更换文档检测模型
This commit is contained in:
437
paddle_detection/configs/semi_det/README.md
Normal file
437
paddle_detection/configs/semi_det/README.md
Normal file
@@ -0,0 +1,437 @@
|
||||
简体中文 | [English](README_en.md)
|
||||
|
||||
# Semi-Supervised Detection (Semi DET) 半监督检测
|
||||
|
||||
## 内容
|
||||
- [简介](#简介)
|
||||
- [模型库](#模型库)
|
||||
- [Baseline](#Baseline)
|
||||
- [DenseTeacher](#DenseTeacher)
|
||||
- [ARSL](#ARSL)
|
||||
- [半监督数据集准备](#半监督数据集准备)
|
||||
- [半监督检测配置](#半监督检测配置)
|
||||
- [训练集配置](#训练集配置)
|
||||
- [预训练配置](#预训练配置)
|
||||
- [全局配置](#全局配置)
|
||||
- [模型配置](#模型配置)
|
||||
- [数据增强配置](#数据增强配置)
|
||||
- [其他配置](#其他配置)
|
||||
- [使用说明](#使用说明)
|
||||
- [训练](#训练)
|
||||
- [评估](#评估)
|
||||
- [预测](#预测)
|
||||
- [部署](#部署)
|
||||
- [引用](#引用)
|
||||
|
||||
## 简介
|
||||
半监督目标检测(Semi DET)是**同时使用有标注数据和无标注数据**进行训练的目标检测，既可以极大地节省标注成本，也可以充分利用无标注数据进一步提高检测精度。PaddleDetection团队提供了[DenseTeacher](denseteacher/)和[ARSL](arsl/)等最前沿的半监督检测算法，用户可以下载使用。
|
||||
|
||||
## 模型库
|
||||
|
||||
### [Baseline](baseline)
|
||||
|
||||
**纯监督数据**模型的训练和模型库，请参照[Baseline](baseline)；
|
||||
|
||||
|
||||
### [DenseTeacher](denseteacher)
|
||||
|
||||
| 模型 | 监督数据比例 | Sup Baseline | Sup Epochs (Iters) | Sup mAP<sup>val<br>0.5:0.95 | Semi mAP<sup>val<br>0.5:0.95 | Semi Epochs (Iters) | 模型下载 | 配置文件 |
|
||||
| :------------: | :---------: | :---------------------: | :---------------------: |:---------------------------: |:----------------------------: | :------------------: |:--------: |:----------: |
|
||||
| DenseTeacher-FCOS | 5% | [sup_config](./baseline/fcos_r50_fpn_2x_coco_sup005.yml) | 24 (8712) | 21.3 | **30.6** | 240 (87120) | [download](https://paddledet.bj.bcebos.com/models/denseteacher_fcos_r50_fpn_coco_semi005.pdparams) | [config](denseteacher/denseteacher_fcos_r50_fpn_coco_semi005.yml) |
|
||||
| DenseTeacher-FCOS | 10% | [sup_config](./baseline/fcos_r50_fpn_2x_coco_sup010.yml) | 24 (17424) | 26.3 | **35.1** | 240 (174240) | [download](https://paddledet.bj.bcebos.com/models/denseteacher_fcos_r50_fpn_coco_semi010.pdparams) | [config](denseteacher/denseteacher_fcos_r50_fpn_coco_semi010.yml) |
|
||||
| DenseTeacher-FCOS(LSJ)| 10% | [sup_config](./baseline/fcos_r50_fpn_2x_coco_sup010.yml) | 24 (17424) | 26.3 | **37.1(LSJ)** | 240 (174240) | [download](https://paddledet.bj.bcebos.com/models/denseteacher_fcos_r50_fpn_coco_semi010_lsj.pdparams) | [config](denseteacher/denseteacher_fcos_r50_fpn_coco_semi010_lsj.yml) |
|
||||
| DenseTeacher-FCOS |100%(full)| [sup_config](./../fcos/fcos_r50_fpn_iou_multiscale_2x_coco.ymll) | 24 (175896) | 42.6 | **44.2** | 24 (175896)| [download](https://paddledet.bj.bcebos.com/models/denseteacher_fcos_r50_fpn_coco_full.pdparams) | [config](denseteacher/denseteacher_fcos_r50_fpn_coco_full.yml) |
|
||||
|
||||
| 模型 | 监督数据比例 | Sup Baseline | Sup Epochs (Iters) | Sup mAP<sup>val<br>0.5:0.95 | Semi mAP<sup>val<br>0.5:0.95 | Semi Epochs (Iters) | 模型下载 | 配置文件 |
|
||||
| :------------: | :---------: | :---------------------: | :---------------------: |:---------------------------: |:----------------------------: | :------------------: |:--------: |:----------: |
|
||||
| DenseTeacher-PPYOLOE+_s | 5% | [sup_config](./baseline/ppyoloe_plus_crn_s_80e_coco_sup005.yml) | 80 (14480) | 32.8 | **34.0** | 200 (36200) | [download](https://paddledet.bj.bcebos.com/models/denseteacher_ppyoloe_plus_crn_s_coco_semi005.pdparams) | [config](denseteacher/denseteacher_ppyoloe_plus_crn_s_coco_semi005.yml) |
|
||||
| DenseTeacher-PPYOLOE+_s | 10% | [sup_config](./baseline/ppyoloe_plus_crn_s_80e_coco_sup010.yml) | 80 (14480) | 35.3 | **37.5** | 200 (36200) | [download](https://paddledet.bj.bcebos.com/models/denseteacher_ppyoloe_plus_crn_s_coco_semi010.pdparams) | [config](denseteacher/denseteacher_ppyoloe_plus_crn_s_coco_semi010.yml) |
|
||||
| DenseTeacher-PPYOLOE+_l | 5% | [sup_config](./baseline/ppyoloe_plus_crn_s_80e_coco_sup005.yml) | 80 (14480) | 42.9 | **45.4** | 200 (36200) | [download](https://paddledet.bj.bcebos.com/models/denseteacher_ppyoloe_plus_crn_l_coco_semi005.pdparams) | [config](denseteacher/denseteacher_ppyoloe_plus_crn_l_coco_semi005.yml) |
|
||||
| DenseTeacher-PPYOLOE+_l | 10% | [sup_config](./baseline/ppyoloe_plus_crn_l_80e_coco_sup010.yml) | 80 (14480) | 45.7 | **47.4** | 200 (36200) | [download](https://paddledet.bj.bcebos.com/models/denseteacher_ppyoloe_plus_crn_l_coco_semi010.pdparams) | [config](denseteacher/denseteacher_ppyoloe_plus_crn_l_coco_semi010.yml) |
|
||||
|
||||
|
||||
### [ARSL](arsl)
|
||||
|
||||
| 模型 | COCO监督数据比例 | Semi mAP<sup>val<br>0.5:0.95 | Semi Epochs (Iters) | 模型下载 | 配置文件 |
|
||||
| :------------: | :---------:|:----------------------------: | :------------------: |:--------: |:----------: |
|
||||
| ARSL-FCOS | 1% | **22.8** | 240 (87120) | [download](https://paddledet.bj.bcebos.com/models/arsl_fcos_r50_fpn_coco_semi001.pdparams) | [config](arsl/arsl_fcos_r50_fpn_coco_semi001.yml) |
|
||||
| ARSL-FCOS | 5% | **33.1** | 240 (174240) | [download](https://paddledet.bj.bcebos.com/models/arsl_fcos_r50_fpn_coco_semi005.pdparams) | [config](arsl/arsl_fcos_r50_fpn_coco_semi005.yml ) |
|
||||
| ARSL-FCOS | 10% | **36.9** | 240 (174240) | [download](https://paddledet.bj.bcebos.com/models/arsl_fcos_r50_fpn_coco_semi010.pdparams) | [config](arsl/arsl_fcos_r50_fpn_coco_semi010.yml ) |
|
||||
| ARSL-FCOS | 10% | **38.5(LSJ)** | 240 (174240) | [download](https://paddledet.bj.bcebos.com/models/arsl_fcos_r50_fpn_coco_semi010_lsj.pdparams) | [config](arsl/arsl_fcos_r50_fpn_coco_semi010_lsj.yml ) |
|
||||
| ARSL-FCOS | full(100%) | **45.1** | 240 (174240) | [download](https://paddledet.bj.bcebos.com/models/arsl_fcos_r50_fpn_coco_full.pdparams) | [config](arsl/arsl_fcos_r50_fpn_coco_full.yml ) |
|
||||
|
||||
|
||||
|
||||
## 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：
|
||||
```bash
|
||||
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`:
|
||||
```shell
|
||||
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:
|
||||
|
||||
```shell
|
||||
# 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:
|
||||
|
||||
<details>
|
||||
<summary> COCO unlabeled 标注转换代码：</summary>
|
||||
|
||||
```python
|
||||
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)
|
||||
```
|
||||
|
||||
</details>
|
||||
|
||||
|
||||
<details open>
|
||||
<summary> 解压后的数据集目录如下：</summary>
|
||||
|
||||
```
|
||||
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
|
||||
```
|
||||
|
||||
</details>
|
||||
|
||||
## 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：
|
||||
|
||||
```python
|
||||
_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：
|
||||
|
||||
```python
|
||||
_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**：
|
||||
|
||||
```python
|
||||
# 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**：
|
||||
|
||||
```python
|
||||
# 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
|
||||
|
||||
```python
|
||||
### pretrain and warmup config, choose one and comment 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](../../../docs/feature_models/SSLD_PRETRAINED_MODEL.md)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`：
|
||||
|
||||
```python
|
||||
### 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**。
|
||||
|
||||
```python
|
||||
_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`：
|
||||
```python
|
||||
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：
|
||||
|
||||
```python
|
||||
### 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：
|
||||
|
||||
```python
|
||||
### 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
|
||||
|
||||
```bash
|
||||
# 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
|
||||
|
||||
```bash
|
||||
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
|
||||
|
||||
```bash
|
||||
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。
|
||||
|
||||
```bash
|
||||
# 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}
|
||||
}
|
||||
```
|
||||
@@ -0,0 +1,31 @@
|
||||
metric: COCO
|
||||
num_classes: 80
|
||||
|
||||
# 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
|
||||
|
||||
EvalDataset:
|
||||
!COCODataSet
|
||||
image_dir: val2017
|
||||
anno_path: annotations/instances_val2017.json
|
||||
dataset_dir: dataset/coco
|
||||
allow_empty: true
|
||||
|
||||
TestDataset:
|
||||
!ImageFolder
|
||||
anno_path: annotations/instances_val2017.json # also support txt (like VOC's label_list.txt)
|
||||
dataset_dir: dataset/coco # if set, anno_path will be 'dataset_dir/anno_path'
|
||||
@@ -0,0 +1,31 @@
|
||||
metric: COCO
|
||||
num_classes: 80
|
||||
|
||||
# partial labeled COCO, use `SemiCOCODataSet` rather than `COCODataSet`
|
||||
TrainDataset:
|
||||
!SemiCOCODataSet
|
||||
image_dir: train2017
|
||||
anno_path: semi_annotations/instances_train2017.1@1.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@1-unlabeled.json
|
||||
dataset_dir: dataset/coco
|
||||
data_fields: ['image']
|
||||
supervised: False
|
||||
|
||||
EvalDataset:
|
||||
!COCODataSet
|
||||
image_dir: val2017
|
||||
anno_path: annotations/instances_val2017.json
|
||||
dataset_dir: dataset/coco
|
||||
allow_empty: true
|
||||
|
||||
TestDataset:
|
||||
!ImageFolder
|
||||
anno_path: annotations/instances_val2017.json # also support txt (like VOC's label_list.txt)
|
||||
dataset_dir: dataset/coco # if set, anno_path will be 'dataset_dir/anno_path'
|
||||
@@ -0,0 +1,31 @@
|
||||
metric: COCO
|
||||
num_classes: 80
|
||||
|
||||
# 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
|
||||
|
||||
EvalDataset:
|
||||
!COCODataSet
|
||||
image_dir: val2017
|
||||
anno_path: annotations/instances_val2017.json
|
||||
dataset_dir: dataset/coco
|
||||
allow_empty: true
|
||||
|
||||
TestDataset:
|
||||
!ImageFolder
|
||||
anno_path: annotations/instances_val2017.json # also support txt (like VOC's label_list.txt)
|
||||
dataset_dir: dataset/coco # if set, anno_path will be 'dataset_dir/anno_path'
|
||||
@@ -0,0 +1,31 @@
|
||||
metric: COCO
|
||||
num_classes: 80
|
||||
|
||||
# partial labeled COCO, use `SemiCOCODataSet` rather than `COCODataSet`
|
||||
TrainDataset:
|
||||
!SemiCOCODataSet
|
||||
image_dir: train2017
|
||||
anno_path: semi_annotations/instances_train2017.1@5.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@5-unlabeled.json
|
||||
dataset_dir: dataset/coco
|
||||
data_fields: ['image']
|
||||
supervised: False
|
||||
|
||||
EvalDataset:
|
||||
!COCODataSet
|
||||
image_dir: val2017
|
||||
anno_path: annotations/instances_val2017.json
|
||||
dataset_dir: dataset/coco
|
||||
allow_empty: true
|
||||
|
||||
TestDataset:
|
||||
!ImageFolder
|
||||
anno_path: annotations/instances_val2017.json # also support txt (like VOC's label_list.txt)
|
||||
dataset_dir: dataset/coco # if set, anno_path will be 'dataset_dir/anno_path'
|
||||
@@ -0,0 +1,31 @@
|
||||
metric: COCO
|
||||
num_classes: 20
|
||||
# before training, change VOC to COCO format by 'python voc2coco.py'
|
||||
# partial labeled COCO, use `SemiCOCODataSet` rather than `COCODataSet`
|
||||
TrainDataset:
|
||||
!SemiCOCODataSet
|
||||
image_dir: VOC2007/JPEGImages
|
||||
anno_path: PseudoAnnotations/VOC2007_trainval.json
|
||||
dataset_dir: dataset/voc/VOCdevkit
|
||||
data_fields: ['image', 'gt_bbox', 'gt_class', 'is_crowd']
|
||||
|
||||
# partial unlabeled COCO, use `SemiCOCODataSet` rather than `COCODataSet`
|
||||
UnsupTrainDataset:
|
||||
!SemiCOCODataSet
|
||||
image_dir: VOC2012/JPEGImages
|
||||
anno_path: PseudoAnnotations/VOC2012_trainval.json
|
||||
dataset_dir: dataset/voc/VOCdevkit
|
||||
data_fields: ['image']
|
||||
supervised: False
|
||||
|
||||
EvalDataset:
|
||||
!COCODataSet
|
||||
image_dir: VOC2007/JPEGImages
|
||||
anno_path: PseudoAnnotations/VOC2007_test.json
|
||||
dataset_dir: dataset/voc/VOCdevkit/
|
||||
allow_empty: true
|
||||
|
||||
TestDataset:
|
||||
!ImageFolder
|
||||
anno_path: PseudoAnnotations/VOC2007_test.json # also support txt (like VOC's label_list.txt)
|
||||
dataset_dir: dataset/voc/VOCdevkit/ # if set, anno_path will be 'dataset_dir/anno_path'
|
||||
213
paddle_detection/configs/semi_det/_base_/voc2coco.py
Normal file
213
paddle_detection/configs/semi_det/_base_/voc2coco.py
Normal file
@@ -0,0 +1,213 @@
|
||||
# convert VOC xml to COCO format json
|
||||
import xml.etree.ElementTree as ET
|
||||
import os
|
||||
import json
|
||||
import argparse
|
||||
|
||||
|
||||
# create and init coco json, img set, and class set
|
||||
def init_json():
|
||||
# create coco json
|
||||
coco = dict()
|
||||
coco['images'] = []
|
||||
coco['type'] = 'instances'
|
||||
coco['annotations'] = []
|
||||
coco['categories'] = []
|
||||
# voc classes
|
||||
voc_class = [
|
||||
'aeroplane', 'bicycle', 'bird', 'boat', 'bottle', 'bus', 'car', 'cat',
|
||||
'chair', 'cow', 'diningtable', 'dog', 'horse', 'motorbike', 'person',
|
||||
'pottedplant', 'sheep', 'sofa', 'train', 'tvmonitor'
|
||||
]
|
||||
# init json categories
|
||||
image_set = set()
|
||||
class_set = dict()
|
||||
for cat_id, cat_name in enumerate(voc_class):
|
||||
cat_item = dict()
|
||||
cat_item['supercategory'] = 'none'
|
||||
cat_item['id'] = cat_id
|
||||
cat_item['name'] = cat_name
|
||||
coco['categories'].append(cat_item)
|
||||
class_set[cat_name] = cat_id
|
||||
return coco, class_set, image_set
|
||||
|
||||
|
||||
def getImgItem(file_name, size, img_id):
|
||||
if file_name is None:
|
||||
raise Exception('Could not find filename tag in xml file.')
