模型 |
监督数据比例 |
Sup Baseline |
Sup Epochs (Iters) |
Sup mAPval 0.5:0.95
| Semi mAPval 0.5:0.95
| Semi Epochs (Iters) |
模型下载 |
配置文件 |
DenseTeacher-FCOS |
5% |
sup_config |
24 (8712) |
21.3 |
30.6 |
240 (87120) |
download |
config |
DenseTeacher-FCOS |
10% |
sup_config |
24 (17424) |
26.3 |
35.1 |
240 (174240) |
download |
config |
DenseTeacher-FCOS(LSJ) |
10% |
sup_config |
24 (17424) |
26.3 |
37.1(LSJ) |
240 (174240) |
download |
config |
DenseTeacher-FCOS |
100%(full) |
sup_config |
24 (175896) |
42.6 |
44.2 |
24 (175896) |
download |
config |
半监督数据集准备
半监督目标检测同时需要有标注数据和无标注数据,且无标注数据量一般远多于有标注数据量。
对于COCO数据集一般有两种常规设置:
(1)抽取部分比例的原始训练集train2017
作为标注数据和无标注数据;
从train2017
中按固定百分比(1%、2%、5%、10%等)抽取,由于抽取方法会对半监督训练的结果影响较大,所以采用五折交叉验证来评估。运行数据集划分制作的脚本如下:
python tools/gen_semi_coco.py
会按照 1%、2%、5%、10% 的监督数据比例来划分train2017
全集,为了交叉验证每一种划分会随机重复5次,生成的半监督标注文件如下:
- 标注数据集标注:
instances_train2017.{fold}@{percent}.json
- 无标注数据集标注:
instances_train2017.{fold}@{percent}-unlabeled.json
其中,fold
表示交叉验证,percent
表示有标注数据的百分比。
注意如果根据txt_file
生成,需要下载COCO_supervision.txt
:
wget https://bj.bcebos.com/v1/paddledet/data/coco/COCO_supervision.txt
(2)使用全量原始训练集train2017
作为有标注数据 和 全量原始无标签图片集unlabeled2017
作为无标注数据;
下载链接
PaddleDetection团队提供了COCO数据集全部的标注文件,请下载并解压存放至对应目录:
# 下载COCO全量数据集图片和标注
# 包括 train2017, val2017, annotations
wget https://bj.bcebos.com/v1/paddledet/data/coco.tar
# 下载PaddleDetection团队整理的COCO部分比例数据的标注文件
wget https://bj.bcebos.com/v1/paddledet/data/coco/semi_annotations.zip
# unlabeled2017是可选,如果不需要训‘full’则无需下载
# 下载COCO全量 unlabeled 无标注数据集
wget https://bj.bcebos.com/v1/paddledet/data/coco/unlabeled2017.zip
wget https://bj.bcebos.com/v1/paddledet/data/coco/image_info_unlabeled2017.zip
# 下载转换完的 unlabeled2017 无标注json文件
wget https://bj.bcebos.com/v1/paddledet/data/coco/instances_unlabeled2017.zip
如果需要用到COCO全量unlabeled无标注数据集,需要将原版的image_info_unlabeled2017.json
进行格式转换,运行以下代码:
COCO unlabeled 标注转换代码:
import json
anns_train = json.load(open('annotations/instances_train2017.json', 'r'))
anns_unlabeled = json.load(open('annotations/image_info_unlabeled2017.json', 'r'))
unlabeled_json = {
'images': anns_unlabeled['images'],
'annotations': [],
'categories': anns_train['categories'],
}
path = 'annotations/instances_unlabeled2017.json'
with open(path, 'w') as f:
json.dump(unlabeled_json, f)
解压后的数据集目录如下:
PaddleDetection
├── dataset
│ ├── coco
│ │ ├── annotations
│ │ │ ├── instances_train2017.json
│ │ │ ├── instances_unlabeled2017.json
│ │ │ ├── instances_val2017.json
│ │ ├── semi_annotations
│ │ │ ├── instances_train2017.1@1.json
│ │ │ ├── instances_train2017.1@1-unlabeled.json
│ │ │ ├── instances_train2017.1@2.json
│ │ │ ├── instances_train2017.1@2-unlabeled.json
│ │ │ ├── instances_train2017.1@5.json
│ │ │ ├── instances_train2017.1@5-unlabeled.json
│ │ │ ├── instances_train2017.1@10.json
│ │ │ ├── instances_train2017.1@10-unlabeled.json
│ │ ├── train2017
│ │ ├── unlabeled2017
│ │ ├── val2017
半监督检测配置
配置半监督检测,需要基于选用的基础检测器的配置文件,如:
_BASE_: [
'../../fcos/fcos_r50_fpn_iou_multiscale_2x_coco.yml',
'../_base_/coco_detection_percent_10.yml',
]
log_iter: 50
snapshot_epoch: 5
epochs: &epochs 240
weights: output/denseteacher_fcos_r50_fpn_coco_semi010/model_final
并依次做出如下几点改动:
训练集配置
首先可以直接引用已经配置好的半监督训练集,如:
_BASE_: [
'../_base_/coco_detection_percent_10.yml',
]
具体来看,构建半监督数据集,需要同时配置监督数据集TrainDataset
和无监督数据集UnsupTrainDataset
的路径,注意必须选用SemiCOCODataSet
类而不是COCODataSet
类,如以下所示:
COCO-train2017部分比例数据集:
# partial labeled COCO, use `SemiCOCODataSet` rather than `COCODataSet`
TrainDataset:
!SemiCOCODataSet
image_dir: train2017
anno_path: semi_annotations/instances_train2017.1@10.json
dataset_dir: dataset/coco
data_fields: ['image', 'gt_bbox', 'gt_class', 'is_crowd']
# partial unlabeled COCO, use `SemiCOCODataSet` rather than `COCODataSet`
UnsupTrainDataset:
!SemiCOCODataSet
image_dir: train2017
anno_path: semi_annotations/instances_train2017.1@10-unlabeled.json
dataset_dir: dataset/coco
data_fields: ['image']
supervised: False
或者 COCO-train2017 full全量数据集:
# full labeled COCO, use `SemiCOCODataSet` rather than `COCODataSet`
TrainDataset:
!SemiCOCODataSet
image_dir: train2017
anno_path: annotations/instances_train2017.json
dataset_dir: dataset/coco
data_fields: ['image', 'gt_bbox', 'gt_class', 'is_crowd']
# full unlabeled COCO, use `SemiCOCODataSet` rather than `COCODataSet`
UnsupTrainDataset:
!SemiCOCODataSet
image_dir: unlabeled2017
anno_path: annotations/instances_unlabeled2017.json
dataset_dir: dataset/coco
data_fields: ['image']
supervised: False
验证集EvalDataset
和测试集TestDataset
的配置不需要更改,且还是采用COCODataSet
类。
预训练配置
### pretrain and warmup config, choose one and coment another
pretrain_weights: https://paddledet.bj.bcebos.com/models/pretrained/ResNet50_cos_pretrained.pdparams
semi_start_iters: 5000
ema_start_iters: 3000
use_warmup: &use_warmup True
注意:
Dense Teacher
原文使用R50-va-caffe
预训练,PaddleDetection中默认使用R50-vb
预训练,如果使用R50-vd
结合SSLD的预训练模型,可进一步显著提升检测精度,同时backbone部分配置也需要做出相应更改,如:
python
pretrain_weights: https://paddledet.bj.bcebos.com/models/pretrained/ResNet50_vd_ssld_v2_pretrained.pdparams
ResNet:
depth: 50
variant: d
norm_type: bn
freeze_at: 0
return_idx: [1, 2, 3]
num_stages: 4
lr_mult_list: [0.05, 0.05, 0.1, 0.15]
全局配置
需要在配置文件中添加如下全局配置,并且注意 DenseTeacher 模型需要使用use_simple_ema: True
而不是use_ema: True
:
### global config
use_simple_ema: True
ema_decay: 0.9996
ssod_method: DenseTeacher
DenseTeacher:
train_cfg:
sup_weight: 1.0
unsup_weight: 1.0
loss_weight: {distill_loss_cls: 4.0, distill_loss_box: 1.0, distill_loss_quality: 1.0}
concat_sup_data: True
suppress: linear
ratio: 0.01
gamma: 2.0
test_cfg:
inference_on: teacher
模型配置
如果没有特殊改动,则直接继承自基础检测器里的模型配置。
以 DenseTeacher
为例,选择 fcos_r50_fpn_iou_multiscale_2x_coco.yml
作为基础检测器进行半监督训练,teacher网络的结构和student网络的结构均为基础检测器的结构,且结构相同。
_BASE_: [
'../../fcos/fcos_r50_fpn_iou_multiscale_2x_coco.yml',
]
数据增强配置
构建半监督训练集的Reader,需要在原先TrainReader
的基础上,新增加weak_aug
,strong_aug
,sup_batch_transforms
和unsup_batch_transforms
,并且需要注意:
- 如果有
NormalizeImage
,需要单独从sample_transforms
中抽出来放在weak_aug
和strong_aug
中;
sample_transforms
为公用的基础数据增强;
- 完整的弱数据增强为
sample_transforms + weak_aug
,完整的强数据增强为sample_transforms + strong_aug
;
如以下所示:
原纯监督模型的TrainReader
:
TrainReader:
sample_transforms:
- Decode: {}
- RandomResize: {target_size: [[640, 1333], [672, 1333], [704, 1333], [736, 1333], [768, 1333], [800, 1333]], keep_ratio: True, interp: 1}
- RandomFlip: {}
- NormalizeImage: {mean: [0.485, 0.456, 0.406], std: [0.229, 0.224, 0.225], is_scale: True}
batch_transforms:
- Permute: {}
- PadBatch: {pad_to_stride: 32}
- Gt2FCOSTarget:
object_sizes_boundary: [64, 128, 256, 512]
center_sampling_radius: 1.5
downsample_ratios: [8, 16, 32, 64, 128]
norm_reg_targets: True
batch_size: 2
shuffle: True
drop_last: True
更改后的半监督TrainReader:
### reader config
SemiTrainReader:
sample_transforms:
- Decode: {}
