[transformer] add Deformable DETR base code (#3718)

ppdet/modeling/heads/detr_head.py

@@ -21,9 +21,10 @@ import paddle.nn as nn
 import paddle.nn.functional as F
 from ppdet.core.workspace import register
 import pycocotools.mask as mask_util
-from ..initializer import linear_init_
+from ..initializer import linear_init_, constant_
+from ..transformers.utils import inverse_sigmoid
 
-__all__ = ['DETRHead']
+__all__ = ['DETRHead', 'DeformableDETRHead']
 
 
 class MLP(nn.Layer):
@@ -275,3 +276,77 @@ class DETRHead(nn.Layer):
                 gt_mask=gt_mask)
         else:
             return (outputs_bbox[-1], outputs_logit[-1], outputs_seg)
+
+
+@register
+class DeformableDETRHead(nn.Layer):
+    __shared__ = ['num_classes', 'hidden_dim']
+    __inject__ = ['loss']
+
+    def __init__(self,
+                 num_classes=80,
+                 hidden_dim=512,
+                 nhead=8,
+                 num_mlp_layers=3,
+                 loss='DETRLoss'):
+        super(DeformableDETRHead, self).__init__()
+        self.num_classes = num_classes
+        self.hidden_dim = hidden_dim
+        self.nhead = nhead
+        self.loss = loss
+
+        self.score_head = nn.Linear(hidden_dim, self.num_classes)
+        self.bbox_head = MLP(hidden_dim,
+                             hidden_dim,
+                             output_dim=4,
+                             num_layers=num_mlp_layers)
+
+        self._reset_parameters()
+
+    def _reset_parameters(self):
+        linear_init_(self.score_head)
+        constant_(self.score_head.bias, -4.595)
+        constant_(self.bbox_head.layers[-1].weight)
+        bias = paddle.zeros_like(self.bbox_head.layers[-1].bias)
+        bias[2:] = -2.0
+        self.bbox_head.layers[-1].bias.set_value(bias)
+
+    @classmethod
+    def from_config(cls, cfg, hidden_dim, nhead, input_shape):
+        return {'hidden_dim': hidden_dim, 'nhead': nhead}
+
+    def forward(self, out_transformer, body_feats, inputs=None):
+        r"""
+        Args:
+            out_transformer (Tuple): (feats: [num_levels, batch_size,
+                                                num_queries, hidden_dim],
+                            memory: [batch_size,
+                                \sum_{l=0}^{L-1} H_l \cdot W_l, hidden_dim],
+                            reference_points: [batch_size, num_queries, 2])
+            body_feats (List(Tensor)): list[[B, C, H, W]]
+            inputs (dict): dict(inputs)
+        """
+        feats, memory, reference_points = out_transformer
+        reference_points = inverse_sigmoid(reference_points.unsqueeze(0))
+        outputs_bbox = self.bbox_head(feats)
+
+        # It's equivalent to "outputs_bbox[:, :, :, :2] += reference_points",
+        # but the gradient is wrong in paddle.
+        outputs_bbox = paddle.concat(
+            [
+                outputs_bbox[:, :, :, :2] + reference_points,
+                outputs_bbox[:, :, :, 2:]
+            ],
+            axis=-1)
+
+        outputs_bbox = F.sigmoid(outputs_bbox)
+        outputs_logit = self.score_head(feats)
+
+        if self.training:
+            assert inputs is not None
+            assert 'gt_bbox' in inputs and 'gt_class' in inputs
+
+            return self.loss(outputs_bbox, outputs_logit, inputs['gt_bbox'],
+                             inputs['gt_class'])
+        else:
+            return (outputs_bbox[-1], outputs_logit[-1], None)

ppdet/modeling/post_process.py

@@ -532,12 +532,23 @@ class DETRBBoxPostProcess(object):
 
         scores = F.sigmoid(logits) if self.use_focal_loss else F.softmax(
             logits)[:, :, :-1]
-        scores, labels = scores.max(-1), scores.argmax(-1)
 