|
||||
if size['width'] is None:
|
||||
raise Exception('Could not find width tag in xml file.')
|
||||
if size['height'] is None:
|
||||
raise Exception('Could not find height tag in xml file.')
|
||||
image_item = dict()
|
||||
image_item['id'] = img_id
|
||||
image_item['file_name'] = file_name
|
||||
image_item['width'] = size['width']
|
||||
image_item['height'] = size['height']
|
||||
return image_item
|
||||
|
||||
|
||||
def getAnnoItem(object_name, image_id, ann_id, category_id, bbox):
|
||||
annotation_item = dict()
|
||||
annotation_item['segmentation'] = []
|
||||
seg = []
|
||||
# bbox[] is x,y,w,h
|
||||
# left_top
|
||||
seg.append(bbox[0])
|
||||
seg.append(bbox[1])
|
||||
# left_bottom
|
||||
seg.append(bbox[0])
|
||||
seg.append(bbox[1] + bbox[3])
|
||||
# right_bottom
|
||||
seg.append(bbox[0] + bbox[2])
|
||||
seg.append(bbox[1] + bbox[3])
|
||||
# right_top
|
||||
seg.append(bbox[0] + bbox[2])
|
||||
seg.append(bbox[1])
|
||||
|
||||
annotation_item['segmentation'].append(seg)
|
||||
|
||||
annotation_item['area'] = bbox[2] * bbox[3]
|
||||
annotation_item['iscrowd'] = 0
|
||||
annotation_item['ignore'] = 0
|
||||
annotation_item['image_id'] = image_id
|
||||
annotation_item['bbox'] = bbox
|
||||
annotation_item['category_id'] = category_id
|
||||
annotation_item['id'] = ann_id
|
||||
return annotation_item
|
||||
|
||||
|
||||
def convert_voc_to_coco(txt_path, json_path, xml_path):
|
||||
|
||||
# create and init coco json, img set, and class set
|
||||
coco_json, class_set, image_set = init_json()
|
||||
|
||||
### collect img and ann info into coco json
|
||||
# read img_name in txt, e.g., 000005 for voc2007, 2008_000002 for voc2012
|
||||
img_txt = open(txt_path, 'r')
|
||||
img_line = img_txt.readline().strip()
|
||||
|
||||
# loop xml
|
||||
img_id = 0
|
||||
ann_id = 0
|
||||
while img_line:
|
||||
print('img_id:', img_id)
|
||||
|
||||
# find corresponding xml
|
||||
xml_name = img_line.split('Annotations/', 1)[1]
|
||||
xml_file = os.path.join(xml_path, xml_name)
|
||||
if not os.path.exists(xml_file):
|
||||
print('{} is not exists.'.format(xml_name))
|
||||
img_line = img_txt.readline().strip()
|
||||
continue
|
||||
|
||||
# decode xml
|
||||
tree = ET.parse(xml_file)
|
||||
root = tree.getroot()
|
||||
if root.tag != 'annotation':
|
||||
raise Exception(
|
||||
'xml {} root element should be annotation, rather than {}'.
|
||||
format(xml_name, root.tag))
|
||||
|
||||
# init img and ann info
|
||||
bndbox = dict()
|
||||
size = dict()
|
||||
size['width'] = None
|
||||
size['height'] = None
|
||||
size['depth'] = None
|
||||
|
||||
# filename
|
||||
fileNameNode = root.find('filename')
|
||||
file_name = fileNameNode.text
|
||||
|
||||
# img size
|
||||
sizeNode = root.find('size')
|
||||
if not sizeNode:
|
||||
raise Exception('xml {} structure broken at size tag.'.format(
|
||||
xml_name))
|
||||
for subNode in sizeNode:
|
||||
size[subNode.tag] = int(subNode.text)
|
||||
|
||||
# add img into json
|
||||
if file_name not in image_set:
|
||||
img_id += 1
|
||||
format_img_id = int("%04d" % img_id)
|
||||
# print('line 120. format_img_id:', format_img_id)
|
||||
image_item = getImgItem(file_name, size, img_id)
|
||||
image_set.add(file_name)
|
||||
coco_json['images'].append(image_item)
|
||||
else:
|
||||
raise Exception(' xml {} duplicated image: {}'.format(xml_name,
|
||||
file_name))
|
||||
|
||||
### add objAnn into json
|
||||
objectAnns = root.findall('object')
|
||||
for objectAnn in objectAnns:
|
||||
bndbox['xmin'] = None
|
||||
bndbox['xmax'] = None
|
||||
bndbox['ymin'] = None
|
||||
bndbox['ymax'] = None
|
||||
|
||||
#add obj category
|
||||
object_name = objectAnn.find('name').text
|
||||
if object_name not in class_set:
|
||||
raise Exception('xml {} Unrecognized category: {}'.format(
|
||||
xml_name, object_name))
|
||||
else:
|
||||
current_category_id = class_set[object_name]
|
||||
|
||||
#add obj bbox ann
|
||||
objectBboxNode = objectAnn.find('bndbox')
|
||||
for coordinate in objectBboxNode:
|
||||
if bndbox[coordinate.tag] is not None:
|
||||
raise Exception('xml {} structure corrupted at bndbox tag.'.
|
||||
format(xml_name))
|
||||
bndbox[coordinate.tag] = int(float(coordinate.text))
|
||||
bbox = []
|
||||
# x
|
||||
bbox.append(bndbox['xmin'])
|
||||
# y
|
||||
bbox.append(bndbox['ymin'])
|
||||
# w
|
||||
bbox.append(bndbox['xmax'] - bndbox['xmin'])
|
||||
# h
|
||||
bbox.append(bndbox['ymax'] - bndbox['ymin'])
|
||||
ann_id += 1
|
||||
ann_item = getAnnoItem(object_name, img_id, ann_id,
|
||||
current_category_id, bbox)
|
||||
coco_json['annotations'].append(ann_item)
|
||||
|
||||
img_line = img_txt.readline().strip()
|
||||
|
||||
print('Saving json.')
|
||||
json.dump(coco_json, open(json_path, 'w'))
|
||||
|
||||
|
||||
if __name__ == '__main__':
|
||||
parser = argparse.ArgumentParser()
|
||||
parser.add_argument(
|
||||
'--type', type=str, default='VOC2007_test', help="data type")
|
||||
parser.add_argument(
|
||||
'--base_path',
|
||||
type=str,
|
||||
default='dataset/voc/VOCdevkit',
|
||||
help="base VOC path.")
|
||||
args = parser.parse_args()
|
||||
|
||||
# image info path
|
||||
txt_name = args.type + '.txt'
|
||||
json_name = args.type + '.json'
|
||||
txt_path = os.path.join(args.base_path, 'PseudoAnnotations', txt_name)
|
||||
json_path = os.path.join(args.base_path, 'PseudoAnnotations', json_name)
|
||||
|
||||
# xml path
|
||||
xml_path = os.path.join(args.base_path,
|
||||
args.type.split('_')[0], 'Annotations')
|
||||
|
||||
print('txt_path:', txt_path)
|
||||
print('json_path:', json_path)
|
||||
print('xml_path:', xml_path)
|
||||
|
||||
print('Converting {} to COCO json.'.format(args.type))
|
||||
convert_voc_to_coco(txt_path, json_path, xml_path)
|
||||
print('Finished.')
|
||||
48
paddle_detection/configs/semi_det/arsl/README.md
Normal file
48
paddle_detection/configs/semi_det/arsl/README.md
Normal file
@@ -0,0 +1,48 @@
|
||||
简体中文 | [English](README_en.md)
|
||||
|
||||
# Ambiguity-Resistant Semi-Supervised Learning for Dense Object Detection (ARSL)
|
||||
|
||||
## ARSL-FCOS 模型库
|
||||
|
||||
| 模型 | COCO监督数据比例 | Semi mAP<sup>val<br>0.5:0.95 | Semi Epochs (Iters) | 模型下载 | 配置文件 |
|
||||
| :------------: | :---------:|:----------------------------: | :------------------: |:--------: |:----------: |
|
||||
| ARSL-FCOS | 1% | **22.8** | 240 (87120) | [download](https://paddledet.bj.bcebos.com/models/arsl_fcos_r50_fpn_coco_semi001.pdparams) | [config](./arsl_fcos_r50_fpn_coco_semi001.yml) |
|
||||
| ARSL-FCOS | 5% | **33.1** | 240 (174240) | [download](https://paddledet.bj.bcebos.com/models/arsl_fcos_r50_fpn_coco_semi005.pdparams) | [config](./arsl_fcos_r50_fpn_coco_semi005.yml ) |
|
||||
| ARSL-FCOS | 10% | **36.9** | 240 (174240) | [download](https://paddledet.bj.bcebos.com/models/arsl_fcos_r50_fpn_coco_semi010.pdparams) | [config](./arsl_fcos_r50_fpn_coco_semi010.yml ) |
|
||||
| ARSL-FCOS | 10% | **38.5(LSJ)** | 240 (174240) | [download](https://paddledet.bj.bcebos.com/models/arsl_fcos_r50_fpn_coco_semi010_lsj.pdparams) | [config](./arsl_fcos_r50_fpn_coco_semi010_lsj.yml ) |
|
||||
| ARSL-FCOS | full(100%) | **45.1** | 240 (174240) | [download](https://paddledet.bj.bcebos.com/models/arsl_fcos_r50_fpn_coco_full.pdparams) | [config](./arsl_fcos_r50_fpn_coco_full.yml ) |
|
||||
|
||||
|
||||
|
||||
## 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
|
||||
|
||||
```bash
|
||||
# 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/arsl/arsl_fcos_r50_fpn_coco_semi010.yml --eval
|
||||
|
||||
# 多卡训练
|
||||
python -m paddle.distributed.launch --log_dir=arsl_fcos_r50_fpn_coco_semi010/ --gpus 0,1,2,3,4,5,6,7 tools/train.py -c configs/semi_det/arsl/arsl_fcos_r50_fpn_coco_semi010.yml --eval
|
||||
```
|
||||
|
||||
### 评估
|
||||
|
||||
```bash
|
||||
CUDA_VISIBLE_DEVICES=0 python tools/eval.py -c configs/semi_det/arsl/arsl_fcos_r50_fpn_coco_semi010.yml -o weights=output/arsl_fcos_r50_fpn_coco_semi010/model_final.pdparams
|
||||
```
|
||||
|
||||
### 预测
|
||||
|
||||
```bash
|
||||
CUDA_VISIBLE_DEVICES=0 python tools/infer.py -c configs/semi_det/arsl/arsl_fcos_r50_fpn_coco_semi010.yml -o weights=output/arsl_fcos_r50_fpn_coco_semi010/model_final.pdparams --infer_img=demo/000000014439.jpg
|
||||
```
|
||||
|
||||
|
||||
## 引用
|
||||
|
||||
```
|
||||
|
||||
```
|
||||
@@ -0,0 +1,56 @@
|
||||
architecture: ARSL_FCOS
|
||||
pretrain_weights: https://paddledet.bj.bcebos.com/models/pretrained/ResNet50_cos_pretrained.pdparams
|
||||
|
||||
ARSL_FCOS:
|
||||
backbone: ResNet
|
||||
neck: FPN
|
||||
fcos_head: FCOSHead_ARSL
|
||||
fcos_cr_loss: FCOSLossCR
|
||||
|
||||
|
||||
ResNet:
|
||||
depth: 50
|
||||
norm_type: bn
|
||||
freeze_at: 0
|
||||
return_idx: [1,2,3]
|
||||
num_stages: 4
|
||||
|
||||
FPN:
|
||||
out_channel: 256
|
||||
spatial_scales: [0.125, 0.0625, 0.03125]
|
||||
extra_stage: 2
|
||||
has_extra_convs: true
|
||||
use_c5: false
|
||||
|
||||
FCOSHead_ARSL:
|
||||
fcos_feat:
|
||||
name: FCOSFeat
|
||||
feat_in: 256
|
||||
feat_out: 256
|
||||
num_convs: 4
|
||||
norm_type: "gn"
|
||||
use_dcn: false
|
||||
fpn_stride: [8, 16, 32, 64, 128]
|
||||
prior_prob: 0.01
|
||||
norm_reg_targets: True
|
||||
centerness_on_reg: True
|
||||
fcos_loss:
|
||||
name: FCOSLossMILC
|
||||
loss_alpha: 0.25
|
||||
loss_gamma: 2.0
|
||||
iou_loss_type: "giou"
|
||||
reg_weights: 1.0
|
||||
nms:
|
||||
name: MultiClassNMS
|
||||
nms_top_k: 1000
|
||||
keep_top_k: 100
|
||||
score_threshold: 0.025
|
||||
nms_threshold: 0.6
|
||||
|
||||
|
||||
FCOSLossCR:
|
||||
iou_loss_type: "giou"
|
||||
cls_weight: 2.0
|
||||
reg_weight: 2.0
|
||||
iou_weight: 0.5
|
||||
hard_neg_mining_flag: true
|
||||
@@ -0,0 +1,55 @@
|
||||
worker_num: 2
|
||||
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]
|
||||
num_shift: 0. # default 0.5
|
||||
multiply_strides_reg_targets: False
|
||||
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
|
||||
|
||||
|
||||
EvalReader:
|
||||
sample_transforms:
|
||||
- Decode: {}
|
||||
- Resize: {interp: 2, target_size: [800, 1333], keep_ratio: True}
|
||||
- NormalizeImage: {mean: [0.485, 0.456, 0.406], std: [0.229, 0.224, 0.225], is_scale: true}
|
||||
- Permute: {}
|
||||
batch_transforms:
|
||||
- PadBatch: {pad_to_stride: 32}
|
||||
batch_size: 1
|
||||
|
||||
|
||||
TestReader:
|
||||
sample_transforms:
|
||||
- Decode: {}
|
||||
- Resize: {interp: 2, target_size: [800, 1333], keep_ratio: True}
|
||||
- NormalizeImage: {mean: [0.485, 0.456, 0.406], std: [0.229, 0.224, 0.225], is_scale: true}
|
||||
- Permute: {}
|
||||
batch_transforms:
|
||||
- PadBatch: {pad_to_stride: 32}
|
||||
batch_size: 1
|
||||
@@ -0,0 +1,29 @@
|
||||
epoch: 120 # employ iter to control shedule
|
||||
LearningRate:
|
||||
base_lr: 0.02 # 0.02 for 8*(4+4) batch
|
||||
schedulers:
|
||||
- !PiecewiseDecay
|
||||
gamma: 0.1
|
||||
milestones: [3000] # do not decay lr
|
||||
- !LinearWarmup
|
||||
start_factor: 0.3333333333333333
|
||||
steps: 1000
|
||||
|
||||
max_iter: 360000 # 360k for 32 batch, 720k for 16 batch
|
||||
epoch_iter: 1000 # set epoch_iter for saving checkpoint and eval
|
||||
optimize_rate: 1
|
||||
SEMISUPNET:
|
||||
BBOX_THRESHOLD: 0.5 # # not used
|
||||
TEACHER_UPDATE_ITER: 1
|
||||
BURN_UP_STEP: 30000
|
||||
EMA_KEEP_RATE: 0.9996
|
||||
UNSUP_LOSS_WEIGHT: 1.0 # detailed weights for cls and loc task can be seen in cr_loss
|
||||
PSEUDO_WARM_UP_STEPS: 2000
|
||||
|
||||
OptimizerBuilder:
|
||||
optimizer:
|
||||
momentum: 0.9
|
||||
type: Momentum
|
||||
regularizer:
|
||||
factor: 0.0001
|
||||
type: L2
|
||||
@@ -0,0 +1,30 @@
|
||||
epoch: 30 # employ iter to control shedule
|
||||
LearningRate:
|
||||
base_lr: 0.02 # 0.02 for 8*(4+4) batch
|
||||
schedulers:
|
||||
- !PiecewiseDecay
|
||||
gamma: 0.1
|
||||
milestones: [300] # do not decay lr
|
||||
- !LinearWarmup
|
||||
start_factor: 0.3333333333333333
|
||||
steps: 1000
|
||||
|
||||
max_iter: 90000 # 90k for 32 batch, 180k for 16 batch
|
||||
epoch_iter: 1000 # set epoch_iter for saving checkpoint and eval
|
||||
# update student params according to loss_grad every X iter.
|
||||
optimize_rate: 1
|
||||
SEMISUPNET:
|
||||
BBOX_THRESHOLD: 0.5 # not used
|
||||
TEACHER_UPDATE_ITER: 1
|
||||
BURN_UP_STEP: 9000
|
||||
EMA_KEEP_RATE: 0.9996
|
||||
UNSUP_LOSS_WEIGHT: 1.0 # detailed weights for cls and loc task can be seen in cr_loss
|
||||
PSEUDO_WARM_UP_STEPS: 2000
|
||||
|
||||
OptimizerBuilder:
|
||||
optimizer:
|
||||
momentum: 0.9
|
||||
type: Momentum
|
||||
regularizer:
|
||||
factor: 0.0001
|
||||
type: L2
|
||||
@@ -0,0 +1,12 @@
|
||||
_BASE_: [
|
||||
'../_base_/coco_detection_full.yml',
|
||||
'../../runtime.yml',
|
||||
'_base_/arsl_fcos_r50_fpn.yml',
|
||||
'_base_/optimizer_360k.yml',
|
||||
'_base_/arsl_fcos_reader.yml',
|
||||
]
|
||||
|
||||
weights: output/fcos_r50_fpn_arsl_360k_coco_full/model_final
|
||||
|
||||
#semi detector type
|
||||
ssod_method: ARSL
|
||||
@@ -0,0 +1,12 @@
|
||||
_BASE_: [
|
||||
'../_base_/coco_detection_percent_1.yml',
|
||||
'../../runtime.yml',
|
||||
'_base_/arsl_fcos_r50_fpn.yml',
|
||||
'_base_/optimizer_90k.yml',
|
||||
'_base_/arsl_fcos_reader.yml',
|
||||
]
|
||||
|
||||
weights: output/arsl_fcos_r50_fpn_coco_semi001/model_final
|
||||
|
||||
#semi detector type
|
||||
ssod_method: ARSL
|
||||
@@ -0,0 +1,12 @@
|
||||
_BASE_: [
|
||||
'../_base_/coco_detection_percent_5.yml',
|
||||
'../../runtime.yml',
|
||||
'_base_/arsl_fcos_r50_fpn.yml',
|
||||
'_base_/optimizer_90k.yml',
|
||||
'_base_/arsl_fcos_reader.yml',
|
||||
]
|
||||
|
||||
weights: output/arsl_fcos_r50_fpn_coco_semi005/model_final
|
||||
|
||||
#semi detector type
|
||||
ssod_method: ARSL
|
||||
@@ -0,0 +1,12 @@
|
||||
_BASE_: [
|
||||
'../_base_/coco_detection_percent_10.yml',
|
||||
'../../runtime.yml',
|
||||
'_base_/arsl_fcos_r50_fpn.yml',
|
||||
'_base_/optimizer_360k.yml',
|
||||
'_base_/arsl_fcos_reader.yml',
|
||||
]
|
||||
|
||||
weights: output/arsl_fcos_r50_fpn_coco_semi010/model_final
|
||||
|
||||
#semi detector type
|
||||
ssod_method: ARSL
|
||||
@@ -0,0 +1,47 @@
|
||||
_BASE_: [
|
||||
'../_base_/coco_detection_percent_10.yml',
|
||||
'../../runtime.yml',
|
||||
'_base_/arsl_fcos_r50_fpn.yml',
|
||||
'_base_/optimizer_360k.yml',
|
||||
'_base_/arsl_fcos_reader.yml',
|
||||
]
|
||||
|
||||
weights: output/arsl_fcos_r50_fpn_coco_semi010/model_final
|
||||
|
||||
#semi detector type
|
||||
ssod_method: ARSL
|
||||
|
||||
worker_num: 2
|
||||
SemiTrainReader:
|
||||
sample_transforms:
|
||||
- Decode: {}
|
||||
# large-scale jittering
|
||||
- RandomResize: {target_size: [[400, 1333], [1200, 1333]], keep_ratio: True, interp: 1, random_range: True}
|
||||
- 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]
|
||||
num_shift: 0. # default 0.5
|
||||
multiply_strides_reg_targets: False
|
||||
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
|
||||
93
paddle_detection/configs/semi_det/baseline/README.md
Normal file
93
paddle_detection/configs/semi_det/baseline/README.md
Normal file
@@ -0,0 +1,93 @@
|
||||
# Supervised Baseline 纯监督模型基线
|
||||
|
||||
## COCO数据集模型库
|
||||
|
||||
### [FCOS](../../fcos)
|
||||
|
||||
| 基础模型 | 监督数据比例 | Epochs (Iters) | mAP<sup>val<br>0.5:0.95 | 模型下载 | 配置文件 |
|
||||
| :---------------: | :-------------: | :---------------: |:---------------------: |:--------: | :---------: |
|
||||
| FCOS ResNet50-FPN | 5% | 24 (8712) | 21.3 | [download](https://paddledet.bj.bcebos.com/models/fcos_r50_fpn_2x_coco_sup005.pdparams) | [config](fcos_r50_fpn_2x_coco_sup005.yml) |
|
||||
| FCOS ResNet50-FPN | 10% | 24 (17424) | 26.3 | [download](https://paddledet.bj.bcebos.com/models/fcos_r50_fpn_2x_coco_sup010.pdparams) | [config](fcos_r50_fpn_2x_coco_sup010.yml) |
|
||||
| FCOS ResNet50-FPN | full | 24 (175896) | 42.6 | [download](https://paddledet.bj.bcebos.com/models/fcos_r50_fpn_iou_multiscale_2x_coco.pdparams) | [config](../../fcos/fcos_r50_fpn_iou_multiscale_2x_coco.yml) |
|
||||
|
||||
**注意:**
|
||||
- 以上模型训练默认使用8 GPUs，总batch_size默认为16，默认初始学习率为0.01。如果改动了总batch_size，请按线性比例相应地调整学习率。
|
||||
|
||||
|
||||
### [PP-YOLOE+](../../ppyoloe)
|
||||
|
||||
| 基础模型 | 监督数据比例 | Epochs (Iters) | mAP<sup>val<br>0.5:0.95 | 模型下载 | 配置文件 |
|
||||
| :---------------: | :-------------: | :---------------: | :---------------------: |:--------: | :---------: |
|
||||
| PP-YOLOE+_s | 5% | 80 (7200) | 32.8 | [download](https://paddledet.bj.bcebos.com/models/ppyoloe_plus_crn_s_80e_coco_sup005.pdparams) | [config](ppyoloe_plus_crn_s_80e_coco_sup005.yml) |
|
||||
| PP-YOLOE+_s | 10% | 80 (14480) | 35.3 | [download](https://paddledet.bj.bcebos.com/models/ppyoloe_plus_crn_s_80e_coco_sup010.pdparams) | [config](ppyoloe_plus_crn_s_80e_coco_sup010.yml) |
|
||||
| PP-YOLOE+_s | full | 80 (146560) | 43.7 | [download](https://paddledet.bj.bcebos.com/models/ppyoloe_plus_crn_s_80e_coco.pdparams) | [config](../../ppyoloe/ppyoloe_plus_crn_s_80e_coco.yml) |
|
||||
| PP-YOLOE+_l | 5% | 80 (7200) | 42.9 | [download](https://paddledet.bj.bcebos.com/models/ppyoloe_plus_crn_l_80e_coco_sup005.pdparams) | [config](ppyoloe_plus_crn_l_80e_coco_sup005.yml) |
|
||||
| PP-YOLOE+_l | 10% | 80 (14480) | 45.7 | [download](https://paddledet.bj.bcebos.com/models/ppyoloe_plus_crn_l_80e_coco_sup010.pdparams) | [config](ppyoloe_plus_crn_l_80e_coco_sup010.yml) |
|
||||
| PP-YOLOE+_l | full | 80 (146560) | 49.8 | [download](https://paddledet.bj.bcebos.com/models/ppyoloe_plus_crn_l_80e_coco.pdparams) | [config](../../ppyoloe/ppyoloe_plus_crn_l_80e_coco.yml) |
|
||||
|
||||
**注意:**
|
||||
- 以上模型训练默认使用8 GPUs，总batch_size默认为64，默认初始学习率为0.001。如果改动了总batch_size，请按线性比例相应地调整学习率。
|
||||
|
||||
|
||||
### [Faster R-CNN](../../faster_rcnn)
|
||||
|
||||
| 基础模型 | 监督数据比例 | Epochs (Iters) | mAP<sup>val<br>0.5:0.95 | 模型下载 | 配置文件 |
|
||||
| :---------------: | :-------------: | :---------------: | :---------------------: |:--------: | :---------: |
|
||||
| Faster R-CNN ResNet50-FPN | 5% | 24 (8712) | 20.7 | [download](https://paddledet.bj.bcebos.com/models/faster_rcnn_r50_fpn_2x_coco_sup005.pdparams) | [config](faster_rcnn_r50_fpn_2x_coco_sup005.yml) |
|
||||
| Faster R-CNN ResNet50-FPN | 10% | 24 (17424) | 25.6 | [download](https://paddledet.bj.bcebos.com/models/faster_rcnn_r50_fpn_2x_coco_sup010.pdparams) | [config](faster_rcnn_r50_fpn_2x_coco_sup010.yml) |
|
||||
| Faster R-CNN ResNet50-FPN | full | 24 (175896) | 40.0 | [download](https://paddledet.bj.bcebos.com/models/faster_rcnn_r50_fpn_2x_coco.pdparams) | [config](../../configs/faster_rcnn/faster_rcnn_r50_fpn_2x_coco.yml) |
|
||||
|
||||
**注意:**
|
||||
- 以上模型训练默认使用8 GPUs，总batch_size默认为16，默认初始学习率为0.02。如果改动了总batch_size，请按线性比例相应地调整学习率。
|
||||
|
||||
|
||||
### [RetinaNet](../../retinanet)
|
||||
|
||||
| 基础模型 | 监督数据比例 | Epochs (Iters) | mAP<sup>val<br>0.5:0.95 | 模型下载 | 配置文件 |
|
||||
| :---------------: | :-------------: | :---------------: | :---------------------: |:--------: | :---------: |
|
||||
| RetinaNet ResNet50-FPN | 5% | 24 (8712) | 13.9 | [download](https://paddledet.bj.bcebos.com/models/retinanet_r50_fpn_2x_coco_sup005.pdparams) | [config](retinanet_r50_fpn_2x_coco_sup005.yml) |
|
||||
| RetinaNet ResNet50-FPN | 10% | 24 (17424) | 23.6 | [download](https://paddledet.bj.bcebos.com/models/retinanet_r50_fpn_2x_coco_sup010.pdparams) | [config](retinanet_r50_fpn_2x_coco_sup010.yml) |
|
||||
| RetinaNet ResNet50-FPN | full | 24 (175896) | 39.1 | [download](https://paddledet.bj.bcebos.com/models/retinanet_r50_fpn_2x_coco.pdparams) | [config](../../configs/retinanet/retinanet_r50_fpn_2x_coco.yml) |
|
||||
|
||||
**注意:**
|
||||
- 以上模型训练默认使用8 GPUs，总batch_size默认为16，默认初始学习率为0.01。如果改动了总batch_size，请按线性比例相应地调整学习率。
|
||||
|
||||
|
||||
|
||||
### [RT-DETR](../../rtdetr)
|
||||
|
||||
| 基础模型 | 监督数据比例 | mAP<sup>val<br>0.5:0.95 | 模型下载 | 配置文件 |
|
||||
| :---------------: | :-------------: | :---------------------: |:--------: | :---------: |
|
||||
| RT-DETR ResNet5vd | 5% | 39.1 | [download](https://bj.bcebos.com/v1/paddledet/data/semidet/rtdetr_ssod/baseline/rtdetr_r50vd_6x_coco_sup005.pdparams) | [config](rtdetr_r50vd_6x_coco_sup005.yml) |
|
||||
| RT-DETR ResNet5vd | 10% | 42.3 | [download](https://bj.bcebos.com/v1/paddledet/data/semidet/rtdetr_ssod/baseline/rtdetr_r50vd_6x_coco_sup010.pdparams) | [config](rtdetr_r50vd_6x_coco_sup010.yml) |
|
||||
