|
| 1 | +# Copyright (c) OpenMMLab. All rights reserved. |
| 2 | +import torch |
| 3 | +import torch.nn as nn |
| 4 | +import torch.nn.functional as F |
| 5 | +from mmcv.cnn import ConvModule |
| 6 | + |
| 7 | +from mmdet.core import multi_apply |
| 8 | +from mmdet.models.builder import HEADS |
| 9 | +from mmdet.models.dense_heads import YOLOXHead |
| 10 | +from mmdet.models.dense_heads.tood_head import TaskDecomposition |
| 11 | + |
| 12 | + |
| 13 | +@HEADS.register_module() |
| 14 | +class YOLOXTOODHead(YOLOXHead): |
| 15 | + """YOLOXTOODHead head used in `YOLOX-PAI. |
| 16 | +
|
| 17 | + <https://arxiv.org/abs/2208.13040>`_. |
| 18 | +
|
| 19 | + Args: |
| 20 | + tood_stacked_convs (int): Number of conv layers in TOOD head. |
| 21 | + Default: 3. |
| 22 | + la_down_rate (int): Downsample rate of layer attention. |
| 23 | + Default: 32. |
| 24 | + tood_norm_cfg (dict): Config dict for normalization layer. |
| 25 | + """ |
| 26 | + |
| 27 | + def __init__(self, |
| 28 | + *args, |
| 29 | + tood_stacked_convs=3, |
| 30 | + la_down_rate=32, |
| 31 | + tood_norm_cfg=dict( |
| 32 | + type='GN', num_groups=32, requires_grad=True), |
| 33 | + **kwargs): |
| 34 | + super().__init__(*args, **kwargs) |
| 35 | + self.tood_stacked_convs = tood_stacked_convs |
| 36 | + self.la_down_rate = la_down_rate |
| 37 | + self.tood_norm_cfg = tood_norm_cfg |
| 38 | + |
| 39 | + self._init_tood_layers() |
| 40 | + |
| 41 | + def _init_tood_layers(self): |
| 42 | + self.multi_level_cls_decomps = nn.ModuleList() |
| 43 | + self.multi_level_reg_decomps = nn.ModuleList() |
| 44 | + for _ in self.strides: |
| 45 | + self.multi_level_cls_decomps.append( |
| 46 | + TaskDecomposition(self.in_channels, self.tood_stacked_convs, |
| 47 | + self.tood_stacked_convs * self.la_down_rate, |
| 48 | + self.conv_cfg, self.tood_norm_cfg)) |
| 49 | + self.multi_level_reg_decomps.append( |
| 50 | + TaskDecomposition(self.in_channels, self.tood_stacked_convs, |
| 51 | + self.tood_stacked_convs * self.la_down_rate, |
| 52 | + self.conv_cfg, self.tood_norm_cfg)) |
| 53 | + |
| 54 | + self.inter_convs = nn.ModuleList() |
| 55 | + for _ in range(self.tood_stacked_convs): |
| 56 | + self.inter_convs.append( |
| 57 | + ConvModule( |
| 58 | + self.in_channels, |
| 59 | + self.in_channels, |
| 60 | + 3, |
| 61 | + stride=1, |
| 62 | + padding=1, |
| 63 | + conv_cfg=self.conv_cfg, |
| 64 | + norm_cfg=self.tood_norm_cfg)) |
| 65 | + |
| 66 | + def forward_single(self, x, cls_convs, reg_convs, conv_cls, conv_reg, |
| 67 | + conv_obj, cls_decomp, reg_decomp): |
| 68 | + """Forward feature of a single scale level.""" |
| 69 | + |
| 70 | + inter_feats = [] |
| 71 | + for inter_conv in self.inter_convs: |
| 72 | + x = inter_conv(x) |
| 73 | + inter_feats.append(x) |
| 74 | + feat = torch.cat(inter_feats, 1) |
| 75 | + |
| 76 | + avg_feat = F.adaptive_avg_pool2d(feat, (1, 1)) |
| 77 | + cls_x = cls_decomp(feat, avg_feat) |
| 78 | + reg_x = reg_decomp(feat, avg_feat) |
| 79 | + |
| 80 | + cls_feat = cls_convs(cls_x) |
| 81 | + reg_feat = reg_convs(reg_x) |
| 82 | + |
| 83 | + cls_score = conv_cls(cls_feat) |
| 84 | + bbox_pred = conv_reg(reg_feat) |
| 85 | + objectness = conv_obj(reg_feat) |
| 86 | + |
| 87 | + return cls_score, bbox_pred, objectness |
| 88 | + |
| 89 | + def forward(self, feats): |
| 90 | + """Forward features from the upstream network. |
| 91 | +
|
| 92 | + Args: |
| 93 | + feats (tuple[Tensor]): Features from the upstream network, each is |
| 94 | + a 4D-tensor. |
| 95 | + Returns: |
| 96 | + tuple[Tensor]: A tuple of multi-level predication map, each is a |
| 97 | + 4D-tensor of shape (batch_size, 5+num_classes, height, width). |
| 98 | + """ |
| 99 | + |
| 100 | + return multi_apply( |
| 101 | + self.forward_single, feats, self.multi_level_cls_convs, |
| 102 | + self.multi_level_reg_convs, self.multi_level_conv_cls, |
| 103 | + self.multi_level_conv_reg, self.multi_level_conv_obj, |
| 104 | + self.multi_level_cls_decomps, self.multi_level_reg_decomps) |
0 commit comments