paddle.nn.functional as F
  21. from paddle import ParamAttr
  22. from paddlex.ppdet.core.workspace import register
  23. from ..layers import MultiHeadAttention
  24. from .position_encoding import PositionEmbedding
  25. from .utils import _get_clones, deformable_attention_core_func
  26. from ..initializer import linear_init_, constant_, xavier_uniform_, normal_
  27. __all__ = ['DeformableTransformer']
  28. class MSDeformableAttention(nn.Layer):
  29. def __init__(self,
  30. embed_dim=256,
  31. num_heads=8,
  32. num_levels=4,
  33. num_points=4,
  34. lr_mult=0.1):
  35. """
  36. Multi-Scale Deformable Attention Module
  37. """
  38. super(MSDeformableAttention, self).__init__()
  39. self.embed_dim = embed_dim
  40. self.num_heads = num_heads
  41. self.num_levels = num_levels
  42. self.num_points = num_points
  43. self.total_points = num_heads * num_levels * num_points
  44. self.head_dim = embed_dim // num_heads
  45. assert self.head_dim * num_heads == self.embed_dim, "embed_dim must be divisible by num_heads"
  46. self.sampling_offsets = nn.Linear(
  47. embed_dim,
  48. self.total_points * 2,
  49. weight_attr=ParamAttr(learning_rate=lr_mult),
  50. bias_attr=ParamAttr(learning_rate=lr_mult))
  51. self.attention_weights = nn.Linear(embed_dim, self.total_points)
  52. self.value_proj = nn.Linear(embed_dim, embed_dim)
  53. self.output_proj = nn.Linear(embed_dim, embed_dim)
  54. self._reset_parameters()
  55. def _reset_parameters(self):
  56. # sampling_offsets
  57. constant_(self.sampling_offsets.weight)
  58. thetas = paddle.arange(
  59. self.num_heads,
  60. dtype=paddle.float32) * (2.0 * math.pi / self.num_heads)
  61. grid_init = paddle.stack([thetas.cos(), thetas.sin()], -1)
  62. grid_init = grid_init / grid_init.abs().max(-1, keepdim=True)
  63. grid_init = grid_init.reshape([self.num_heads, 1, 1, 2]).tile(
  64. [1, self.num_levels, self.num_points, 1])
  65. scaling = paddle.arange(
  66. 1, self.num_points + 1,
  67. dtype=paddle.float32).reshape([1, 1, -1, 1])
  68. grid_init *= scaling
  69. self.sampling_offsets.bias.set_value(grid_init.flatten())
  70. # attention_weights
  71. constant_(self.attention_weights.weight)
  72. constant_(self.attention_weights.bias)
  73. # proj
  74. xavier_uniform_(self.value_proj.weight)
  75. constant_(self.value_proj.bias)
  76. xavier_uniform_(self.output_proj.weight)
  77. constant_(self.output_proj.bias)
  78. def forward(self,
  79. query,
  80. reference_points,
  81. value,
  82. value_spatial_shapes,
  83. value_mask=None):
  84. """
  85. Args:
  86. query (Tensor): [bs, query_length, C]
  87. reference_points (Tensor): [bs, query_length, n_levels, 2], range in [0, 1], top-left (0,0),
  88. bottom-right (1, 1), including padding area
  89. value (Tensor): [bs, value_length, C]
  90. value_spatial_shapes (Tensor): [n_levels, 2], [(H_0, W_0), (H_1, W_1), ..., (H_{L-1}, W_{L-1})]
  91. value_mask (Tensor): [bs, value_length], True for non-padding elements, False for padding elements
  92. Returns:
  93. output (Tensor): [bs, Length_{query}, C]
  94. """
  95. bs, Len_q = query.shape[:2]
  96. Len_v = value.shape[1]
  97. assert int(value_spatial_shapes.prod(1).sum()) == Len_v
  98. value = self.value_proj(value)
  99. if value_mask is not None:
  100. value_mask = value_mask.astype(value.dtype).unsqueeze(-1)
  101. value *= value_mask