-        if scores.shape[1] > self.num_top_queries:
-            scores, index = paddle.topk(scores, self.num_top_queries, axis=-1)
-            labels = paddle.stack(
-                [paddle.gather(l, i) for l, i in zip(labels, index)])
+        if not self.use_focal_loss:
+            scores, labels = scores.max(-1), scores.argmax(-1)
+            if scores.shape[1] > self.num_top_queries:
+                scores, index = paddle.topk(
+                    scores, self.num_top_queries, axis=-1)
+                labels = paddle.stack(
+                    [paddle.gather(l, i) for l, i in zip(labels, index)])
+                bbox_pred = paddle.stack(
+                    [paddle.gather(b, i) for b, i in zip(bbox_pred, index)])
+        else:
+            scores, index = paddle.topk(
+                scores.reshape([logits.shape[0], -1]),
+                self.num_top_queries,
+                axis=-1)
+            labels = index % logits.shape[2]
+            index = index // logits.shape[2]
             bbox_pred = paddle.stack(
                 [paddle.gather(b, i) for b, i in zip(bbox_pred, index)])
 

ppdet/modeling/transformers/__init__.py

@@ -16,8 +16,10 @@ from . import detr_transformer
 from . import utils
 from . import matchers
 from . import position_encoding
+from . import deformable_transformer
 
 from .detr_transformer import *
 from .utils import *
 from .matchers import *
 from .position_encoding import *
+from .deformable_transformer import *

