detr_head.py 13.3 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23
# 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 paddle
import paddle.nn as nn
import paddle.nn.functional as F
from ppdet.core.workspace import register
import pycocotools.mask as mask_util
24 25
from ..initializer import linear_init_, constant_
from ..transformers.utils import inverse_sigmoid
26

27
__all__ = ['DETRHead', 'DeformableDETRHead']
28 29 30


class MLP(nn.Layer):
31 32 33 34
    """This code is based on
        https://github.com/facebookresearch/detr/blob/main/models/detr.py
    """

35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54
    def __init__(self, input_dim, hidden_dim, output_dim, num_layers):
        super().__init__()
        self.num_layers = num_layers
        h = [hidden_dim] * (num_layers - 1)
        self.layers = nn.LayerList(
            nn.Linear(n, k) for n, k in zip([input_dim] + h, h + [output_dim]))

        self._reset_parameters()

    def _reset_parameters(self):
        for l in self.layers:
            linear_init_(l)

    def forward(self, x):
        for i, layer in enumerate(self.layers):
            x = F.relu(layer(x)) if i < self.num_layers - 1 else layer(x)
        return x


class MultiHeadAttentionMap(nn.Layer):
55 56 57 58 59
    """This code is based on
        https://github.com/facebookresearch/detr/blob/main/models/segmentation.py

        This is a 2D attention module, which only returns the attention softmax (no multiplication by value)
    """
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

    def __init__(self, query_dim, hidden_dim, num_heads, dropout=0.0,
                 bias=True):
        super().__init__()
        self.num_heads = num_heads
        self.hidden_dim = hidden_dim
        self.dropout = nn.Dropout(dropout)

        weight_attr = paddle.ParamAttr(
            initializer=paddle.nn.initializer.XavierUniform())
        bias_attr = paddle.framework.ParamAttr(
            initializer=paddle.nn.initializer.Constant()) if bias else False

        self.q_proj = nn.Linear(query_dim, hidden_dim, weight_attr, bias_attr)
        self.k_proj = nn.Conv2D(
            query_dim,
            hidden_dim,
            1,
            weight_attr=weight_attr,
            bias_attr=bias_attr)

        self.normalize_fact = float(hidden_dim / self.num_heads)**-0.5

    def forward(self, q, k, mask=None):
        q = self.q_proj(q)
        k = self.k_proj(k)
        bs, num_queries, n, c, h, w = q.shape[0], q.shape[1], self.num_heads,\
                                      self.hidden_dim // self.num_heads, k.shape[-2], k.shape[-1]
        qh = q.reshape([bs, num_queries, n, c])
        kh = k.reshape([bs, n, c, h, w])
        # weights = paddle.einsum("bqnc,bnchw->bqnhw", qh * self.normalize_fact, kh)
        qh = qh.transpose([0, 2, 1, 3]).reshape([-1, num_queries, c])
        kh = kh.reshape([-1, c, h * w])
        weights = paddle.bmm(qh * self.normalize_fact, kh).reshape(
            [bs, n, num_queries, h, w]).transpose([0, 2, 1, 3, 4])

        if mask is not None:
            weights += mask
        # fix a potenial bug: https://github.com/facebookresearch/detr/issues/247
        weights = F.softmax(weights.flatten(3), axis=-1).reshape(weights.shape)
        weights = self.dropout(weights)
        return weights


class MaskHeadFPNConv(nn.Layer):
105 106 107 108 109
    """This code is based on
        https://github.com/facebookresearch/detr/blob/main/models/segmentation.py

        Simple convolutional head, using group norm.
        Upsampling is done using a FPN approach
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
    """

    def __init__(self, input_dim, fpn_dims, context_dim, num_groups=8):
        super().__init__()

        inter_dims = [input_dim,
                      ] + [context_dim // (2**i) for i in range(1, 5)]
        weight_attr = paddle.ParamAttr(
            initializer=paddle.nn.initializer.KaimingUniform())
        bias_attr = paddle.framework.ParamAttr(
            initializer=paddle.nn.initializer.Constant())

        self.conv0 = self._make_layers(input_dim, input_dim, 3, num_groups,
                                       weight_attr, bias_attr)
        self.conv_inter = nn.LayerList()
        for in_dims, out_dims in zip(inter_dims[:-1], inter_dims[1:]):
            self.conv_inter.append(
                self._make_layers(in_dims, out_dims, 3, num_groups, weight_attr,
                                  bias_attr))

        self.conv_out = nn.Conv2D(
            inter_dims[-1],
            1,
            3,
            padding=1,
            weight_attr=weight_attr,
            bias_attr=bias_attr)

        self.adapter = nn.LayerList()
        for i in range(len(fpn_dims)):
            self.adapter.append(
                nn.Conv2D(
                    fpn_dims[i],
                    inter_dims[i + 1],
                    1,
                    weight_attr=weight_attr,
                    bias_attr=bias_attr))