  102. value = value.reshape([bs, Len_v, self.num_heads, self.head_dim])
  103. sampling_offsets = self.sampling_offsets(query).reshape(
  104. [bs, Len_q, self.num_heads, self.num_levels, self.num_points, 2])
  105. attention_weights = self.attention_weights(query).reshape(
  106. [bs, Len_q, self.num_heads, self.num_levels * self.num_points])
  107. attention_weights = F.softmax(attention_weights, -1).reshape(
  108. [bs, Len_q, self.num_heads, self.num_levels, self.num_points])
  109. offset_normalizer = value_spatial_shapes.flip([1]).reshape(
  110. [1, 1, 1, self.num_levels, 1, 2])
  111. sampling_locations = reference_points.reshape([
  112. bs, Len_q, 1, self.num_levels, 1, 2
  113. ]) + sampling_offsets / offset_normalizer
  114. output = deformable_attention_core_func(
  115. value, value_spatial_shapes, sampling_locations, attention_weights)
  116. output = self.output_proj(output)
  117. return output
  118. class DeformableTransformerEncoderLayer(nn.Layer):
  119. def __init__(self,
  120. d_model=256,
  121. n_head=8,
  122. dim_feedforward=1024,
  123. dropout=0.1,
  124. activation="relu",
  125. n_levels=4,
  126. n_points=4,
  127. weight_attr=None,
  128. bias_attr=None):
  129. super(DeformableTransformerEncoderLayer, self).__init__()
  130. # self attention
  131. self.self_attn = MSDeformableAttention(d_model, n_head, n_levels,
  132. n_points)
  133. self.dropout1 = nn.Dropout(dropout)
  134. self.norm1 = nn.LayerNorm(d_model)
  135. # ffn
  136. self.linear1 = nn.Linear(d_model, dim_feedforward, weight_attr,
  137. bias_attr)
  138. self.activation = getattr(F, activation)
  139. self.dropout2 = nn.Dropout(dropout)
  140. self.linear2 = nn.Linear(dim_feedforward, d_model, weight_attr,
  141. bias_attr)
  142. self.dropout3 = nn.Dropout(dropout)
  143. self.norm2 = nn.LayerNorm(d_model)
  144. self._reset_parameters()
  145. def _reset_parameters(self):
  146. linear_init_(self.linear1)
  147. linear_init_(self.linear2)
  148. xavier_uniform_(self.linear1.weight)
  149. xavier_uniform_(self.linear2.weight)
  150. def with_pos_embed(self, tensor, pos):
  151. return tensor if pos is None else tensor + pos
  152. def forward_ffn(self, src):
  153. src2 = self.linear2(self.dropout2(self.activation(self.linear1(src))))
  154. src = src + self.dropout3(src2)
  155. src = self.norm2(src)
  156. return src
  157. def forward(self,
  158. src,
  159. reference_points,
  160. spatial_shapes,
  161. src_mask=None,
  162. pos_embed=None):
  163. # self attention
  164. src2 = self.self_attn(
  165. self.with_pos_embed(src, pos_embed), reference_points, src,
  166. spatial_shapes, src_mask)
  167. src = src + self.dropout1(src2)
  168. src = self.norm1(src)
  169. # ffn
  170. src = self.forward_ffn(src)
  171. return src
  172. class DeformableTransformerEncoder(nn.Layer):
  173. def __init__(self, encoder_layer, num_layers):
  174. super(DeformableTransformerEncoder, self).__init__()
  175. self.layers = _get_clones(encoder_layer, num_layers)
  176. self.num_layers = num_layers
  177. @staticmethod
  178. def get_reference_points(spatial_shapes, valid_ratios):
  179. valid_ratios = valid_ratios.unsqueeze(1)
  180. reference_points = []
  181. for i, (H, W) in enumerate(spatial_shapes.tolist()):
  182. ref_y, ref_x = paddle.meshgrid(
  183. paddle.linspace(0.5, H - 0.5, H),
  184. paddle.linspace(0.5, W - 0.5, W))