ppdet/modeling/transformers/deformable_transformer.py

+# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import math
+import paddle
+import paddle.nn as nn
+import paddle.nn.functional as F
+from paddle import ParamAttr
+
+from ppdet.core.workspace import register
+from ..layers import MultiHeadAttention
+from .position_encoding import PositionEmbedding
+from .utils import _get_clones, deformable_attention_core_func
+from ..initializer import linear_init_, constant_, xavier_uniform_, normal_
+
+__all__ = ['DeformableTransformer']
+
+
+class MSDeformableAttention(nn.Layer):
+    def __init__(self,
+                 embed_dim=256,
+                 num_heads=8,
+                 num_levels=4,
+                 num_points=4,
+                 lr_mult=0.1):
+        """
+        Multi-Scale Deformable Attention Module
+        """
+        super(MSDeformableAttention, self).__init__()
+        self.embed_dim = embed_dim
+        self.num_heads = num_heads
+        self.num_levels = num_levels
+        self.num_points = num_points
+        self.total_points = num_heads * num_levels * num_points
+
+        self.head_dim = embed_dim // num_heads
+        assert self.head_dim * num_heads == self.embed_dim, "embed_dim must be divisible by num_heads"
+
+        self.sampling_offsets = nn.Linear(
+            embed_dim,
+            self.total_points * 2,
+            weight_attr=ParamAttr(learning_rate=lr_mult),
+            bias_attr=ParamAttr(learning_rate=lr_mult))
+
+        self.attention_weights = nn.Linear(embed_dim, self.total_points)
+        self.value_proj = nn.Linear(embed_dim, embed_dim)
+        self.output_proj = nn.Linear(embed_dim, embed_dim)
+
+        self._reset_parameters()
+
+    def _reset_parameters(self):
+        # sampling_offsets
+        constant_(self.sampling_offsets.weight)
+        thetas = paddle.arange(
+            self.num_heads,
+            dtype=paddle.float32) * (2.0 * math.pi / self.num_heads)
+        grid_init = paddle.stack([thetas.cos(), thetas.sin()], -1)
+        grid_init = grid_init / grid_init.abs().max(-1, keepdim=True)
+        grid_init = grid_init.reshape([self.num_heads, 1, 1, 2]).tile(
+            [1, self.num_levels, self.num_points, 1])
+        scaling = paddle.arange(
+            1, self.num_points + 1,
+            dtype=paddle.float32).reshape([1, 1, -1, 1])
+        grid_init *= scaling
+        self.sampling_offsets.bias.set_value(grid_init.flatten())
+        # attention_weights
+        constant_(self.attention_weights.weight)
+        constant_(self.attention_weights.bias)
+        # proj
+        xavier_uniform_(self.value_proj.weight)
+        constant_(self.value_proj.bias)
+        xavier_uniform_(self.output_proj.weight)
+        constant_(self.output_proj.bias)
+
+    def forward(self,
+                query,
+                reference_points,
+                value,
+                value_spatial_shapes,
+                value_mask=None):
+        """
+        Args:
+            query (Tensor): [bs, query_length, C]
+            reference_points (Tensor): [bs, query_length, n_levels, 2], range in [0, 1], top-left (0,0),
+                bottom-right (1, 1), including padding area
+            value (Tensor): [bs, value_length, C]
+            value_spatial_shapes (Tensor): [n_levels, 2], [(H_0, W_0), (H_1, W_1), ..., (H_{L-1}, W_{L-1})]
+            value_mask (Tensor): [bs, value_length], True for non-padding elements, False for padding elements
+
+        Returns:
+            output (Tensor): [bs, Length_{query}, C]
+        """
+        bs, Len_q = query.shape[:2]
+        Len_v = value.shape[1]
+        assert int(value_spatial_shapes.prod(1).sum()) == Len_v
+
+        value = self.value_proj(value)
+        if value_mask is not None:
+            value_mask = value_mask.astype(value.dtype).unsqueeze(-1)
+            value *= value_mask
+        value = value.reshape([bs, Len_v, self.num_heads, self.head_dim])
+
+        sampling_offsets = self.sampling_offsets(query).reshape(
+            [bs, Len_q, self.num_heads, self.num_levels, self.num_points, 2])
+        attention_weights = self.attention_weights(query).reshape(
+            [bs, Len_q, self.num_heads, self.num_levels * self.num_points])
+        attention_weights = F.softmax(attention_weights, -1).reshape(
+            [bs, Len_q, self.num_heads, self.num_levels, self.num_points])
+
+        offset_normalizer = value_spatial_shapes.flip([1]).reshape(
+            [1, 1, 1, self.num_levels, 1, 2])
+        sampling_locations = reference_points.reshape([
+            bs, Len_q, 1, self.num_levels, 1, 2
+        ]) + sampling_offsets / offset_normalizer
+
+        output = deformable_attention_core_func(
+            value, value_spatial_shapes, sampling_locations, attention_weights)
+        output = self.output_proj(output)