| RT-DETR ResNet5vd | VOC2007 | 62.7 | [download](https://bj.bcebos.com/v1/paddledet/data/semidet/rtdetr_ssod/baseline/rtdetr_r50vd_6x_voc2007.pdparams) | [config](rtdetr_r50vd_6x_voc2007.yml) |
|
||||
|
||||
**注意:**
|
||||
- RT-DETR模型训练默认使用4 GPUs，总batch_size默认为16，默认初始学习率为0.0001。如果改动了总batch_size，请按线性比例相应地调整学习率。
|
||||
|
||||
### 注意事项
|
||||
- COCO部分监督数据集请参照 [数据集准备](../README.md) 去下载和准备，各个比例的训练集均为**从train2017中抽取部分百分比的子集**，默认使用`fold`号为1的划分子集，`sup010`表示抽取10%的监督数据训练，`sup005`表示抽取5%，`full`表示全部train2017，验证集均为val2017全量；
|
||||
- 抽取部分百分比的监督数据的抽法不同，或使用的`fold`号不同，精度都会因此而有约0.5 mAP之多的差异；
|
||||
- PP-YOLOE+ 使用Objects365预训练，其余模型均使用ImageNet预训练；
|
||||
- 线型比例相应调整学习率，参照公式: **lr<sub>new</sub> = lr<sub>default</sub> * (batch_size<sub>new</sub> * GPU_number<sub>new</sub>) / (batch_size<sub>default</sub> * GPU_number<sub>default</sub>)**。
|
||||
|
||||
|
||||
## 使用教程
|
||||
|
||||
将以下命令写在一个脚本文件里如```run.sh```，一键运行命令为：```sh run.sh```，也可命令行一句句去运行：
|
||||
|
||||
```bash
|
||||
model_type=semi_det/baseline
|
||||
job_name=ppyoloe_plus_crn_s_80e_coco_sup010 # 可修改，如 fcos_r50_fpn_2x_coco_sup010
|
||||
|
||||
config=configs/${model_type}/${job_name}.yml
|
||||
log_dir=log_dir/${job_name}
|
||||
weights=output/${job_name}/model_final.pdparams
|
||||
|
||||
# 1.training
|
||||
# CUDA_VISIBLE_DEVICES=0 python tools/train.py -c ${config}
|
||||
python -m paddle.distributed.launch --log_dir=${log_dir} --gpus 0,1,2,3,4,5,6,7 tools/train.py -c ${config} --eval --amp
|
||||
|
||||
# 2.eval
|
||||
CUDA_VISIBLE_DEVICES=0 python tools/eval.py -c ${config} -o weights=${weights}
|
||||
```
|
||||
@@ -0,0 +1,42 @@
|
||||
_BASE_: [
|
||||
'../../faster_rcnn/faster_rcnn_r50_fpn_2x_coco.yml',
|
||||
]
|
||||
log_iter: 50
|
||||
snapshot_epoch: 2
|
||||
weights: output/faster_rcnn_r50_fpn_2x_coco_sup005/model_final
|
||||
|
||||
|
||||
TrainDataset:
|
||||
!COCODataSet
|
||||
image_dir: train2017
|
||||
anno_path: semi_annotations/instances_train2017.1@5.json
|
||||
dataset_dir: dataset/coco
|
||||
data_fields: ['image', 'gt_bbox', 'gt_class']
|
||||
|
||||
|
||||
worker_num: 2
|
||||
TrainReader:
|
||||
sample_transforms:
|
||||
- Decode: {}
|
||||
- RandomResize: {target_size: [[640, 1333], [672, 1333], [704, 1333], [736, 1333], [768, 1333], [800, 1333]], interp: 2, keep_ratio: True}
|
||||
- RandomFlip: {}
|
||||
- NormalizeImage: {is_scale: true, mean: [0.485,0.456,0.406], std: [0.229, 0.224,0.225]}
|
||||
- Permute: {}
|
||||
batch_transforms:
|
||||
- PadBatch: {pad_to_stride: 32}
|
||||
batch_size: 2
|
||||
shuffle: true
|
||||
drop_last: true
|
||||
collate_batch: false
|
||||
|
||||
|
||||
epoch: 24
|
||||
LearningRate:
|
||||
base_lr: 0.01
|
||||
schedulers:
|
||||
- !PiecewiseDecay
|
||||
gamma: 0.1
|
||||
milestones: [16, 22]
|
||||
- !LinearWarmup
|
||||
start_factor: 0.1
|
||||
epochs: 1
|
||||
@@ -0,0 +1,42 @@
|
||||
_BASE_: [
|
||||
'../../faster_rcnn/faster_rcnn_r50_fpn_2x_coco.yml',
|
||||
]
|
||||
log_iter: 50
|
||||
snapshot_epoch: 2
|
||||
weights: output/faster_rcnn_r50_fpn_2x_coco_sup010/model_final
|
||||
|
||||
|
||||
TrainDataset:
|
||||
!COCODataSet
|
||||
image_dir: train2017
|
||||
anno_path: semi_annotations/instances_train2017.1@10.json
|
||||
dataset_dir: dataset/coco
|
||||
data_fields: ['image', 'gt_bbox', 'gt_class']
|
||||
|
||||
|
||||
worker_num: 2
|
||||
TrainReader:
|
||||
sample_transforms:
|
||||
- Decode: {}
|
||||
- RandomResize: {target_size: [[640, 1333], [672, 1333], [704, 1333], [736, 1333], [768, 1333], [800, 1333]], interp: 2, keep_ratio: True}
|
||||
- RandomFlip: {}
|
||||
- NormalizeImage: {is_scale: true, mean: [0.485,0.456,0.406], std: [0.229, 0.224,0.225]}
|
||||
- Permute: {}
|
||||
batch_transforms:
|
||||
- PadBatch: {pad_to_stride: 32}
|
||||
batch_size: 2
|
||||
shuffle: true
|
||||
drop_last: true
|
||||
collate_batch: false
|
||||
|
||||
|
||||
epoch: 24
|
||||
LearningRate:
|
||||
base_lr: 0.02
|
||||
schedulers:
|
||||
- !PiecewiseDecay
|
||||
gamma: 0.1
|
||||
milestones: [16, 22]
|
||||
- !LinearWarmup
|
||||
start_factor: 0.1
|
||||
epochs: 1
|
||||
@@ -0,0 +1,26 @@
|
||||
_BASE_: [
|
||||
'../../fcos/fcos_r50_fpn_iou_multiscale_2x_coco.yml',
|
||||
]
|
||||
log_iter: 50
|
||||
snapshot_epoch: 2
|
||||
weights: output/fcos_r50_fpn_2x_coco_sup005/model_final
|
||||
|
||||
|
||||
TrainDataset:
|
||||
!COCODataSet
|
||||
image_dir: train2017
|
||||
anno_path: semi_annotations/instances_train2017.1@5.json
|
||||
dataset_dir: dataset/coco
|
||||
data_fields: ['image', 'gt_bbox', 'gt_class']
|
||||
|
||||
|
||||
epoch: 24
|
||||
LearningRate:
|
||||
base_lr: 0.01
|
||||
schedulers:
|
||||
- !PiecewiseDecay
|
||||
gamma: 0.1
|
||||
milestones: [16, 22]
|
||||
- !LinearWarmup
|
||||
start_factor: 0.001
|
||||
epochs: 1
|
||||
@@ -0,0 +1,26 @@
|
||||
_BASE_: [
|
||||
'../../fcos/fcos_r50_fpn_iou_multiscale_2x_coco.yml',
|
||||
]
|
||||
log_iter: 50
|
||||
snapshot_epoch: 2
|
||||
weights: output/fcos_r50_fpn_2x_coco_sup010/model_final
|
||||
|
||||
|
||||
TrainDataset:
|
||||
!COCODataSet
|
||||
image_dir: train2017
|
||||
anno_path: semi_annotations/instances_train2017.1@10.json
|
||||
dataset_dir: dataset/coco
|
||||
data_fields: ['image', 'gt_bbox', 'gt_class']
|
||||
|
||||
|
||||
epoch: 24
|
||||
LearningRate:
|
||||
base_lr: 0.01
|
||||
schedulers:
|
||||
- !PiecewiseDecay
|
||||
gamma: 0.1
|
||||
milestones: [16, 22]
|
||||
- !LinearWarmup
|
||||
start_factor: 0.001
|
||||
epochs: 1
|
||||
@@ -0,0 +1,29 @@
|
||||
_BASE_: [
|
||||
'../../ppyoloe/ppyoloe_plus_crn_l_80e_coco.yml',
|
||||
]
|
||||
log_iter: 50
|
||||
snapshot_epoch: 5
|
||||
weights: output/ppyoloe_plus_crn_l_80e_coco_sup005/model_final
|
||||
|
||||
pretrain_weights: https://bj.bcebos.com/v1/paddledet/models/pretrained/ppyoloe_crn_l_obj365_pretrained.pdparams
|
||||
depth_mult: 1.0
|
||||
width_mult: 1.0
|
||||
|
||||
|
||||
TrainDataset:
|
||||
!COCODataSet
|
||||
image_dir: train2017
|
||||
anno_path: semi_annotations/instances_train2017.1@5.json
|
||||
dataset_dir: dataset/coco
|
||||
data_fields: ['image', 'gt_bbox', 'gt_class']
|
||||
|
||||
|
||||
epoch: 80
|
||||
LearningRate:
|
||||
base_lr: 0.001
|
||||
schedulers:
|
||||
- !CosineDecay
|
||||
max_epochs: 96
|
||||
- !LinearWarmup
|
||||
start_factor: 0.
|
||||
epochs: 5
|
||||
@@ -0,0 +1,29 @@
|
||||
_BASE_: [
|
||||
'../../ppyoloe/ppyoloe_plus_crn_l_80e_coco.yml',
|
||||
]
|
||||
log_iter: 50
|
||||
snapshot_epoch: 5
|
||||
weights: output/ppyoloe_plus_crn_l_80e_coco_sup010/model_final
|
||||
|
||||
pretrain_weights: https://bj.bcebos.com/v1/paddledet/models/pretrained/ppyoloe_crn_l_obj365_pretrained.pdparams
|
||||
depth_mult: 1.0
|
||||
width_mult: 1.0
|
||||
|
||||
|
||||
TrainDataset:
|
||||
!COCODataSet
|
||||
image_dir: train2017
|
||||
anno_path: semi_annotations/instances_train2017.1@10.json
|
||||
dataset_dir: dataset/coco
|
||||
data_fields: ['image', 'gt_bbox', 'gt_class']
|
||||
|
||||
|
||||
epoch: 80
|
||||
LearningRate:
|
||||
base_lr: 0.001
|
||||
schedulers:
|
||||
- !CosineDecay
|
||||
max_epochs: 96
|
||||
- !LinearWarmup
|
||||
start_factor: 0.
|
||||
epochs: 5
|
||||
@@ -0,0 +1,29 @@
|
||||
_BASE_: [
|
||||
'../../ppyoloe/ppyoloe_plus_crn_s_80e_coco.yml',
|
||||
]
|
||||
log_iter: 50
|
||||
snapshot_epoch: 5
|
||||
weights: output/ppyoloe_plus_crn_s_80e_coco_sup005/model_final
|
||||
|
||||
pretrain_weights: https://bj.bcebos.com/v1/paddledet/models/pretrained/ppyoloe_crn_s_obj365_pretrained.pdparams
|
||||
depth_mult: 0.33
|
||||
width_mult: 0.50
|
||||
|
||||
|
||||
TrainDataset:
|
||||
!COCODataSet
|
||||
image_dir: train2017
|
||||
anno_path: semi_annotations/instances_train2017.1@5.json
|
||||
dataset_dir: dataset/coco
|
||||
data_fields: ['image', 'gt_bbox', 'gt_class']
|
||||
|
||||
|
||||
epoch: 80
|
||||
LearningRate:
|
||||
base_lr: 0.001
|
||||
schedulers:
|
||||
- !CosineDecay
|
||||
max_epochs: 96
|
||||
- !LinearWarmup
|
||||
start_factor: 0.
|
||||
epochs: 5
|
||||
@@ -0,0 +1,29 @@
|
||||
_BASE_: [
|
||||
'../../ppyoloe/ppyoloe_plus_crn_s_80e_coco.yml',
|
||||
]
|
||||
log_iter: 50
|
||||
snapshot_epoch: 5
|
||||
weights: output/ppyoloe_plus_crn_s_80e_coco_sup010/model_final
|
||||
|
||||
pretrain_weights: https://bj.bcebos.com/v1/paddledet/models/pretrained/ppyoloe_crn_s_obj365_pretrained.pdparams
|
||||
depth_mult: 0.33
|
||||
width_mult: 0.50
|
||||
|
||||
|
||||
TrainDataset:
|
||||
!COCODataSet
|
||||
image_dir: train2017
|
||||
anno_path: semi_annotations/instances_train2017.1@10.json
|
||||
dataset_dir: dataset/coco
|
||||
data_fields: ['image', 'gt_bbox', 'gt_class']
|
||||
|
||||
|
||||
epoch: 80
|
||||
LearningRate:
|
||||
base_lr: 0.001
|
||||
schedulers:
|
||||
- !CosineDecay
|
||||
max_epochs: 96
|
||||
- !LinearWarmup
|
||||
start_factor: 0.