  185. ref_y = ref_y.flatten().unsqueeze(0) / (valid_ratios[:, :, i, 1] *
  186. H)
  187. ref_x = ref_x.flatten().unsqueeze(0) / (valid_ratios[:, :, i, 0] *
  188. W)
  189. reference_points.append(paddle.stack((ref_x, ref_y), axis=-1))
  190. reference_points = paddle.concat(reference_points, 1).unsqueeze(2)
  191. reference_points = reference_points * valid_ratios
  192. return reference_points
  193. def forward(self,
  194. src,
  195. spatial_shapes,
  196. src_mask=None,
  197. pos_embed=None,
  198. valid_ratios=None):
  199. output = src
  200. if valid_ratios is None:
  201. valid_ratios = paddle.ones(
  202. [src.shape[0], spatial_shapes.shape[0], 2])
  203. reference_points = self.get_reference_points(spatial_shapes,
  204. valid_ratios)
  205. for layer in self.layers:
  206. output = layer(output, reference_points, spatial_shapes, src_mask,
  207. pos_embed)
  208. return output
  209. class DeformableTransformerDecoderLayer(nn.Layer):
  210. def __init__(self,
  211. d_model=256,
  212. n_head=8,
  213. dim_feedforward=1024,
  214. dropout=0.1,
  215. activation="relu",
  216. n_levels=4,
  217. n_points=4,
  218. weight_attr=None,
  219. bias_attr=None):
  220. super(DeformableTransformerDecoderLayer, self).__init__()
  221. # self attention
  222. self.self_attn = MultiHeadAttention(d_model, n_head, dropout=dropout)
  223. self.dropout1 = nn.Dropout(dropout)
  224. self.norm1 = nn.LayerNorm(d_model)
  225. # cross attention
  226. self.cross_attn = MSDeformableAttention(d_model, n_head, n_levels,
  227. n_points)
  228. self.dropout2 = nn.Dropout(dropout)
  229. self.norm2 = nn.LayerNorm(d_model)
  230. # ffn
  231. self.linear1 = nn.Linear(d_model, dim_feedforward, weight_attr,
  232. bias_attr)
  233. self.activation = getattr(F, activation)
  234. self.dropout3 = nn.Dropout(dropout)
  235. self.linear2 = nn.Linear(dim_feedforward, d_model, weight_attr,
  236. bias_attr)
  237. self.dropout4 = nn.Dropout(dropout)
  238. self.norm3 = nn.LayerNorm(d_model)
  239. self._reset_parameters()
  240. def _reset_parameters(self):
  241. linear_init_(self.linear1)
  242. linear_init_(self.linear2)
  243. xavier_uniform_(self.linear1.weight)
  244. xavier_uniform_(self.linear2.weight)
  245. def with_pos_embed(self, tensor, pos):
  246. return tensor if pos is None else tensor + pos
  247. def forward_ffn(self, tgt):
  248. tgt2 = self.linear2(self.dropout3(self.activation(self.linear1(tgt))))
  249. tgt = tgt + self.dropout4(tgt2)
  250. tgt = self.norm3(tgt)
  251. return tgt
  252. def forward(self,
  253. tgt,
  254. reference_points,
  255. memory,
  256. memory_spatial_shapes,
  257. memory_mask=None,
  258. query_pos_embed=None):
  259. # self attention
  260. q = k = self.with_pos_embed(tgt, query_pos_embed)
  261. tgt2 = self.self_attn(q, k, value=tgt)
  262. tgt = tgt + self.dropout1(tgt2)
  263. tgt = self.norm1(tgt)
  264. # cross attention
  265. tgt2 = self.cross_attn(
  266. self.with_pos_embed(tgt, query_pos_embed), reference_points,
  267. memory, memory_spatial_shapes, memory_mask)
  268. tgt = tgt + self.dropout2(tgt2)
  269. tgt = self.norm2(tgt)
  270. # ffn
  271. tgt = self.forward_ffn(tgt)
  272. return tgt
  273. class DeformableTransformerDecoder(nn.Layer):
  274. def __init__(self, decoder_layer, num_layers, return_intermediate=False):
  275. super(DeformableTransformerDecoder, self).__init__()