+
+        return output
+
+
+class DeformableTransformerEncoderLayer(nn.Layer):
+    def __init__(self,
+                 d_model=256,
+                 n_head=8,
+                 dim_feedforward=1024,
+                 dropout=0.1,
+                 activation="relu",
+                 n_levels=4,
+                 n_points=4,
+                 weight_attr=None,
+                 bias_attr=None):
+        super(DeformableTransformerEncoderLayer, self).__init__()
+        # self attention
+        self.self_attn = MSDeformableAttention(d_model, n_head, n_levels,
+                                               n_points)
+        self.dropout1 = nn.Dropout(dropout)
+        self.norm1 = nn.LayerNorm(d_model)
+        # ffn
+        self.linear1 = nn.Linear(d_model, dim_feedforward, weight_attr,
+                                 bias_attr)
+        self.activation = getattr(F, activation)
+        self.dropout2 = nn.Dropout(dropout)
+        self.linear2 = nn.Linear(dim_feedforward, d_model, weight_attr,
+                                 bias_attr)
+        self.dropout3 = nn.Dropout(dropout)
+        self.norm2 = nn.LayerNorm(d_model)
+        self._reset_parameters()
+
+    def _reset_parameters(self):
+        linear_init_(self.linear1)
+        linear_init_(self.linear2)
+        xavier_uniform_(self.linear1.weight)
+        xavier_uniform_(self.linear2.weight)
+
+    def with_pos_embed(self, 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,
+                reference_points,
+                spatial_shapes,
+                src_mask=None,
+                pos_embed=None):
+        # self attention
+        src2 = self.self_attn(
+            self.with_pos_embed(src, pos_embed), reference_points, src,
+            spatial_shapes, src_mask)
+        src = src + self.dropout1(src2)
+        src = self.norm1(src)
+        # ffn
+        src = self.forward_ffn(src)
+
+        return src
+
+
+class DeformableTransformerEncoder(nn.Layer):
+    def __init__(self, encoder_layer, num_layers):
+        super(DeformableTransformerEncoder, self).__init__()
+        self.layers = _get_clones(encoder_layer, num_layers)
+        self.num_layers = num_layers
+
+    @staticmethod
+    def get_reference_points(spatial_shapes, valid_ratios):
+        valid_ratios = valid_ratios.unsqueeze(1)
+        reference_points = []
+        for i, (H, W) in enumerate(spatial_shapes.tolist()):
+            ref_y, ref_x = paddle.meshgrid(
+                paddle.linspace(0.5, H - 0.5, H),
+                paddle.linspace(0.5, W - 0.5, W))
+            ref_y = ref_y.flatten().unsqueeze(0) / (valid_ratios[:, :, i, 1] *
+                                                    H)
+            ref_x = ref_x.flatten().unsqueeze(0) / (valid_ratios[:, :, i, 0] *
+                                                    W)
+            reference_points.append(paddle.stack((ref_x, ref_y), axis=-1))
+        reference_points = paddle.concat(reference_points, 1).unsqueeze(2)
+        reference_points = reference_points * valid_ratios
+        return reference_points
+
+    def forward(self,
+                src,
+                spatial_shapes,
+                src_mask=None,
+                pos_embed=None,
+                valid_ratios=None):
+        output = src
+        if valid_ratios is None:
+            valid_ratios = paddle.ones(
+                [src.shape[0], spatial_shapes.shape[0], 2])
+        reference_points = self.get_reference_points(spatial_shapes,
+                                                     valid_ratios)
+        for layer in self.layers:
+            output = layer(output, reference_points, spatial_shapes, src_mask,
+                           pos_embed)
+
+        return output
+
+
+class DeformableTransformerDecoderLayer(nn.Layer):
+    def __init__(self,
+                 d_model=256,
+                 n_head=8,
+                 dim_feedforward=1024,
+                 dropout=0.1,
+                 activation="relu",
+                 n_levels=4,
+                 n_points=4,
+                 weight_attr=None,
+                 bias_attr=None):
+        super(DeformableTransformerDecoderLayer, self).__init__()
+
+        # self attention
+        self.self_attn = MultiHeadAttention(d_model, n_head, dropout=dropout)
+        self.dropout1 = nn.Dropout(dropout)
+        self.norm1 = nn.LayerNorm(d_model)
+
+        # cross attention
+        self.cross_attn = MSDeformableAttention(d_model, n_head, n_levels,
+                                                n_points)
+        self.dropout2 = nn.Dropout(dropout)
+        self.norm2 = nn.LayerNorm(d_model)
+
+        # ffn
+        self.linear1 = nn.Linear(d_model, dim_feedforward, weight_attr,
+                                 bias_attr)
+        self.activation = getattr(F, activation)