|
||||
epochs: 5
|
||||
@@ -0,0 +1,26 @@
|
||||
_BASE_: [
|
||||
'../../retinanet/retinanet_r50_fpn_2x_coco.yml',
|
||||
]
|
||||
log_iter: 50
|
||||
snapshot_epoch: 2
|
||||
weights: output/retinanet_r50_fpn_2x_coco_sup005/model_final
|
||||
|
||||
|
||||
TrainDataset:
|
||||
!COCODataSet
|
||||
image_dir: train2017
|
||||
anno_path: semi_annotations/instances_train2017.1@5.json
|
||||
dataset_dir: dataset/coco
|
||||
data_fields: ['image', 'gt_bbox', 'gt_class']
|
||||
|
||||
|
||||
epoch: 24
|
||||
LearningRate:
|
||||
base_lr: 0.01
|
||||
schedulers:
|
||||
- !PiecewiseDecay
|
||||
gamma: 0.1
|
||||
milestones: [16, 22]
|
||||
- !LinearWarmup
|
||||
start_factor: 0.001
|
||||
epochs: 1
|
||||
@@ -0,0 +1,26 @@
|
||||
_BASE_: [
|
||||
'../../retinanet/retinanet_r50_fpn_2x_coco.yml',
|
||||
]
|
||||
log_iter: 50
|
||||
snapshot_epoch: 2
|
||||
weights: output/retinanet_r50_fpn_2x_coco_sup010/model_final
|
||||
|
||||
|
||||
TrainDataset:
|
||||
!COCODataSet
|
||||
image_dir: train2017
|
||||
anno_path: semi_annotations/instances_train2017.1@10.json
|
||||
dataset_dir: dataset/coco
|
||||
data_fields: ['image', 'gt_bbox', 'gt_class']
|
||||
|
||||
|
||||
epoch: 24
|
||||
LearningRate:
|
||||
base_lr: 0.01
|
||||
schedulers:
|
||||
- !PiecewiseDecay
|
||||
gamma: 0.1
|
||||
milestones: [16, 22]
|
||||
- !LinearWarmup
|
||||
start_factor: 0.001
|
||||
epochs: 1
|
||||
@@ -0,0 +1,35 @@
|
||||
_BASE_: [
|
||||
'../../rtdetr/rtdetr_r50vd_6x_coco.yml',
|
||||
]
|
||||
log_iter: 50
|
||||
snapshot_epoch: 2
|
||||
weights: output/rtdetr_r50vd_6x_coco/model_final
|
||||
|
||||
pretrain_weights: https://paddledet.bj.bcebos.com/models/pretrained/ResNet50_vd_ssld_v2_pretrained.pdparams
|
||||
TrainDataset:
|
||||
!COCODataSet
|
||||
image_dir: train2017
|
||||
anno_path: semi_annotations/instances_train2017.1@5.json
|
||||
dataset_dir: dataset/coco
|
||||
data_fields: ['image', 'gt_bbox', 'gt_class']
|
||||
|
||||
|
||||
worker_num: 4
|
||||
TrainReader:
|
||||
sample_transforms:
|
||||
- Decode: {}
|
||||
- RandomDistort: {prob: 0.8}
|
||||
- RandomExpand: {fill_value: [123.675, 116.28, 103.53]}
|
||||
- RandomCrop: {prob: 0.8}
|
||||
- RandomFlip: {}
|
||||
batch_transforms:
|
||||
- BatchRandomResize: {target_size: [480, 512, 544, 576, 608, 640, 640, 640, 672, 704, 736, 768, 800], random_size: True, random_interp: True, keep_ratio: False}
|
||||
- NormalizeImage: {mean: [0., 0., 0.], std: [1., 1., 1.], norm_type: none}
|
||||
- NormalizeBox: {}
|
||||
- BboxXYXY2XYWH: {}
|
||||
- Permute: {}
|
||||
batch_size: 4
|
||||
shuffle: true
|
||||
drop_last: true
|
||||
collate_batch: false
|
||||
use_shared_memory: false
|
||||
@@ -0,0 +1,35 @@
|
||||
_BASE_: [
|
||||
'../../rtdetr/rtdetr_r50vd_6x_coco.yml',
|
||||
]
|
||||
log_iter: 50
|
||||
snapshot_epoch: 2
|
||||
weights: output/rtdetr_r50vd_6x_coco/model_final
|
||||
|
||||
pretrain_weights: https://paddledet.bj.bcebos.com/models/pretrained/ResNet50_vd_ssld_v2_pretrained.pdparams
|
||||
TrainDataset:
|
||||
!COCODataSet
|
||||
image_dir: train2017
|
||||
anno_path: semi_annotations/instances_train2017.1@5.json
|
||||
dataset_dir: dataset/coco
|
||||
data_fields: ['image', 'gt_bbox', 'gt_class']
|
||||
|
||||
|
||||
worker_num: 4
|
||||
TrainReader:
|
||||
sample_transforms:
|
||||
- Decode: {}
|
||||
- RandomDistort: {prob: 0.8}
|
||||
- RandomExpand: {fill_value: [123.675, 116.28, 103.53]}
|
||||
- RandomCrop: {prob: 0.8}
|
||||
- RandomFlip: {}
|
||||
batch_transforms:
|
||||
- BatchRandomResize: {target_size: [480, 512, 544, 576, 608, 640, 640, 640, 672, 704, 736, 768, 800], random_size: True, random_interp: True, keep_ratio: False}
|
||||
- NormalizeImage: {mean: [0., 0., 0.], std: [1., 1., 1.], norm_type: none}
|
||||
- NormalizeBox: {}
|
||||
- BboxXYXY2XYWH: {}
|
||||
- Permute: {}
|
||||
batch_size: 4
|
||||
shuffle: true
|
||||
drop_last: true
|
||||
collate_batch: false
|
||||
use_shared_memory: false
|
||||
101
paddle_detection/configs/semi_det/denseteacher/README.md
Normal file
101
paddle_detection/configs/semi_det/denseteacher/README.md
Normal file
@@ -0,0 +1,101 @@
|
||||
简体中文 | [English](README_en.md)
|
||||
|
||||
# Dense Teacher: Dense Pseudo-Labels for Semi-supervised Object Detection
|
||||
|
||||
## FCOS模型库
|
||||
|
||||
| 模型 | 监督数据比例 | Sup Baseline | Sup Epochs (Iters) | Sup mAP<sup>val<br>0.5:0.95 | Semi mAP<sup>val<br>0.5:0.95 | Semi Epochs (Iters) | 模型下载 | 配置文件 |
|
||||
| :------------: | :---------: | :---------------------: | :---------------------: |:---------------------------: |:----------------------------: | :------------------: |:--------: |:----------: |
|
||||
| DenseTeacher-FCOS | 5% | [sup_config](../baseline/fcos_r50_fpn_2x_coco_sup005.yml) | 24 (8712) | 21.3 | **30.6** | 240 (87120) | [download](https://paddledet.bj.bcebos.com/models/denseteacher_fcos_r50_fpn_coco_semi005.pdparams) | [config](./denseteacher_fcos_r50_fpn_coco_semi005.yml) |
|
||||
| DenseTeacher-FCOS | 10% | [sup_config](../baseline/fcos_r50_fpn_2x_coco_sup010.yml) | 24 (17424) | 26.3 | **35.1** | 240 (174240) | [download](https://paddledet.bj.bcebos.com/models/denseteacher_fcos_r50_fpn_coco_semi010.pdparams) | [config](./denseteacher_fcos_r50_fpn_coco_semi010.yml) |
|
||||
| DenseTeacher-FCOS(LSJ)| 10% | [sup_config](../baseline/fcos_r50_fpn_2x_coco_sup010.yml) | 24 (17424) | 26.3 | **37.1(LSJ)** | 240 (174240) | [download](https://paddledet.bj.bcebos.com/models/denseteacher_fcos_r50_fpn_coco_semi010_lsj.pdparams) | [config](./denseteacher_fcos_r50_fpn_coco_semi010_lsj.yml) |
|
||||
| DenseTeacher-FCOS |100%(full)| [sup_config](../../fcos/fcos_r50_fpn_iou_multiscale_2x_coco.ymll) | 24 (175896) | 42.6 | **44.2** | 24 (175896)| [download](https://paddledet.bj.bcebos.com/models/denseteacher_fcos_r50_fpn_coco_full.pdparams) | [config](./denseteacher_fcos_r50_fpn_coco_full.yml) |
|
||||
|
||||
|
||||
**注意:**
|
||||
- 以上模型训练默认使用8 GPUs，监督数据总batch_size默认为16，无监督数据总batch_size默认也为16，默认初始学习率为0.01。如果改动了总batch_size，请按线性比例相应地调整学习率；
|
||||
- **监督数据比例**是指使用的有标签COCO数据集占 COCO train2017 全量训练集的百分比，使用的无标签COCO数据集一般也是相同比例，但具体图片和有标签数据的图片不重合；
|
||||
- `Semi Epochs (Iters)`表示**半监督训练**的模型的 Epochs (Iters)，如果使用**自定义数据集**，需自行根据Iters换算到对应的Epochs调整，最好保证总Iters 和COCO数据集的设置较为接近；
|
||||
- `Sup mAP`是**只使用有监督数据训练**的模型的精度，请参照**基础检测器的配置文件** 和 [baseline](../baseline)；
|
||||
- `Semi mAP`是**半监督训练**的模型的精度，模型下载和配置文件的链接均为**半监督模型**；
|
||||
- `LSJ`表示 **large-scale jittering**，表示使用更大范围的多尺度训练，可进一步提升精度，但训练速度也会变慢；
|
||||
- 半监督检测的配置讲解，请参照[文档](../README.md/#半监督检测配置)；
|
||||
- `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](../../../docs/feature_models/SSLD_PRETRAINED_MODEL.md)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]
|
||||
```
|
||||
|
||||
|
||||
## PPYOLOE+ 模型库
|
||||
|
||||
| 模型 | 监督数据比例 | Sup Baseline | Sup Epochs (Iters) | Sup mAP<sup>val<br>0.5:0.95 | Semi mAP<sup>val<br>0.5:0.95 | Semi Epochs (Iters) | 模型下载 | 配置文件 |
|
||||
| :------------: | :---------: | :---------------------: | :---------------------: |:---------------------------: |:----------------------------: | :------------------: |:--------: |:----------: |
|
||||
| DenseTeacher-PPYOLOE+_s | 5% | [sup_config](../baseline/ppyoloe_plus_crn_s_80e_coco_sup005.yml) | 80 (14480) | 32.8 | **34.0** | 200 (36200) | [download](https://paddledet.bj.bcebos.com/models/denseteacher_ppyoloe_plus_crn_s_coco_semi005.pdparams) | [config](./denseteacher_ppyoloe_plus_crn_s_coco_semi005.yml) |
|
||||
| DenseTeacher-PPYOLOE+_s | 10% | [sup_config](../baseline/ppyoloe_plus_crn_s_80e_coco_sup010.yml) | 80 (14480) | 35.3 | **37.5** | 200 (36200) | [download](https://paddledet.bj.bcebos.com/models/denseteacher_ppyoloe_plus_crn_s_coco_semi010.pdparams) | [config](./denseteacher_ppyoloe_plus_crn_s_coco_semi010.yml) |
|
||||
| DenseTeacher-PPYOLOE+_l | 5% | [sup_config](../baseline/ppyoloe_plus_crn_s_80e_coco_sup005.yml) | 80 (14480) | 42.9 | **45.4** | 200 (36200) | [download](https://paddledet.bj.bcebos.com/models/denseteacher_ppyoloe_plus_crn_l_coco_semi005.pdparams) | [config](./denseteacher_ppyoloe_plus_crn_l_coco_semi005.yml) |
|
||||
| DenseTeacher-PPYOLOE+_l | 10% | [sup_config](../baseline/ppyoloe_plus_crn_l_80e_coco_sup010.yml) | 80 (14480) | 45.7 | **47.4** | 200 (36200) | [download](https://paddledet.bj.bcebos.com/models/denseteacher_ppyoloe_plus_crn_l_coco_semi010.pdparams) | [config](./denseteacher_ppyoloe_plus_crn_l_coco_semi010.yml) |
|
||||
|
||||
|
||||
## 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
|
||||
|
||||
```bash
|
||||
# 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
|
||||
|
||||
```bash
|
||||
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
|
||||
|
||||
```bash
|
||||
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。
|
||||
|
||||
```bash
|
||||
# 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}
|
||||
}
|
||||
```
|
||||
@@ -0,0 +1,166 @@
|
||||
_BASE_: [
|
||||
'denseteacher_fcos_r50_fpn_coco_semi010.yml',
|
||||
'../_base_/coco_detection_full.yml',
|
||||
]
|
||||
log_iter: 100
|
||||
snapshot_epoch: 2
|
||||
epochs: &epochs 24
|
||||
weights: output/denseteacher_fcos_r50_fpn_coco_full/model_final
|
||||
|
||||
|
||||
### pretrain and warmup config, choose one and comment another
|
||||
# pretrain_weights: https://bj.bcebos.com/v1/paddledet/models/fcos_r50_fpn_iou_multiscale_2x_coco.pdparams # mAP=42.6
|
||||
# semi_start_iters: 0
|
||||
# ema_start_iters: 0
|
||||
# use_warmup: &use_warmup False
|
||||
|
||||
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
|
||||
|
||||
|
||||
### 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: 2.0, 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
|
||||
|
||||
|
||||
### reader config
|
||||
worker_num: 2
|
||||
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]
|
||||
num_shift: 0.5
|
||||
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
|
||||
|
||||
EvalReader:
|
||||
sample_transforms:
|
||||
- Decode: {}
|
||||
- Resize: {target_size: [800, 1333], keep_ratio: True, interp: 1}
|
||||
- NormalizeImage: {mean: [0.485, 0.456, 0.406], std: [0.229, 0.224, 0.225], is_scale: True}
|
||||
- Permute: {}
|
||||
batch_transforms:
|
||||
- PadBatch: {pad_to_stride: 32}
|
||||
batch_size: 1
|
||||
|
||||
TestReader:
|
||||
sample_transforms:
|
||||
- Decode: {}
|
||||
- Resize: {target_size: [800, 1333], keep_ratio: True, interp: 1}
|
||||
- NormalizeImage: {mean: [0.485, 0.456, 0.406], std: [0.229, 0.224, 0.225], is_scale: True}
|
||||
- Permute: {}
|
||||
batch_transforms:
|
||||
- PadBatch: {pad_to_stride: 32}
|
||||
batch_size: 1
|
||||
fuse_normalize: True
|
||||
|
||||
|
||||
### model config
|
||||
architecture: FCOS
|
||||
FCOS:
|
||||
backbone: ResNet
|
||||
neck: FPN
|
||||
fcos_head: FCOSHead
|
||||
|
||||
ResNet:
|
||||
depth: 50
|
||||
variant: 'b'
|
||||
norm_type: bn
|
||||
freeze_at: 0 # res2
|
||||
return_idx: [1, 2, 3]
|
||||
num_stages: 4
|
||||
|
||||
FPN:
|
||||
out_channel: 256
|
||||
spatial_scales: [0.125, 0.0625, 0.03125]
|
||||
extra_stage: 2
|
||||
has_extra_convs: True
|
||||
use_c5: False
|
||||
|
||||
FCOSHead:
|
||||
fcos_feat:
|
||||
name: FCOSFeat
|
||||
feat_in: 256
|
||||
feat_out: 256
|
||||
num_convs: 4
|
||||
norm_type: "gn"
|
||||
use_dcn: False
|
||||
fpn_stride: [8, 16, 32, 64, 128]
|
||||
prior_prob: 0.01
|
||||
norm_reg_targets: True
|
||||
centerness_on_reg: True
|
||||
num_shift: 0.5
|
||||
fcos_loss:
|
||||
name: FCOSLoss
|
||||
loss_alpha: 0.25
|
||||
loss_gamma: 2.0
|
||||
iou_loss_type: "giou"
|
||||
reg_weights: 1.0
|
||||
quality: "iou"
|
||||
nms:
|
||||
name: MultiClassNMS
|
||||
nms_top_k: 1000
|
||||
keep_top_k: 100
|
||||
score_threshold: 0.025
|
||||
nms_threshold: 0.6
|
||||
|
||||
|
||||
### other config
|
||||
epoch: *epochs
|
||||
LearningRate:
|
||||
base_lr: 0.01
|
||||
schedulers:
|
||||
- !PiecewiseDecay
|
||||
gamma: 0.1
|
||||
milestones: [*epochs]
|
||||
use_warmup: *use_warmup
|
||||
- !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
|
||||
@@ -0,0 +1,164 @@
|
||||
_BASE_: [
|
||||
'../../fcos/fcos_r50_fpn_iou_multiscale_2x_coco.yml',
|
||||
'../_base_/coco_detection_percent_5.yml',
|
||||
]
|
||||
log_iter: 20
|
||||
snapshot_epoch: 5