  276. self.layers = _get_clones(decoder_layer, num_layers)
  277. self.num_layers = num_layers
  278. self.return_intermediate = return_intermediate
  279. def forward(self,
  280. tgt,
  281. reference_points,
  282. memory,
  283. memory_spatial_shapes,
  284. memory_mask=None,
  285. query_pos_embed=None):
  286. output = tgt
  287. intermediate = []
  288. for lid, layer in enumerate(self.layers):
  289. output = layer(output, reference_points, memory,
  290. memory_spatial_shapes, memory_mask, query_pos_embed)
  291. if self.return_intermediate:
  292. intermediate.append(output)
  293. if self.return_intermediate:
  294. return paddle.stack(intermediate)
  295. return output.unsqueeze(0)
  296. @register
  297. class DeformableTransformer(nn.Layer):
  298. __shared__ = ['hidden_dim']
  299. def __init__(self,
  300. num_queries=300,
  301. position_embed_type='sine',
  302. return_intermediate_dec=True,
  303. backbone_num_channels=[512, 1024, 2048],
  304. num_feature_levels=4,
  305. num_encoder_points=4,
  306. num_decoder_points=4,
  307. hidden_dim=256,
  308. nhead=8,
  309. num_encoder_layers=6,
  310. num_decoder_layers=6,
  311. dim_feedforward=1024,
  312. dropout=0.1,
  313. activation="relu",
  314. lr_mult=0.1,
  315. weight_attr=None,
  316. bias_attr=None):
  317. super(DeformableTransformer, self).__init__()
  318. assert position_embed_type in ['sine', 'learned'], \
  319. f'ValueError: position_embed_type not supported {position_embed_type}!'
  320. assert len(backbone_num_channels) <= num_feature_levels
  321. self.hidden_dim = hidden_dim
  322. self.nhead = nhead
  323. self.num_feature_levels = num_feature_levels
  324. encoder_layer = DeformableTransformerEncoderLayer(
  325. hidden_dim, nhead, dim_feedforward, dropout, activation,
  326. num_feature_levels, num_encoder_points, weight_attr, bias_attr)
  327. self.encoder = DeformableTransformerEncoder(encoder_layer,
  328. num_encoder_layers)
  329. decoder_layer = DeformableTransformerDecoderLayer(
  330. hidden_dim, nhead, dim_feedforward, dropout, activation,
  331. num_feature_levels, num_decoder_points, weight_attr, bias_attr)
  332. self.decoder = DeformableTransformerDecoder(
  333. decoder_layer, num_decoder_layers, return_intermediate_dec)
  334. self.level_embed = nn.Embedding(num_feature_levels, hidden_dim)
  335. self.tgt_embed = nn.Embedding(num_queries, hidden_dim)
  336. self.query_pos_embed = nn.Embedding(num_queries, hidden_dim)
  337. self.reference_points = nn.Linear(
  338. hidden_dim,
  339. 2,
  340. weight_attr=ParamAttr(learning_rate=lr_mult),
  341. bias_attr=ParamAttr(learning_rate=lr_mult))
  342. self.input_proj = nn.LayerList()
  343. for in_channels in backbone_num_channels:
  344. self.input_proj.append(
  345. nn.Sequential(
  346. nn.Conv2D(
  347. in_channels,
  348. hidden_dim,
  349. kernel_size=1,
  350. weight_attr=weight_attr,
  351. bias_attr=bias_attr),
  352. nn.GroupNorm(32, hidden_dim)))
  353. in_channels = backbone_num_channels[-1]
  354. for _ in range(num_feature_levels - len(backbone_num_channels)):
  355. self.input_proj.append(
  356. nn.Sequential(
  357. nn.Conv2D(
  358. in_channels,
  359. hidden_dim,
  360. kernel_size=3,
  361. stride=2,
  362. padding=1,
  363. weight_attr=weight_attr,
  364. bias_attr=bias_attr),
  365. nn.GroupNorm(32, hidden_dim)))
  366. in_channels = hidden_dim