+        self.dropout3 = nn.Dropout(dropout)
+        self.linear2 = nn.Linear(dim_feedforward, d_model, weight_attr,
+                                 bias_attr)
+        self.dropout4 = nn.Dropout(dropout)
+        self.norm3 = nn.LayerNorm(d_model)
+        self._reset_parameters()
+
+    def _reset_parameters(self):
+        linear_init_(self.linear1)
+        linear_init_(self.linear2)
+        xavier_uniform_(self.linear1.weight)
+        xavier_uniform_(self.linear2.weight)
+
+    def with_pos_embed(self, tensor, pos):
+        return tensor if pos is None else tensor + pos
+
+    def forward_ffn(self, tgt):
+        tgt2 = self.linear2(self.dropout3(self.activation(self.linear1(tgt))))
+        tgt = tgt + self.dropout4(tgt2)
+        tgt = self.norm3(tgt)
+        return tgt
+
+    def forward(self,
+                tgt,
+                reference_points,
+                memory,
+                memory_spatial_shapes,
+                memory_mask=None,
+                query_pos_embed=None):
+        # self attention
+        q = k = self.with_pos_embed(tgt, query_pos_embed)
+        tgt2 = self.self_attn(q, k, value=tgt)
+        tgt = tgt + self.dropout1(tgt2)
+        tgt = self.norm1(tgt)
+
+        # cross attention
+        tgt2 = self.cross_attn(
+            self.with_pos_embed(tgt, query_pos_embed), reference_points, memory,
+            memory_spatial_shapes, memory_mask)
+        tgt = tgt + self.dropout2(tgt2)
+        tgt = self.norm2(tgt)
+
+        # ffn
+        tgt = self.forward_ffn(tgt)
+
+        return tgt
+
+
+class DeformableTransformerDecoder(nn.Layer):
+    def __init__(self, decoder_layer, num_layers, return_intermediate=False):
+        super(DeformableTransformerDecoder, self).__init__()
+        self.layers = _get_clones(decoder_layer, num_layers)
+        self.num_layers = num_layers
+        self.return_intermediate = return_intermediate
+
+    def forward(self,
+                tgt,
+                reference_points,
+                memory,
+                memory_spatial_shapes,
+                memory_mask=None,
+                query_pos_embed=None):
+        output = tgt
+        intermediate = []
+        for lid, layer in enumerate(self.layers):
+            output = layer(output, reference_points, memory,
+                           memory_spatial_shapes, memory_mask, query_pos_embed)
+
+            if self.return_intermediate:
+                intermediate.append(output)
+
+        if self.return_intermediate:
+            return paddle.stack(intermediate)
+
+        return output.unsqueeze(0)
+
+
+@register
+class DeformableTransformer(nn.Layer):
+    __shared__ = ['hidden_dim']
+
+    def __init__(self,
+                 num_queries=300,
+                 position_embed_type='sine',
+                 return_intermediate_dec=True,
+                 backbone_num_channels=[512, 1024, 2048],
+                 num_feature_levels=4,
+                 num_encoder_points=4,
+                 num_decoder_points=4,
+                 hidden_dim=256,
+                 nhead=8,
+                 num_encoder_layers=6,
+                 num_decoder_layers=6,
+                 dim_feedforward=1024,
+                 dropout=0.1,
+                 activation="relu",
+                 lr_mult=0.1,
+                 weight_attr=None,
+                 bias_attr=None):
+        super(DeformableTransformer, self).__init__()
+        assert position_embed_type in ['sine', 'learned'], \
+            f'ValueError: position_embed_type not supported {position_embed_type}!'
+        assert len(backbone_num_channels) <= num_feature_levels
+
+        self.hidden_dim = hidden_dim
+        self.nhead = nhead
+        self.num_feature_levels = num_feature_levels
+
+        encoder_layer = DeformableTransformerEncoderLayer(
+            hidden_dim, nhead, dim_feedforward, dropout, activation,
+            num_feature_levels, num_encoder_points, weight_attr, bias_attr)
+        self.encoder = DeformableTransformerEncoder(encoder_layer,
+                                                    num_encoder_layers)
+
+        decoder_layer = DeformableTransformerDecoderLayer(
+            hidden_dim, nhead, dim_feedforward, dropout, activation,
+            num_feature_levels, num_decoder_points, weight_attr, bias_attr)
+        self.decoder = DeformableTransformerDecoder(
+            decoder_layer, num_decoder_layers, return_intermediate_dec)
+
+        self.level_embed = nn.Embedding(num_feature_levels, hidden_dim)
+        self.tgt_embed = nn.Embedding(num_queries, hidden_dim)
+        self.query_pos_embed = nn.Embedding(num_queries, hidden_dim)
+
+        self.reference_points = nn.Linear(
+            hidden_dim,
+            2,
+            weight_attr=ParamAttr(learning_rate=lr_mult),
+            bias_attr=ParamAttr(learning_rate=lr_mult))
+