|
||||
epochs: &epochs 240 # 480 will be better
|
||||
weights: output/denseteacher_fcos_r50_fpn_coco_semi005/model_final
|
||||
|
||||
|
||||
### pretrain and warmup config, choose one and comment another
|
||||
# pretrain_weights: https://bj.bcebos.com/v1/paddledet/models/fcos_r50_fpn_2x_coco_sup005.pdparams # mAP=21.3
|
||||
# semi_start_iters: 0
|
||||
# ema_start_iters: 0
|
||||
# use_warmup: &use_warmup False
|
||||
|
||||
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
|
||||
|
||||
|
||||
### 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
|
||||
|
||||
|
||||
### reader config
|
||||
worker_num: 2
|
||||
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
|
||||
|
||||
EvalReader:
|
||||
sample_transforms:
|
||||
- Decode: {}
|
||||
- Resize: {target_size: [800, 1333], keep_ratio: True, interp: 1}
|
||||
- NormalizeImage: {mean: [0.485, 0.456, 0.406], std: [0.229, 0.224, 0.225], is_scale: True}
|
||||
- Permute: {}
|
||||
batch_transforms:
|
||||
- PadBatch: {pad_to_stride: 32}
|
||||
batch_size: 1
|
||||
|
||||
TestReader:
|
||||
sample_transforms:
|
||||
- Decode: {}
|
||||
- Resize: {target_size: [800, 1333], keep_ratio: True, interp: 1}
|
||||
- NormalizeImage: {mean: [0.485, 0.456, 0.406], std: [0.229, 0.224, 0.225], is_scale: True}
|
||||
- Permute: {}
|
||||
batch_transforms:
|
||||
- PadBatch: {pad_to_stride: 32}
|
||||
batch_size: 1
|
||||
fuse_normalize: True
|
||||
|
||||
|
||||
### model config
|
||||
architecture: FCOS
|
||||
FCOS:
|
||||
backbone: ResNet
|
||||
neck: FPN
|
||||
fcos_head: FCOSHead
|
||||
|
||||
ResNet:
|
||||
depth: 50
|
||||
variant: 'b'
|
||||
norm_type: bn
|
||||
freeze_at: 0 # res2
|
||||
return_idx: [1, 2, 3]
|
||||
num_stages: 4
|
||||
|
||||
FPN:
|
||||
out_channel: 256
|
||||
spatial_scales: [0.125, 0.0625, 0.03125]
|
||||
extra_stage: 2
|
||||
has_extra_convs: True
|
||||
use_c5: False
|
||||
|
||||
FCOSHead:
|
||||
fcos_feat:
|
||||
name: FCOSFeat
|
||||
feat_in: 256
|
||||
feat_out: 256
|
||||
num_convs: 4
|
||||
norm_type: "gn"
|
||||
use_dcn: False
|
||||
fpn_stride: [8, 16, 32, 64, 128]
|
||||
prior_prob: 0.01
|
||||
norm_reg_targets: True
|
||||
centerness_on_reg: True
|
||||
fcos_loss:
|
||||
name: FCOSLoss
|
||||
loss_alpha: 0.25
|
||||
loss_gamma: 2.0
|
||||
iou_loss_type: "giou"
|
||||
reg_weights: 1.0
|
||||
quality: "iou"
|
||||
nms:
|
||||
name: MultiClassNMS
|
||||
nms_top_k: 1000
|
||||
keep_top_k: 100
|
||||
score_threshold: 0.025
|
||||
nms_threshold: 0.6
|
||||
|
||||
|
||||
### other config
|
||||
epoch: *epochs
|
||||
LearningRate:
|
||||
base_lr: 0.01
|
||||
schedulers:
|
||||
- !PiecewiseDecay
|
||||
gamma: 0.1
|
||||
milestones: [*epochs]
|
||||
use_warmup: *use_warmup
|
||||
- !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
|
||||
@@ -0,0 +1,169 @@
|
||||
_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
|
||||
|
||||
|
||||
### pretrain and warmup config, choose one and comment another
|
||||
# pretrain_weights: https://bj.bcebos.com/v1/paddledet/models/fcos_r50_fpn_2x_coco_sup010.pdparams # mAP=26.3
|
||||
# semi_start_iters: 0
|
||||
# ema_start_iters: 0
|
||||
# use_warmup: &use_warmup False
|
||||
|
||||
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
|
||||
|
||||
|
||||
### 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
|
||||
|
||||
|
||||
### reader config
|
||||
worker_num: 2
|
||||
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]
|
||||
num_shift: 0. # default 0.5
|
||||
multiply_strides_reg_targets: False
|
||||
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
|
||||
|
||||
EvalReader:
|
||||
sample_transforms:
|
||||
- Decode: {}
|
||||
- Resize: {target_size: [800, 1333], keep_ratio: True, interp: 1}
|
||||
- NormalizeImage: {mean: [0.485, 0.456, 0.406], std: [0.229, 0.224, 0.225], is_scale: True}
|
||||
- Permute: {}
|
||||
batch_transforms:
|
||||
- PadBatch: {pad_to_stride: 32}
|
||||
batch_size: 1
|
||||
|
||||
TestReader:
|
||||
sample_transforms:
|
||||
- Decode: {}
|
||||
- Resize: {target_size: [800, 1333], keep_ratio: True, interp: 1}
|
||||
- NormalizeImage: {mean: [0.485, 0.456, 0.406], std: [0.229, 0.224, 0.225], is_scale: True}
|
||||
- Permute: {}
|
||||
batch_transforms:
|
||||
- PadBatch: {pad_to_stride: 32}
|
||||
batch_size: 1
|
||||
fuse_normalize: True
|
||||
|
||||
|
||||
### model config
|
||||
architecture: FCOS
|
||||
FCOS:
|
||||
backbone: ResNet
|
||||
neck: FPN
|
||||
fcos_head: FCOSHead
|
||||
|
||||
ResNet:
|
||||
depth: 50
|
||||
variant: 'b'
|
||||
norm_type: bn
|
||||
freeze_at: 0 # res2
|
||||
return_idx: [1, 2, 3]
|
||||
num_stages: 4
|
||||
|
||||
FPN:
|
||||
out_channel: 256
|
||||
spatial_scales: [0.125, 0.0625, 0.03125]
|
||||
extra_stage: 2
|
||||
has_extra_convs: True
|
||||
use_c5: False
|
||||
|
||||
FCOSHead:
|
||||
fcos_feat:
|
||||
name: FCOSFeat
|
||||
feat_in: 256
|
||||
feat_out: 256
|
||||
num_convs: 4
|
||||
norm_type: "gn"
|
||||
use_dcn: False
|
||||
fpn_stride: [8, 16, 32, 64, 128]
|
||||
prior_prob: 0.01
|
||||
norm_reg_targets: True
|
||||
centerness_on_reg: True
|
||||
num_shift: 0. # default 0.5
|
||||
multiply_strides_reg_targets: False
|
||||
sqrt_score: False
|
||||
fcos_loss:
|
||||
name: FCOSLoss
|
||||
loss_alpha: 0.25
|
||||
loss_gamma: 2.0
|
||||
iou_loss_type: "giou"
|
||||
reg_weights: 1.0
|
||||
quality: "iou"
|
||||
nms:
|
||||
name: MultiClassNMS
|
||||
nms_top_k: 1000
|
||||
keep_top_k: 100
|
||||
score_threshold: 0.025
|
||||
nms_threshold: 0.6
|
||||
|
||||
|
||||
### other config
|
||||
epoch: *epochs
|
||||
LearningRate:
|
||||
base_lr: 0.01
|
||||
schedulers:
|
||||
- !PiecewiseDecay
|
||||
gamma: 0.1
|
||||
milestones: [*epochs]
|
||||
use_warmup: *use_warmup
|
||||
- !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
|
||||
@@ -0,0 +1,44 @@
|
||||
_BASE_: [
|
||||
'denseteacher_fcos_r50_fpn_coco_semi010.yml',
|
||||
]
|
||||
log_iter: 50
|
||||
snapshot_epoch: 5
|
||||
epochs: &epochs 240
|
||||
weights: output/denseteacher_fcos_r50_fpn_coco_semi010_lsj/model_final
|
||||
|
||||
|
||||
### reader config
|
||||
worker_num: 2
|
||||
SemiTrainReader:
|
||||
sample_transforms:
|
||||
- Decode: {}
|
||||
# large-scale jittering
|
||||
- RandomResize: {target_size: [[400, 1333], [1200, 1333]], keep_ratio: True, interp: 1, random_range: True}
|
||||
- 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]
|
||||
num_shift: 0. # default 0.5
|
||||
multiply_strides_reg_targets: False
|
||||
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
|
||||
@@ -0,0 +1,151 @@
|
||||
_BASE_: [
|
||||
'../../ppyoloe/ppyoloe_plus_crn_l_80e_coco.yml',
|
||||
'../_base_/coco_detection_percent_5.yml',
|
||||
]
|
||||
log_iter: 50
|
||||
snapshot_epoch: 5
|
||||
weights: output/denseteacher_ppyoloe_plus_crn_l_coco_semi005/model_final
|
||||
|
||||
epochs: &epochs 200
|
||||
cosine_epochs: &cosine_epochs 240
|
||||
|
||||
|
||||
### pretrain and warmup config, choose one and comment another
|
||||
pretrain_weights: https://bj.bcebos.com/v1/paddledet/models/ppyoloe_plus_crn_l_80e_coco_sup005.pdparams # mAP=42.9
|
||||
semi_start_iters: 0
|
||||
ema_start_iters: 0
|
||||
use_warmup: &use_warmup False
|
||||
|
||||
# pretrain_weights: https://bj.bcebos.com/v1/paddledet/models/pretrained/ppyoloe_crn_l_obj365_pretrained.pdparams
|
||||
# semi_start_iters: 5000
|
||||
# ema_start_iters: 3000
|
||||
# use_warmup: &use_warmup 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: 1.0, distill_loss_iou: 2.5, distill_loss_dfl: 0., distill_loss_contrast: 0.1}
|
||||
contrast_loss:
|
||||
temperature: 0.2
|
||||
alpha: 0.9
|
||||
smooth_iter: 100
|
||||
concat_sup_data: True
|
||||
suppress: linear
|
||||
ratio: 0.01
|
||||
test_cfg:
|
||||
inference_on: teacher
|
||||
|
||||
|
||||
### reader config
|
||||
batch_size: &batch_size 8
|
||||
worker_num: 2
|
||||
SemiTrainReader:
|
||||
sample_transforms:
|
||||
- Decode: {}
|
||||
- RandomDistort: {}
|
||||
- RandomExpand: {fill_value: [123.675, 116.28, 103.53]}
|
||||
- RandomFlip: {}
|
||||
- RandomCrop: {} # unsup will be fake gt_boxes
|
||||
weak_aug:
|
||||
- NormalizeImage: {mean: [0., 0., 0.], std: [1., 1., 1.], is_scale: true, norm_type: none}
|
||||
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., 0., 0.], std: [1., 1., 1.], is_scale: true, norm_type: none}
|
||||
sup_batch_transforms:
|
||||
- BatchRandomResize: {target_size: [640], random_size: True, random_interp: True, keep_ratio: False}
|
||||
- Permute: {}
|
||||
- PadGT: {}
|
||||
unsup_batch_transforms:
|
||||
- BatchRandomResize: {target_size: [640], random_size: True, random_interp: True, keep_ratio: False}
|
||||
- Permute: {}
|
||||
sup_batch_size: *batch_size
|
||||
unsup_batch_size: *batch_size
|
||||
shuffle: True
|
||||
drop_last: True
|
||||
collate_batch: True
|
||||
|
||||
EvalReader:
|
||||
sample_transforms:
|
||||
- Decode: {}
|
||||
- Resize: {target_size: [640, 640], keep_ratio: False, interp: 2}
|
||||
- NormalizeImage: {mean: [0., 0., 0.], std: [1., 1., 1.], norm_type: none}
|
||||
- Permute: {}
|
||||
batch_size: 2
|
||||
|
||||
TestReader:
|
||||
inputs_def:
|
||||
image_shape: [3, 640, 640]
|
||||
sample_transforms:
|
||||
- Decode: {}
|
||||
- Resize: {target_size: [640, 640], keep_ratio: False, interp: 2}
|
||||
- NormalizeImage: {mean: [0., 0., 0.], std: [1., 1., 1.], norm_type: none}
|
||||
- Permute: {}
|
||||
batch_size: 1
|
||||
|
||||
|
||||
### model config
|
||||
architecture: PPYOLOE
|
||||
norm_type: sync_bn
|
||||
ema_black_list: ['proj_conv.weight']
|
||||
custom_black_list: ['reduce_mean']
|
||||
PPYOLOE:
|
||||
backbone: CSPResNet
|
||||
neck: CustomCSPPAN
|
||||
yolo_head: PPYOLOEHead
|
||||
post_process: ~
|
||||
|
||||
eval_size: ~ # means None, but not str 'None'
|
||||
PPYOLOEHead:
|
||||
fpn_strides: [32, 16, 8]
|
||||
grid_cell_scale: 5.0
|
||||
grid_cell_offset: 0.5
|
||||
static_assigner_epoch: -1 #
|
||||
use_varifocal_loss: True
|
||||
loss_weight: {class: 1.0, iou: 2.5, dfl: 0.5}
|
||||
static_assigner:
|
||||
name: ATSSAssigner
|
||||
topk: 9
|
||||
assigner:
|
||||
name: TaskAlignedAssigner
|
||||
topk: 13
|
||||
alpha: 1.0
|
||||
beta: 6.0
|
||||
nms:
|
||||
name: MultiClassNMS
|
||||
nms_top_k: 1000
|
||||
keep_top_k: 300
|
||||
score_threshold: 0.01
|
||||
nms_threshold: 0.7
|
||||
|
||||
|
||||
### other config
|
||||
epoch: *epochs
|
||||
LearningRate:
|
||||
base_lr: 0.01
|
||||
schedulers:
|
||||
- !CosineDecay
|
||||
max_epochs: *cosine_epochs
|
||||
use_warmup: *use_warmup
|
||||
- !LinearWarmup
|
||||
start_factor: 0.001
|
||||
epochs: 3
|
||||
|
||||
OptimizerBuilder:
|
||||
optimizer:
|
||||
momentum: 0.9
|
||||
type: Momentum
|
||||
regularizer:
|
||||
factor: 0.0005 # dt-fcos 0.0001
|
||||
type: L2
|
||||
clip_grad_by_norm: 1.0 # dt-fcos clip_grad_by_value
|
||||
@@ -0,0 +1,151 @@
|
||||
_BASE_: [
|
||||
'../../ppyoloe/ppyoloe_plus_crn_l_80e_coco.yml',
|
||||
'../_base_/coco_detection_percent_10.yml',
|
||||
]
|
||||
log_iter: 50
|
||||
snapshot_epoch: 5
|
||||
weights: output/denseteacher_ppyoloe_plus_crn_l_coco_semi010/model_final
|
||||
|
||||
epochs: &epochs 200
|
||||
cosine_epochs: &cosine_epochs 240
|
||||
|
||||
|
||||
### pretrain and warmup config, choose one and comment another
|
||||
pretrain_weights: https://bj.bcebos.com/v1/paddledet/models/ppyoloe_plus_crn_l_80e_coco_sup010.pdparams # mAP=45.7
|
||||
semi_start_iters: 0
|
||||
ema_start_iters: 0
|
||||
use_warmup: &use_warmup False
|
||||
|
||||
# pretrain_weights: https://bj.bcebos.com/v1/paddledet/models/pretrained/ppyoloe_crn_l_obj365_pretrained.pdparams
|
||||
# semi_start_iters: 5000
|
||||
# ema_start_iters: 3000
|
||||
# use_warmup: &use_warmup 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: 1.0, distill_loss_iou: 2.5, distill_loss_dfl: 0., distill_loss_contrast: 0.1}
|
||||
contrast_loss:
|
||||
temperature: 0.2
|
||||
alpha: 0.9
|
||||
smooth_iter: 100
|
||||
concat_sup_data: True
|
||||
suppress: linear
|
||||
ratio: 0.01
|
||||
test_cfg:
|
||||
inference_on: teacher
|
||||
|
||||
|
||||
### reader config
|
||||
batch_size: &batch_size 8
|
||||
worker_num: 2
|
||||
SemiTrainReader:
|
||||
sample_transforms:
|
||||