  367. self.position_embedding = PositionEmbedding(
  368. hidden_dim // 2,
  369. normalize=True if position_embed_type == 'sine' else False,
  370. embed_type=position_embed_type,
  371. offset=-0.5)
  372. self._reset_parameters()
  373. def _reset_parameters(self):
  374. normal_(self.level_embed.weight)
  375. normal_(self.tgt_embed.weight)
  376. normal_(self.query_pos_embed.weight)
  377. xavier_uniform_(self.reference_points.weight)
  378. constant_(self.reference_points.bias)
  379. for l in self.input_proj:
  380. xavier_uniform_(l[0].weight)
  381. constant_(l[0].bias)
  382. @classmethod
  383. def from_config(cls, cfg, input_shape):
  384. return {'backbone_num_channels': [i.channels for i in input_shape], }
  385. def _get_valid_ratio(self, mask):
  386. mask = mask.astype(paddle.float32)
  387. _, H, W = mask.shape
  388. valid_ratio_h = paddle.sum(mask[:, :, 0], 1) / H
  389. valid_ratio_w = paddle.sum(mask[:, 0, :], 1) / W
  390. valid_ratio = paddle.stack([valid_ratio_w, valid_ratio_h], -1)
  391. return valid_ratio
  392. def forward(self, src_feats, src_mask=None):
  393. srcs = []
  394. for i in range(len(src_feats)):
  395. srcs.append(self.input_proj[i](src_feats[i]))
  396. if self.num_feature_levels > len(srcs):
  397. len_srcs = len(srcs)
  398. for i in range(len_srcs, self.num_feature_levels):
  399. if i == len_srcs:
  400. srcs.append(self.input_proj[i](src_feats[-1]))
  401. else:
  402. srcs.append(self.input_proj[i](srcs[-1]))
  403. src_flatten = []
  404. mask_flatten = []
  405. lvl_pos_embed_flatten = []
  406. spatial_shapes = []
  407. valid_ratios = []
  408. for level, src in enumerate(srcs):
  409. bs, c, h, w = src.shape
  410. spatial_shapes.append([h, w])
  411. src = src.flatten(2).transpose([0, 2, 1])
  412. src_flatten.append(src)
  413. if src_mask is not None:
  414. mask = F.interpolate(
  415. src_mask.unsqueeze(0).astype(src.dtype),
  416. size=(h, w))[0].astype('bool')
  417. else:
  418. mask = paddle.ones([bs, h, w], dtype='bool')
  419. valid_ratios.append(self._get_valid_ratio(mask))
  420. pos_embed = self.position_embedding(mask).flatten(2).transpose(
  421. [0, 2, 1])
  422. lvl_pos_embed = pos_embed + self.level_embed.weight[level].reshape(
  423. [1, 1, -1])
  424. lvl_pos_embed_flatten.append(lvl_pos_embed)
  425. mask = mask.astype(src.dtype).flatten(1)
  426. mask_flatten.append(mask)
  427. src_flatten = paddle.concat(src_flatten, 1)
  428. mask_flatten = paddle.concat(mask_flatten, 1)
  429. lvl_pos_embed_flatten = paddle.concat(lvl_pos_embed_flatten, 1)
  430. # [l, 2]
  431. spatial_shapes = paddle.to_tensor(spatial_shapes, dtype='int64')
  432. # [b, l, 2]
  433. valid_ratios = paddle.stack(valid_ratios, 1)
  434. # encoder
  435. memory = self.encoder(src_flatten, spatial_shapes, mask_flatten,
  436. lvl_pos_embed_flatten, valid_ratios)
  437. # prepare input for decoder
  438. bs, _, c = memory.shape
  439. query_embed = self.query_pos_embed.weight.unsqueeze(0).tile([bs, 1, 1])
  440. tgt = self.tgt_embed.weight.unsqueeze(0).tile([bs, 1, 1])
  441. reference_points = F.sigmoid(self.reference_points(query_embed))
  442. reference_points_input = reference_points.unsqueeze(
  443. 2) * valid_ratios.unsqueeze(1)
  444. # decoder
  445. hs = self.decoder(tgt, reference_points_input, memory, spatial_shapes,
  446. mask_flatten, query_embed)
  447. return (hs, memory, reference_points)