+        self.input_proj = nn.LayerList()
+        for in_channels in backbone_num_channels:
+            self.input_proj.append(
+                nn.Sequential(
+                    nn.Conv2D(
+                        in_channels,
+                        hidden_dim,
+                        kernel_size=1,
+                        weight_attr=weight_attr,
+                        bias_attr=bias_attr),
+                    nn.GroupNorm(32, hidden_dim)))
+        in_channels = backbone_num_channels[-1]
+        for _ in range(num_feature_levels - len(backbone_num_channels)):
+            self.input_proj.append(
+                nn.Sequential(
+                    nn.Conv2D(
+                        in_channels,
+                        hidden_dim,
+                        kernel_size=3,
+                        stride=2,
+                        padding=1,
+                        weight_attr=weight_attr,
+                        bias_attr=bias_attr),
+                    nn.GroupNorm(32, hidden_dim)))
+            in_channels = hidden_dim
+
+        self.position_embedding = PositionEmbedding(
+            hidden_dim // 2,
+            normalize=True if position_embed_type == 'sine' else False,
+            embed_type=position_embed_type,
+            offset=-0.5)
+
+        self._reset_parameters()
+
+    def _reset_parameters(self):
+        normal_(self.level_embed.weight)
+        normal_(self.tgt_embed.weight)
+        normal_(self.query_pos_embed.weight)
+        xavier_uniform_(self.reference_points.weight)
+        constant_(self.reference_points.bias)
+        for l in self.input_proj:
+            xavier_uniform_(l[0].weight)
+            constant_(l[0].bias)
+
+    @classmethod
+    def from_config(cls, cfg, input_shape):
+        return {'backbone_num_channels': [i.channels for i in input_shape], }
+
+    def _get_valid_ratio(self, mask):
+        mask = mask.astype(paddle.float32)
+        _, H, W = mask.shape
+        valid_ratio_h = paddle.sum(mask[:, :, 0], 1) / H
+        valid_ratio_w = paddle.sum(mask[:, 0, :], 1) / W
+        valid_ratio = paddle.stack([valid_ratio_w, valid_ratio_h], -1)
+        return valid_ratio
+
+    def forward(self, src_feats, src_mask=None):
+        srcs = []
+        for i in range(len(src_feats)):
+            srcs.append(self.input_proj[i](src_feats[i]))
+        if self.num_feature_levels > len(srcs):
+            len_srcs = len(srcs)
+            for i in range(len_srcs, self.num_feature_levels):
+                if i == len_srcs:
+                    srcs.append(self.input_proj[i](src_feats[-1]))
+                else:
+                    srcs.append(self.input_proj[i](srcs[-1]))
+        src_flatten = []
+        mask_flatten = []
+        lvl_pos_embed_flatten = []
+        spatial_shapes = []
+        valid_ratios = []
+        for level, src in enumerate(srcs):
+            bs, c, h, w = src.shape
+            spatial_shapes.append([h, w])
+            src = src.flatten(2).transpose([0, 2, 1])
+            src_flatten.append(src)
+            if src_mask is not None:
+                mask = F.interpolate(
+                    src_mask.unsqueeze(0).astype(src.dtype),
+                    size=(h, w))[0].astype('bool')
+            else:
+                mask = paddle.ones([bs, h, w], dtype='bool')
+            valid_ratios.append(self._get_valid_ratio(mask))
+            pos_embed = self.position_embedding(mask).flatten(2).transpose(
+                [0, 2, 1])
+            lvl_pos_embed = pos_embed + self.level_embed.weight[level].reshape(
+                [1, 1, -1])
+            lvl_pos_embed_flatten.append(lvl_pos_embed)
+            mask = mask.astype(src.dtype).flatten(1)
+            mask_flatten.append(mask)
+        src_flatten = paddle.concat(src_flatten, 1)
+        mask_flatten = paddle.concat(mask_flatten, 1)
+        lvl_pos_embed_flatten = paddle.concat(lvl_pos_embed_flatten, 1)
+        # [l, 2]
+        spatial_shapes = paddle.to_tensor(spatial_shapes, dtype='int64')
+        # [b, l, 2]
+        valid_ratios = paddle.stack(valid_ratios, 1)
+
+        # encoder
+        memory = self.encoder(src_flatten, spatial_shapes, mask_flatten,
+                              lvl_pos_embed_flatten, valid_ratios)
+
+        # prepare input for decoder
+        bs, _, c = memory.shape
+        query_embed = self.query_pos_embed.weight.unsqueeze(0).tile([bs, 1, 1])
+        tgt = self.tgt_embed.weight.unsqueeze(0).tile([bs, 1, 1])
+        reference_points = F.sigmoid(self.reference_points(query_embed))
+        reference_points_input = reference_points.unsqueeze(
+            2) * valid_ratios.unsqueeze(1)
+
+        # decoder
+        hs = self.decoder(tgt, reference_points_input, memory, spatial_shapes,
+                          mask_flatten, query_embed)
+
+        return (hs, memory, reference_points)