    def _make_layers(self,
                     in_dims,
                     out_dims,
                     kernel_size,
                     num_groups,
                     weight_attr=None,
                     bias_attr=None):
        return nn.Sequential(
            nn.Conv2D(
                in_dims,
                out_dims,
                kernel_size,
                padding=kernel_size // 2,
                weight_attr=weight_attr,
                bias_attr=bias_attr),
            nn.GroupNorm(num_groups, out_dims),
            nn.ReLU())

    def forward(self, x, bbox_attention_map, fpns):
        x = paddle.concat([
            x.tile([bbox_attention_map.shape[1], 1, 1, 1]),
            bbox_attention_map.flatten(0, 1)
        ], 1)
        x = self.conv0(x)
        for inter_layer, adapter_layer, feat in zip(self.conv_inter[:-1],
                                                    self.adapter, fpns):
            feat = adapter_layer(feat).tile(
                [bbox_attention_map.shape[1], 1, 1, 1])
            x = inter_layer(x)
            x = feat + F.interpolate(x, size=feat.shape[-2:])

        x = self.conv_inter[-1](x)
        x = self.conv_out(x)
        return x


@register
class DETRHead(nn.Layer):
    __shared__ = ['num_classes', 'hidden_dim', 'use_focal_loss']
    __inject__ = ['loss']

    def __init__(self,
                 num_classes=80,
                 hidden_dim=256,
                 nhead=8,
                 num_mlp_layers=3,
                 loss='DETRLoss',
                 fpn_dims=[1024, 512, 256],
                 with_mask_head=False,
                 use_focal_loss=False):
        super(DETRHead, self).__init__()
        # add background class
        self.num_classes = num_classes if use_focal_loss else num_classes + 1
        self.hidden_dim = hidden_dim
        self.loss = loss
        self.with_mask_head = with_mask_head
        self.use_focal_loss = use_focal_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)
        if self.with_mask_head:
            self.bbox_attention = MultiHeadAttentionMap(hidden_dim, hidden_dim,
                                                        nhead)
            self.mask_head = MaskHeadFPNConv(hidden_dim + nhead, fpn_dims,
                                             hidden_dim)
        self._reset_parameters()

    def _reset_parameters(self):
        linear_init_(self.score_head)

    @classmethod
    def from_config(cls, cfg, hidden_dim, nhead, input_shape):

        return {
            'hidden_dim': hidden_dim,
            'nhead': nhead,
            'fpn_dims': [i.channels for i in input_shape[::-1]][1:]
        }

    @staticmethod
    def get_gt_mask_from_polygons(gt_poly, pad_mask):
        out_gt_mask = []
        for polygons, padding in zip(gt_poly, pad_mask):
            height, width = int(padding[:, 0].sum()), int(padding[0, :].sum())
            masks = []
            for obj_poly in polygons:
                rles = mask_util.frPyObjects(obj_poly, height, width)
                rle = mask_util.merge(rles)
                masks.append(
                    paddle.to_tensor(mask_util.decode(rle)).astype('float32'))
            masks = paddle.stack(masks)
            masks_pad = paddle.zeros(
                [masks.shape[0], pad_mask.shape[1], pad_mask.shape[2]])
            masks_pad[:, :height, :width] = masks
            out_gt_mask.append(masks_pad)
        return out_gt_mask

    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, hidden_dim, h, w],
                            src_proj: [batch_size, h*w, hidden_dim],
                            src_mask: [batch_size, 1, 1, h, w])
            body_feats (List(Tensor)): list[[B, C, H, W]]
            inputs (dict): dict(inputs)
        """
        feats, memory, src_proj, src_mask = out_transformer
        outputs_logit = self.score_head(feats)
        outputs_bbox = F.sigmoid(self.bbox_head(feats))
        outputs_seg = None
        if self.with_mask_head:
            bbox_attention_map = self.bbox_attention(feats[-1], memory,
                                                     src_mask)
            fpn_feats = [a for a in body_feats[::-1]][1:]
            outputs_seg = self.mask_head(src_proj, bbox_attention_map,
                                         fpn_feats)
            outputs_seg = outputs_seg.reshape([
                feats.shape[1], feats.shape[2], outputs_seg.shape[-2],
                outputs_seg.shape[-1]
            ])

        if self.training:
            assert inputs is not None
            assert 'gt_bbox' in inputs and 'gt_class' in inputs
            gt_mask = self.get_gt_mask_from_polygons(
                inputs['gt_poly'],
                inputs['pad_mask']) if 'gt_poly' in inputs else None
            return self.loss(
                outputs_bbox,
                outputs_logit,
                inputs['gt_bbox'],
                inputs['gt_class'],
                masks=outputs_seg,
                gt_mask=gt_mask)
        else:
            return (outputs_bbox[-1], outputs_logit[-1], outputs_seg)
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


@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)
320

W
Wenyu 已提交
321 322 323 324
        with paddle.no_grad():
            bias = paddle.zeros_like(self.bbox_head.layers[-1].bias)
            bias[2:] = -2.0
            self.bbox_head.layers[-1].bias.set_value(bias)
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

    @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)