- Decode: {}
|
||||
- RandomDistort: {}
|
||||
- RandomExpand: {fill_value: [123.675, 116.28, 103.53]}
|
||||
- RandomFlip: {}
|
||||
- RandomCrop: {} # unsup will be fake gt_boxes
|
||||
weak_aug:
|
||||
- NormalizeImage: {mean: [0., 0., 0.], std: [1., 1., 1.], is_scale: true, norm_type: none}
|
||||
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., 0., 0.], std: [1., 1., 1.], is_scale: true, norm_type: none}
|
||||
sup_batch_transforms:
|
||||
- BatchRandomResize: {target_size: [640], random_size: True, random_interp: True, keep_ratio: False}
|
||||
- Permute: {}
|
||||
- PadGT: {}
|
||||
unsup_batch_transforms:
|
||||
- BatchRandomResize: {target_size: [640], random_size: True, random_interp: True, keep_ratio: False}
|
||||
- Permute: {}
|
||||
sup_batch_size: *batch_size
|
||||
unsup_batch_size: *batch_size
|
||||
shuffle: True
|
||||
drop_last: True
|
||||
collate_batch: True
|
||||
|
||||
EvalReader:
|
||||
sample_transforms:
|
||||
- Decode: {}
|
||||
- Resize: {target_size: [640, 640], keep_ratio: False, interp: 2}
|
||||
- NormalizeImage: {mean: [0., 0., 0.], std: [1., 1., 1.], norm_type: none}
|
||||
- Permute: {}
|
||||
batch_size: 2
|
||||
|
||||
TestReader:
|
||||
inputs_def:
|
||||
image_shape: [3, 640, 640]
|
||||
sample_transforms:
|
||||
- Decode: {}
|
||||
- Resize: {target_size: [640, 640], keep_ratio: False, interp: 2}
|
||||
- NormalizeImage: {mean: [0., 0., 0.], std: [1., 1., 1.], norm_type: none}
|
||||
- Permute: {}
|
||||
batch_size: 1
|
||||
|
||||
|
||||
### model config
|
||||
architecture: PPYOLOE
|
||||
norm_type: sync_bn
|
||||
ema_black_list: ['proj_conv.weight']
|
||||
custom_black_list: ['reduce_mean']
|
||||
PPYOLOE:
|
||||
backbone: CSPResNet
|
||||
neck: CustomCSPPAN
|
||||
yolo_head: PPYOLOEHead
|
||||
post_process: ~
|
||||
|
||||
eval_size: ~ # means None, but not str 'None'
|
||||
PPYOLOEHead:
|
||||
fpn_strides: [32, 16, 8]
|
||||
grid_cell_scale: 5.0
|
||||
grid_cell_offset: 0.5
|
||||
static_assigner_epoch: -1 #
|
||||
use_varifocal_loss: True
|
||||
loss_weight: {class: 1.0, iou: 2.5, dfl: 0.5}
|
||||
static_assigner:
|
||||
name: ATSSAssigner
|
||||
topk: 9
|
||||
assigner:
|
||||
name: TaskAlignedAssigner
|
||||
topk: 13
|
||||
alpha: 1.0
|
||||
beta: 6.0
|
||||
nms:
|
||||
name: MultiClassNMS
|
||||
nms_top_k: 1000
|
||||
keep_top_k: 300
|
||||
score_threshold: 0.01
|
||||
nms_threshold: 0.7
|
||||
|
||||
|
||||
### other config
|
||||
epoch: *epochs
|
||||
LearningRate:
|
||||
base_lr: 0.01
|
||||
schedulers:
|
||||
- !CosineDecay
|
||||
max_epochs: *cosine_epochs
|
||||
use_warmup: *use_warmup
|
||||
- !LinearWarmup
|
||||
start_factor: 0.001
|
||||
epochs: 3
|
||||
|
||||
OptimizerBuilder:
|
||||
optimizer:
|
||||
momentum: 0.9
|
||||
type: Momentum
|
||||
regularizer:
|
||||
factor: 0.0005 # dt-fcos 0.0001
|
||||
type: L2
|
||||
clip_grad_by_norm: 1.0 # dt-fcos clip_grad_by_value
|
||||
@@ -0,0 +1,151 @@
|
||||
_BASE_: [
|
||||
'../../ppyoloe/ppyoloe_plus_crn_s_80e_coco.yml',
|
||||
'../_base_/coco_detection_percent_5.yml',
|
||||
]
|
||||
log_iter: 50
|
||||
snapshot_epoch: 5
|
||||
weights: output/denseteacher_ppyoloe_plus_crn_s_coco_semi005/model_final
|
||||
|
||||
epochs: &epochs 200
|
||||
cosine_epochs: &cosine_epochs 240
|
||||
|
||||
|
||||
### pretrain and warmup config, choose one and comment another
|
||||
pretrain_weights: https://bj.bcebos.com/v1/paddledet/models/ppyoloe_plus_crn_s_80e_coco_sup005.pdparams # mAP=32.8
|
||||
semi_start_iters: 0
|
||||
ema_start_iters: 0
|
||||
use_warmup: &use_warmup False
|
||||
|
||||
# pretrain_weights: https://bj.bcebos.com/v1/paddledet/models/pretrained/ppyoloe_crn_s_obj365_pretrained.pdparams
|
||||
# semi_start_iters: 5000
|
||||
# ema_start_iters: 3000
|
||||
# use_warmup: &use_warmup 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: 1.0, distill_loss_iou: 2.5, distill_loss_dfl: 0., distill_loss_contrast: 0.1}
|
||||
contrast_loss:
|
||||
temperature: 0.2
|
||||
alpha: 0.9
|
||||
smooth_iter: 100
|
||||
concat_sup_data: True
|
||||
suppress: linear
|
||||
ratio: 0.01
|
||||
test_cfg:
|
||||
inference_on: teacher
|
||||
|
||||
|
||||
### reader config
|
||||
batch_size: &batch_size 8
|
||||
worker_num: 2
|
||||
SemiTrainReader:
|
||||
sample_transforms:
|
||||
- Decode: {}
|
||||
- RandomDistort: {}
|
||||
- RandomExpand: {fill_value: [123.675, 116.28, 103.53]}
|
||||
- RandomFlip: {}
|
||||
- RandomCrop: {} # unsup will be fake gt_boxes
|
||||
weak_aug:
|
||||
- NormalizeImage: {mean: [0., 0., 0.], std: [1., 1., 1.], is_scale: true, norm_type: none}
|
||||
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., 0., 0.], std: [1., 1., 1.], is_scale: true, norm_type: none}
|
||||
sup_batch_transforms:
|
||||
- BatchRandomResize: {target_size: [640], random_size: True, random_interp: True, keep_ratio: False}
|
||||
- Permute: {}
|
||||
- PadGT: {}
|
||||
unsup_batch_transforms:
|
||||
- BatchRandomResize: {target_size: [640], random_size: True, random_interp: True, keep_ratio: False}
|
||||
- Permute: {}
|
||||
sup_batch_size: *batch_size
|
||||
unsup_batch_size: *batch_size
|
||||
shuffle: True
|
||||
drop_last: True
|
||||
collate_batch: True
|
||||
|
||||
EvalReader:
|
||||
sample_transforms:
|
||||
- Decode: {}
|
||||
- Resize: {target_size: [640, 640], keep_ratio: False, interp: 2}
|
||||
- NormalizeImage: {mean: [0., 0., 0.], std: [1., 1., 1.], norm_type: none}
|
||||
- Permute: {}
|
||||
batch_size: 2
|
||||
|
||||
TestReader:
|
||||
inputs_def:
|
||||
image_shape: [3, 640, 640]
|
||||
sample_transforms:
|
||||
- Decode: {}
|
||||
- Resize: {target_size: [640, 640], keep_ratio: False, interp: 2}
|
||||
- NormalizeImage: {mean: [0., 0., 0.], std: [1., 1., 1.], norm_type: none}
|
||||
- Permute: {}
|
||||
batch_size: 1
|
||||
|
||||
|
||||
### model config
|
||||
architecture: PPYOLOE
|
||||
norm_type: sync_bn
|
||||
ema_black_list: ['proj_conv.weight']
|
||||
custom_black_list: ['reduce_mean']
|
||||
PPYOLOE:
|
||||
backbone: CSPResNet
|
||||
neck: CustomCSPPAN
|
||||
yolo_head: PPYOLOEHead
|
||||
post_process: ~
|
||||
|
||||
eval_size: ~ # means None, but not str 'None'
|
||||
PPYOLOEHead:
|
||||
fpn_strides: [32, 16, 8]
|
||||
grid_cell_scale: 5.0
|
||||
grid_cell_offset: 0.5
|
||||
static_assigner_epoch: -1 #
|
||||
use_varifocal_loss: True
|
||||
loss_weight: {class: 1.0, iou: 2.5, dfl: 0.5}
|
||||
static_assigner:
|
||||
name: ATSSAssigner
|
||||
topk: 9
|
||||
assigner:
|
||||
name: TaskAlignedAssigner
|
||||
topk: 13
|
||||
alpha: 1.0
|
||||
beta: 6.0
|
||||
nms:
|
||||
name: MultiClassNMS
|
||||
nms_top_k: 1000
|
||||
keep_top_k: 300
|
||||
score_threshold: 0.01
|
||||
nms_threshold: 0.7
|
||||
|
||||
|
||||
### other config
|
||||
epoch: *epochs
|
||||
LearningRate:
|
||||
base_lr: 0.01
|
||||
schedulers:
|
||||
- !CosineDecay
|
||||
max_epochs: *cosine_epochs
|
||||
use_warmup: *use_warmup
|
||||
- !LinearWarmup
|
||||
start_factor: 0.001
|
||||
epochs: 3
|
||||
|
||||
OptimizerBuilder:
|
||||
optimizer:
|
||||
momentum: 0.9
|
||||
type: Momentum
|
||||
regularizer:
|
||||
factor: 0.0005 # dt-fcos 0.0001
|
||||
type: L2
|
||||
clip_grad_by_norm: 1.0 # dt-fcos clip_grad_by_value
|
||||
@@ -0,0 +1,151 @@
|
||||
_BASE_: [
|
||||
'../../ppyoloe/ppyoloe_plus_crn_s_80e_coco.yml',
|
||||
'../_base_/coco_detection_percent_10.yml',
|
||||
]
|
||||
log_iter: 50
|
||||
snapshot_epoch: 5
|
||||
weights: output/denseteacher_ppyoloe_plus_crn_s_coco_semi010/model_final
|
||||
|
||||
epochs: &epochs 200
|
||||
cosine_epochs: &cosine_epochs 240
|
||||
|
||||
|
||||
### pretrain and warmup config, choose one and comment another
|
||||
pretrain_weights: https://bj.bcebos.com/v1/paddledet/models/ppyoloe_plus_crn_s_80e_coco_sup010.pdparams # mAP=35.3
|
||||
semi_start_iters: 0
|
||||
ema_start_iters: 0
|
||||
use_warmup: &use_warmup False
|
||||
|
||||
# pretrain_weights: https://bj.bcebos.com/v1/paddledet/models/pretrained/ppyoloe_crn_s_obj365_pretrained.pdparams
|
||||
# semi_start_iters: 5000
|
||||
# ema_start_iters: 3000
|
||||
# use_warmup: &use_warmup 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: 1.0, distill_loss_iou: 2.5, distill_loss_dfl: 0., distill_loss_contrast: 0.1}
|
||||
contrast_loss:
|
||||
temperature: 0.2
|
||||
alpha: 0.9
|
||||
smooth_iter: 100
|
||||
concat_sup_data: True
|
||||
suppress: linear
|
||||
ratio: 0.01
|
||||
test_cfg:
|
||||
inference_on: teacher
|
||||
|
||||
|
||||
### reader config
|
||||
batch_size: &batch_size 8
|
||||
worker_num: 2
|
||||
SemiTrainReader:
|
||||
sample_transforms:
|
||||
- Decode: {}
|
||||
- RandomDistort: {}
|
||||
- RandomExpand: {fill_value: [123.675, 116.28, 103.53]}
|
||||
- RandomFlip: {}
|
||||
- RandomCrop: {} # unsup will be fake gt_boxes
|
||||
weak_aug:
|
||||
- NormalizeImage: {mean: [0., 0., 0.], std: [1., 1., 1.], is_scale: true, norm_type: none}
|
||||
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., 0., 0.], std: [1., 1., 1.], is_scale: true, norm_type: none}
|
||||
sup_batch_transforms:
|
||||
- BatchRandomResize: {target_size: [640], random_size: True, random_interp: True, keep_ratio: False}
|
||||
- Permute: {}
|
||||
- PadGT: {}
|
||||
unsup_batch_transforms:
|
||||
- BatchRandomResize: {target_size: [640], random_size: True, random_interp: True, keep_ratio: False}
|
||||
- Permute: {}
|
||||
sup_batch_size: *batch_size
|
||||
unsup_batch_size: *batch_size
|
||||
shuffle: True
|
||||
drop_last: True
|
||||
collate_batch: True
|
||||
|
||||
EvalReader:
|
||||
sample_transforms:
|
||||
- Decode: {}
|
||||
- Resize: {target_size: [640, 640], keep_ratio: False, interp: 2}
|
||||
- NormalizeImage: {mean: [0., 0., 0.], std: [1., 1., 1.], norm_type: none}
|
||||
- Permute: {}
|
||||
batch_size: 2
|
||||
|
||||
TestReader:
|
||||
inputs_def:
|
||||
image_shape: [3, 640, 640]
|
||||
sample_transforms:
|
||||
- Decode: {}
|
||||
- Resize: {target_size: [640, 640], keep_ratio: False, interp: 2}
|
||||
- NormalizeImage: {mean: [0., 0., 0.], std: [1., 1., 1.], norm_type: none}
|
||||
- Permute: {}
|
||||
batch_size: 1
|
||||
|
||||
|
||||
### model config
|
||||
architecture: PPYOLOE
|
||||
norm_type: sync_bn
|
||||
ema_black_list: ['proj_conv.weight']
|
||||
custom_black_list: ['reduce_mean']
|
||||
PPYOLOE:
|
||||
backbone: CSPResNet
|
||||
neck: CustomCSPPAN
|
||||
yolo_head: PPYOLOEHead
|
||||
post_process: ~
|
||||
|
||||
eval_size: ~ # means None, but not str 'None'
|
||||
PPYOLOEHead:
|
||||
fpn_strides: [32, 16, 8]
|
||||
grid_cell_scale: 5.0
|
||||
grid_cell_offset: 0.5
|
||||
static_assigner_epoch: -1 #
|
||||
use_varifocal_loss: True
|
||||
loss_weight: {class: 1.0, iou: 2.5, dfl: 0.5}
|
||||
static_assigner:
|
||||
name: ATSSAssigner
|
||||
topk: 9
|
||||
assigner:
|
||||
name: TaskAlignedAssigner
|
||||
topk: 13
|
||||
alpha: 1.0
|
||||
beta: 6.0
|
||||
nms:
|
||||
name: MultiClassNMS
|
||||
nms_top_k: 1000
|
||||
keep_top_k: 300
|
||||
score_threshold: 0.01
|
||||
nms_threshold: 0.7
|
||||
|
||||
|
||||
### other config
|
||||
epoch: *epochs
|
||||
LearningRate:
|
||||
base_lr: 0.01
|
||||
schedulers:
|
||||
- !CosineDecay
|
||||
max_epochs: *cosine_epochs
|
||||
use_warmup: *use_warmup
|
||||
- !LinearWarmup
|
||||
start_factor: 0.001
|
||||
epochs: 3
|
||||
|
||||
OptimizerBuilder:
|
||||
optimizer:
|
||||
momentum: 0.9
|
||||
type: Momentum
|
||||
regularizer:
|
||||
factor: 0.0005 # dt-fcos 0.0001
|
||||
type: L2
|
||||
clip_grad_by_norm: 1.0 # dt-fcos clip_grad_by_value
|
||||
109
paddle_detection/configs/semi_det/rtdetr_ssod/README.md
Normal file
109
paddle_detection/configs/semi_det/rtdetr_ssod/README.md
Normal file
@@ -0,0 +1,109 @@
|
||||
简体中文 | [English](README_en.md)
|
||||
|
||||
# RTDETR-SSOD(基于RTDETR的配套半监督目标检测方法)
|
||||
# 复现模型指标注意事项: 模型中指标是采用在监督数据训练饱和后加载监督数据所训练的模型进行半监督训练
|
||||
- 例如 使用 baseline/rtdetr_r50vd_6x_coco_sup005.yml使用5%coco数据训练全监督模型，得到rtdetr_r50vd_6x_coco_sup005.pdparams,在rt_detr_ssod005_coco_no_warmup.yml中设置
|
||||
- pretrain_student_weights: rtdetr_r50vd_6x_coco_sup005.pdparams
|
||||
- pretrain_teacher_weights: rtdetr_r50vd_6x_coco_sup005.pdparams
|
||||
- 1.使用coco数据集5%和10%有标记数据和voc数据集VOC2007trainval 所训练的权重已给出请参考 semi_det/baseline/README.md.