- RandomResize: {target_size: [[640, 1333], [672, 1333], [704, 1333], [736, 1333], [768, 1333], [800, 1333]], keep_ratio: True, interp: 1}
- RandomFlip: {}
weak_aug:
- NormalizeImage: {mean: [0.485, 0.456, 0.406], std: [0.229, 0.224, 0.225], is_scale: true}
strong_aug:
- StrongAugImage: {transforms: [
RandomColorJitter: {prob: 0.8, brightness: 0.4, contrast: 0.4, saturation: 0.4, hue: 0.1},
RandomErasingCrop: {},
RandomGaussianBlur: {prob: 0.5, sigma: [0.1, 2.0]},
RandomGrayscale: {prob: 0.2},
]}
- NormalizeImage: {mean: [0.485, 0.456, 0.406], std: [0.229, 0.224, 0.225], is_scale: true}
sup_batch_transforms:
- Permute: {}
- PadBatch: {pad_to_stride: 32}
- Gt2FCOSTarget:
object_sizes_boundary: [64, 128, 256, 512]
center_sampling_radius: 1.5
downsample_ratios: [8, 16, 32, 64, 128]
norm_reg_targets: True
unsup_batch_transforms:
- Permute: {}
- PadBatch: {pad_to_stride: 32}
sup_batch_size: 2
unsup_batch_size: 2
shuffle: True
drop_last: True
其他配置
训练epoch数需要和全量数据训练时换算总iter数保持一致,如全量训练24 epoch(换算约为180k个iter),则10%监督数据的半监督训练,总epoch数需要为240 epoch左右(换算约为180k个iter)。示例如下:
### other config
epoch: 240
LearningRate:
base_lr: 0.01
schedulers:
- !PiecewiseDecay
gamma: 0.1
milestones: 240
use_warmup: True
- !LinearWarmup
start_factor: 0.001
steps: 1000
OptimizerBuilder:
optimizer:
momentum: 0.9
type: Momentum
regularizer:
factor: 0.0001
type: L2
clip_grad_by_value: 1.0
使用说明
仅训练时必须使用半监督检测的配置文件去训练,评估、预测、部署也可以按基础检测器的配置文件去执行。
训练
# 单卡训练 (不推荐,需按线性比例相应地调整学习率)
CUDA_VISIBLE_DEVICES=0 python tools/train.py -c configs/semi_det/denseteacher/denseteacher_fcos_r50_fpn_coco_semi010.yml --eval
# 多卡训练
python -m paddle.distributed.launch --log_dir=denseteacher_fcos_semi010/ --gpus 0,1,2,3,4,5,6,7 tools/train.py -c configs/semi_det/denseteacher/denseteacher_fcos_r50_fpn_coco_semi010.yml --eval
评估
CUDA_VISIBLE_DEVICES=0 python tools/eval.py -c configs/semi_det/denseteacher/denseteacher_fcos_r50_fpn_coco_semi010.yml -o weights=output/denseteacher_fcos_r50_fpn_coco_semi010/model_final.pdparams
预测
CUDA_VISIBLE_DEVICES=0 python tools/infer.py -c configs/semi_det/denseteacher/denseteacher_fcos_r50_fpn_coco_semi010.yml -o weights=output/denseteacher_fcos_r50_fpn_coco_semi010/model_final.pdparams --infer_img=demo/000000014439.jpg
部署
部署可以使用半监督检测配置文件,也可以使用基础检测器的配置文件去部署和使用。
# 导出模型
CUDA_VISIBLE_DEVICES=0 python tools/export_model.py -c configs/semi_det/denseteacher/denseteacher_fcos_r50_fpn_coco_semi010.yml -o weights=https://paddledet.bj.bcebos.com/models/denseteacher_fcos_r50_fpn_coco_semi010.pdparams
# 导出权重预测
CUDA_VISIBLE_DEVICES=0 python deploy/python/infer.py --model_dir=output_inference/denseteacher_fcos_r50_fpn_coco_semi010 --image_file=demo/000000014439_640x640.jpg --device=GPU
# 部署测速
CUDA_VISIBLE_DEVICES=0 python deploy/python/infer.py --model_dir=output_inference/denseteacher_fcos_r50_fpn_coco_semi010 --image_file=demo/000000014439_640x640.jpg --device=GPU --run_benchmark=True # --run_mode=trt_fp16
# 导出ONNX
paddle2onnx --model_dir output_inference/denseteacher_fcos_r50_fpn_coco_semi010/ --model_filename model.pdmodel --params_filename model.pdiparams --opset_version 12 --save_file denseteacher_fcos_r50_fpn_coco_semi010.onnx
引用
@article{denseteacher2022,
title={Dense Teacher: Dense Pseudo-Labels for Semi-supervised Object Detection},
author={Hongyu Zhou, Zheng Ge, Songtao Liu, Weixin Mao, Zeming Li, Haiyan Yu, Jian Sun},
journal={arXiv preprint arXiv:2207.02541},
year={2022}
}