ppdet/modeling/transformers/position_encoding.py

@@ -32,11 +32,14 @@ class PositionEmbedding(nn.Layer):
                  normalize=True,
                  scale=None,
                  embed_type='sine',
-                 num_embeddings=50):
+                 num_embeddings=50,
+                 offset=0.):
         super(PositionEmbedding, self).__init__()
         assert embed_type in ['sine', 'learned']
 
         self.embed_type = embed_type
+        self.offset = offset
+        self.eps = 1e-6
         if self.embed_type == 'sine':
             self.num_pos_feats = num_pos_feats
             self.temperature = temperature
@@ -65,9 +68,10 @@ class PositionEmbedding(nn.Layer):
             y_embed = mask.cumsum(1, dtype='float32')
             x_embed = mask.cumsum(2, dtype='float32')
             if self.normalize:
-                eps = 1e-6
-                y_embed = y_embed / (y_embed[:, -1:, :] + eps) * self.scale
-                x_embed = x_embed / (x_embed[:, :, -1:] + eps) * self.scale
+                y_embed = (y_embed + self.offset) / (
+                    y_embed[:, -1:, :] + self.eps) * self.scale
+                x_embed = (x_embed + self.offset) / (
+                    x_embed[:, :, -1:] + self.eps) * self.scale
 
             dim_t = 2 * (paddle.arange(self.num_pos_feats) //
                          2).astype('float32')

ppdet/modeling/transformers/utils.py

@@ -25,7 +25,8 @@ from ..bbox_utils import bbox_overlaps
 
 __all__ = [
     '_get_clones', 'bbox_overlaps', 'bbox_cxcywh_to_xyxy',
-    'bbox_xyxy_to_cxcywh', 'sigmoid_focal_loss'
+    'bbox_xyxy_to_cxcywh', 'sigmoid_focal_loss', 'inverse_sigmoid',
+    'deformable_attention_core_func'
 ]
 
 
@@ -55,3 +56,51 @@ def sigmoid_focal_loss(logit, label, normalizer=1.0, alpha=0.25, gamma=2.0):
         alpha_t = alpha * label + (1 - alpha) * (1 - label)
         loss = alpha_t * loss
     return loss.mean(1).sum() / normalizer
+
+
+def inverse_sigmoid(x, eps=1e-6):
+    x = x.clip(min=0., max=1.)
+    return paddle.log(x / (1 - x + eps) + eps)
+
+
+def deformable_attention_core_func(value, value_spatial_shapes,
+                                   sampling_locations, attention_weights):
+    """
+    Args:
+        value (Tensor): [bs, value_length, n_head, c]
+        value_spatial_shapes (Tensor): [n_levels, 2]
+        sampling_locations (Tensor): [bs, query_length, n_head, n_levels, n_points, 2]
+        attention_weights (Tensor): [bs, query_length, n_head, n_levels, n_points]
+
+    Returns:
+        output (Tensor): [bs, Length_{query}, C]
+    """
+    bs, Len_v, n_head, c = value.shape
+    _, Len_q, n_head, n_levels, n_points, _ = sampling_locations.shape
+
+    value_list = value.split(value_spatial_shapes.prod(1).tolist(), axis=1)
+    sampling_grids = 2 * sampling_locations - 1
+    sampling_value_list = []
+    for level, (h, w) in enumerate(value_spatial_shapes.tolist()):
+        # N_, H_*W_, M_, D_ -> N_, H_*W_, M_*D_ -> N_, M_*D_, H_*W_ -> N_*M_, D_, H_, W_
+        value_l_ = value_list[level].flatten(2).transpose(
+            [0, 2, 1]).reshape([bs * n_head, c, h, w])
+        # N_, Lq_, M_, P_, 2 -> N_, M_, Lq_, P_, 2 -> N_*M_, Lq_, P_, 2
+        sampling_grid_l_ = sampling_grids[:, :, :, level].transpose(
+            [0, 2, 1, 3, 4]).flatten(0, 1)
+        # N_*M_, D_, Lq_, P_
+        sampling_value_l_ = F.grid_sample(
+            value_l_,
+            sampling_grid_l_,
+            mode='bilinear',
+            padding_mode='zeros',
+            align_corners=False)
+        sampling_value_list.append(sampling_value_l_)
+    # (N_, Lq_, M_, L_, P_) -> (N_, M_, Lq_, L_, P_) -> (N_*M_, 1, Lq_, L_*P_)
+    attention_weights = attention_weights.transpose([0, 2, 1, 3, 4]).reshape(
+        [bs * n_head, 1, Len_q, n_levels * n_points])
+    output = (paddle.stack(
+        sampling_value_list, axis=-2).flatten(-2) *
+              attention_weights).sum(-1).reshape([bs, n_head * c, Len_q])
+
+    return output.transpose([0, 2, 1])