|
||||
- 2.rt_detr_ssod_voc_no_warmup.yml rt_detr_ssod005_coco_no_warmup.yml rt_detr_ssod010_coco_no_warmup.yml 是使用训练好的全监督权中直接开启半监督训练（推荐）
|
||||
## RTDETR-SSOD模型库
|
||||
|
||||
| 模型 | 监督数据比例 | Sup Baseline | Sup Epochs (Iters) | Sup mAP<sup>val<br>0.5:0.95 | Semi mAP<sup>val<br>0.5:0.95 | Semi Epochs (Iters) | 模型下载 | 配置文件 |
|
||||
| :------------: | :---------: | :---------------------: | :---------------------: |:---------------------------: |:----------------------------: | :------------------: |:--------: |:----------: |
|
||||
| RTDETR-SSOD | 5% | [sup_config](../baseline/rtdetr_r50vd_6x_coco_sup005.yml) | - | 39.0 | **42.3** | - | [download](https://bj.bcebos.com/v1/paddledet/rt_detr_ssod005_coco_no_warmup.pdparams) | [config](./rt_detr_ssod005_coco_no_warmup.yml) |
|
||||
| RTDETR-SSOD | 10% | [sup_config](../baseline/rtdetr_r50vd_6x_coco_sup010.yml) | -| 42.3 | **44.8** | - | [download](https://bj.bcebos.com/v1/paddledet/data/semidet/rtdetr_ssod/rt_detr_ssod010_coco/rt_detr_ssod010_coco_no_warmup.pdparams) | [config](./rt_detr_ssod010_coco_with_warmup.yml) |
|
||||
| RTDETR-SSOD(VOC)| VOC | [sup_config](../baseline/rtdetr_r50vd_6x_coco_voc2007.yml) | - | 62.7 | **65.8(LSJ)** | - | [download](https://bj.bcebos.com/v1/paddledet/data/semidet/rtdetr_ssod/rt_detr_ssod_voc/rt_detr_ssod_voc_no_warmup.pdparams) | [config](./rt_detr_ssod_voc_with_warmup.yml) |
|
||||
|
||||
**注意:**
|
||||
- 以上模型训练默认使用8 GPUs，监督数据总batch_size默认为16，无监督数据总batch_size默认也为16，默认初始学习率为0.01。如果改动了总batch_size，请按线性比例相应地调整学习率；
|
||||
- **监督数据比例**是指使用的有标签COCO数据集占 COCO train2017 全量训练集的百分比，使用的无标签COCO数据集一般也是相同比例，但具体图片和有标签数据的图片不重合；
|
||||
- `Semi Epochs (Iters)`表示**半监督训练**的模型的 Epochs (Iters)，如果使用**自定义数据集**，需自行根据Iters换算到对应的Epochs调整，最好保证总Iters 和COCO数据集的设置较为接近；
|
||||
- `Sup mAP`是**只使用有监督数据训练**的模型的精度，请参照**基础检测器的配置文件** 和 [baseline](../baseline)；
|
||||
- `Semi mAP`是**半监督训练**的模型的精度，模型下载和配置文件的链接均为**半监督模型**；
|
||||
- `LSJ`表示 **large-scale jittering**，表示使用更大范围的多尺度训练，可进一步提升精度，但训练速度也会变慢；
|
||||
- 半监督检测的配置讲解，请参照[文档](../README.md/#半监督检测配置)；
|
||||
- `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](../../../docs/feature_models/SSLD_PRETRAINED_MODEL.md)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]
|
||||
```
|
||||
|
||||
## 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
|
||||
|
||||
```bash
|
||||
# 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/rtdetr_ssod/rt_detr_ssod010_coco_no_warmup.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/rtdetr_ssod/rt_detr_ssod010_coco_no_warmup.yml --eval
|
||||
```
|
||||
|
||||
### 评估
|
||||
|
||||
```bash
|
||||
CUDA_VISIBLE_DEVICES=0 python tools/eval.py -c configs/semi_det/rtdetr_ssod/rt_detr_ssod010_coco_no_warmup.yml -o weights=output/rt_detr_ssod/model_final/model_final.pdparams
|
||||
```
|
||||
|
||||
### 预测
|
||||
|
||||
```bash
|
||||
CUDA_VISIBLE_DEVICES=0 python tools/infer.py -c configs/semi_det/rtdetr_ssod/rt_detr_ssod010_coco_no_warmup.yml -o weights=output/rt_detr_ssod/model_final/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。
|
||||
|
||||
```bash
|
||||
# 导出模型
|
||||
CUDA_VISIBLE_DEVICES=0 python tools/export_model.py -c configs/semi_det/rtdetr_ssod/rt_detr_ssod010_coco_no_warmup.yml -o weights=https://paddledet.bj.bcebos.com/models/rt_detr_ssod010_coco_no_warmup.pdparams
|
||||
|
||||
# 导出权重预测
|
||||
CUDA_VISIBLE_DEVICES=0 python deploy/python/infer.py --model_dir=output_inference/rt_detr_ssod010_coco_no_warmup --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/rt_detr_ssod010_coco_no_warmup --image_file=demo/000000014439_640x640.jpg --device=GPU --run_benchmark=True # --run_mode=trt_fp16
|
||||
|
||||
# 导出ONNX
|
||||
paddle2onnx --model_dir output_inference/drt_detr_ssod010_coco_no_warmup/ --model_filename model.pdmodel --params_filename model.pdiparams --opset_version 12 --save_file rt_detr_ssod010_coco_no_warmup.onnx
|
||||
```
|
||||
|
||||
|
||||
# RTDETR-SSOD 下游任务
|
||||
|
||||
我们验证了RTDETR-SSOD模型强大的泛化能力，在低光、工业、交通等不同场景下游任务检测效果稳定提升!
|
||||
|
||||
voc数据集采用[voc](https://github.com/thsant/wgisd)，是一个广泛使用的计算机视觉数据集，用于目标检测、图像分割和场景理解等任务。该数据集包含20个类别的图像,处理后的COCO格式，包含图片标注训练集5011张，图片无标注训练集11540张，测试集2510张，20个类别；
|
||||
|
||||
低光数据集使用[ExDark](https://github.com/cs-chan/Exclusively-Dark-Image-Dataset/tree/master/Dataset)，该数据集是一个专门在低光照环境下拍摄出针对低光目标检测的数据集，包括从极低光环境到暮光环境等10种不同光照条件下的图片，处理后的COCO格式，包含图片训练集5891张，测试集1472张，12个类别;
|
||||
|
||||
工业数据集使用[PKU-Market-PCB](https://robotics.pkusz.edu.cn/resources/dataset/)，该数据集用于印刷电路板（PCB）的瑕疵检测，提供了6种常见的PCB缺陷;
|
||||
|
||||
商超数据集[SKU110k](https://github.com/eg4000/SKU110K_CVPR19)是商品超市场景下的密集目标检测数据集，包含11,762张图片和超过170个实例。其中包括8,233张用于训练的图像、588张用于验证的图像和2,941张用于测试的图像;
|
||||
|
||||
自动驾驶数据集使用[sslad](https://soda-2d.github.io/index.html);
|
||||
|
||||
交通数据集使用[visdrone](http://aiskyeye.com/home/);
|
||||
|
||||
## 下游数据集实验结果：
|
||||
|
||||
| 数据集 | 业务方向 | 划分 | labeled数据量 | 全监督mAP | 半监督mAP |
|
||||
|----------|-----------|---------------------|-----------------|------------------|--------------|
|
||||
| voc | 通用 | voc07, 12；1:2 | 5000 | 63.1 | 65.8（+2.7） |
|
||||
| visdrone | 无人机交通 | 1:9 | 647 | 19.4 | 20.6 (+1.2) |
|
||||
| pcb | 工业缺陷 | 1:9 | 55 | 22.9 | 26.8 (+3.9) |
|
||||
| sku110k | 商品 | 1:9 | 821 | 38.9 | 52.4 (+13.5) |
|
||||
| sslad | 自动驾驶 | 1:32 | 4967 | 42.1 | 43.3 (+1.2) |
|
||||
| exdark | 低光照 | 1:9 | 589 | 39.6 | 44.1 (+4.5) |
|
||||
@@ -0,0 +1,212 @@
|
||||
_BASE_: [
|
||||
'../../runtime.yml',
|
||||
'../../rtdetr/_base_/rtdetr_r50vd.yml',
|
||||
'../../rtdetr/_base_/rtdetr_reader.yml',
|
||||
]
|
||||
eval_interval: 4000
|
||||
save_interval: 4000
|
||||
weights: output/rt_detr_ssod/model_final
|
||||
find_unused_parameters: True
|
||||
save_dir: output
|
||||
log_iter: 1
|
||||
ssod_method: Semi_RTDETR
|
||||
### global config
|
||||
use_simple_ema: True
|
||||
ema_decay: 0.9996
|
||||
use_gpu: true
|
||||
|
||||
### reader config
|
||||
worker_num: 4
|
||||
|
||||
SemiTrainReader:
|
||||
sample_transforms:
|
||||
- Decode: {}
|
||||
- RandomDistort: {prob: 0.8}
|
||||
- RandomExpand: {fill_value: [0., 0., 0.]}
|
||||
- RandomCrop: {prob: 0.8}
|
||||
- RandomFlip: {}
|
||||
weak_aug:
|
||||
- RandomFlip: {prob: 0.0}
|
||||
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},
|
||||
]}
|
||||
sup_batch_transforms:
|
||||
- BatchRandomResizeForSSOD: {target_size: [480, 512, 544, 576, 608, 640, 640, 640, 672, 704, 736, 768, 800], random_size: True, random_interp: True, keep_ratio: False}
|
||||
- NormalizeImage: {mean: [0., 0., 0.], std: [1., 1., 1.], norm_type: none}
|
||||
- NormalizeBox: {}
|
||||
- BboxXYXY2XYWH: {}
|
||||
- Permute: {}
|
||||
unsup_batch_transforms:
|
||||
- BatchRandomResizeForSSOD: {target_size: [480, 512, 544, 576, 608, 640, 640, 640, 672, 704, 736, 768, 800], random_size: True, random_interp: True, keep_ratio: False}
|
||||
- NormalizeImage: {mean: [0., 0., 0.], std: [1., 1., 1.], norm_type: none}
|
||||
- NormalizeBox: {}
|
||||
- BboxXYXY2XYWH: {}
|
||||
- Permute: {}
|
||||
sup_batch_size: 2
|
||||
unsup_batch_size: 2
|
||||
shuffle: true
|
||||
drop_last: true
|
||||
collate_batch: false
|
||||
use_shared_memory: false
|
||||
|
||||
EvalReader:
|
||||
sample_transforms:
|
||||
- Decode: {}
|
||||
- Resize: {target_size: [640, 640], keep_ratio: False}
|
||||
- NormalizeImage: {mean: [0., 0., 0.], std: [1., 1., 1.], norm_type: none}
|
||||
- Permute: {}
|
||||
batch_size: 2
|
||||
shuffle: false
|
||||
drop_last: false
|
||||
|
||||
|
||||
TestReader:
|
||||
sample_transforms:
|
||||
- Decode: {}
|
||||
- Resize: { target_size: [640, 640], keep_ratio: False }
|
||||
- NormalizeImage: {mean: [0., 0., 0.], std: [1., 1., 1.], norm_type: none}
|
||||
- Permute: {}
|
||||
batch_size: 1
|
||||
shuffle: false
|
||||
drop_last: false
|
||||
|
||||
|
||||
pretrain_student_weights: https://bj.bcebos.com/v1/paddledet/data/semidet/rtdetr_ssod/baseline/rtdetr_r50vd_6x_coco_sup005.pdparams
|
||||
pretrain_teacher_weights: https://bj.bcebos.com/v1/paddledet/data/semidet/rtdetr_ssod/baseline/rtdetr_r50vd_6x_coco_sup005.pdparams
|
||||
|
||||
hidden_dim: 256
|
||||
use_focal_loss: True
|
||||
eval_size: [640, 640]
|
||||
|
||||
architecture: DETR
|
||||
DETR:
|
||||
backbone: ResNet
|
||||
neck: HybridEncoder
|
||||
transformer: RTDETRTransformer
|
||||
detr_head: DINOHead
|
||||
post_process: DETRPostProcess
|
||||
post_process_semi: DETRBBoxSemiPostProcess
|
||||
ResNet:
|
||||
depth: 50
|
||||
variant: d
|
||||
norm_type: bn
|
||||
freeze_at: 0
|
||||
return_idx: [1, 2, 3]
|
||||
lr_mult_list: [0.1, 0.1, 0.1, 0.1]
|
||||
num_stages: 4
|
||||
freeze_stem_only: True
|
||||
|
||||
HybridEncoder:
|
||||
hidden_dim: 256
|
||||
use_encoder_idx: [2]
|
||||
num_encoder_layers: 1
|
||||
encoder_layer:
|
||||
name: TransformerLayer
|
||||
d_model: 256
|
||||
nhead: 8
|
||||
dim_feedforward: 1024
|
||||
dropout: 0.
|
||||
activation: 'gelu'
|
||||
expansion: 1.0
|
||||
|
||||
|
||||
PPDETRTransformer:
|
||||
num_queries: 300
|
||||
position_embed_type: sine
|
||||
feat_strides: [8, 16, 32]
|
||||
num_levels: 3
|
||||
nhead: 8
|
||||
num_decoder_layers: 6
|
||||
dim_feedforward: 1024
|
||||
dropout: 0.0
|
||||
activation: relu
|
||||
num_denoising: 100
|
||||
label_noise_ratio: 0.5
|
||||
box_noise_scale: 1.0
|
||||
learnt_init_query: False
|
||||
|
||||
DINOHead:
|
||||
loss:
|
||||
name: DINOLoss
|
||||
loss_coeff: {class: 1, bbox: 5, giou: 2}
|
||||
aux_loss: True
|
||||
use_vfl: True
|
||||
matcher:
|
||||
name: HungarianMatcher
|
||||
matcher_coeff: {class: 2, bbox: 5, giou: 2}
|
||||
|
||||
DETRPostProcess:
|
||||
num_top_queries: 300
|
||||
|
||||
|
||||
|
||||
SSOD: DETR_SSOD
|
||||
DETR_SSOD:
|
||||
teacher: DETR
|
||||
student: DETR
|
||||
train_cfg:
|
||||
sup_weight: 1.0
|
||||
unsup_weight: 1.0
|
||||
ema_start_iters: -1
|
||||
pseudo_label_initial_score_thr: 0.7
|
||||
min_pseduo_box_size: 0
|
||||
concat_sup_data: True
|
||||
test_cfg:
|
||||
inference_on: teacher
|
||||
|
||||
|
||||
|
||||
metric: COCO
|
||||
num_classes: 80
|
||||
|
||||
# partial labeled COCO, use `SemiCOCODataSet` rather than `COCODataSet`
|
||||
TrainDataset:
|
||||
!SemiCOCODataSet
|
||||
image_dir: train2017
|
||||
anno_path: semi_annotations/instances_train2017.1@5.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@5-unlabeled.json
|
||||
dataset_dir: dataset/coco
|
||||
data_fields: ['image']
|
||||
supervised: False
|
||||
|
||||
EvalDataset:
|
||||
!COCODataSet
|
||||
image_dir: val2017
|
||||
anno_path: annotations/instances_val2017.json
|
||||
dataset_dir: dataset/coco
|
||||
allow_empty: true
|
||||
|
||||
TestDataset:
|
||||
!ImageFolder
|
||||
anno_path: annotations/instances_val2017.json # also support txt (like VOC's label_list.txt)
|
||||
dataset_dir: dataset/coco # if set, anno_path will be 'dataset_dir/anno_path'
|
||||
|
||||
epoch: 400 #epoch: 60
|
||||
|
||||
LearningRate:
|
||||
base_lr: 0.0002
|
||||
schedulers:
|
||||
- !PiecewiseDecay
|
||||
gamma: 1.0
|
||||
milestones: [400]
|
||||
use_warmup: false
|
||||
- !LinearWarmup
|
||||
start_factor: 0.001
|
||||
steps: 2000
|
||||
|
||||
OptimizerBuilder:
|
||||
clip_grad_by_norm: 0.1
|
||||
regularizer: false
|
||||
optimizer:
|
||||
type: AdamW
|
||||
weight_decay: 0.0001
|
||||
@@ -0,0 +1,215 @@
|
||||
_BASE_: [
|
||||
'../../runtime.yml',
|
||||
'../../rtdetr/_base_/rtdetr_r50vd.yml',
|
||||
'../../rtdetr/_base_/rtdetr_reader.yml',
|
||||
]
|
||||
|
||||
#for debug
|
||||
eval_interval: 4000
|
||||
save_interval: 4000
|
||||
weights: output/rt_detr_ssod/model_final
|
||||
find_unused_parameters: True
|
||||
save_dir: output
|
||||
log_iter: 50
|
||||
ssod_method: Semi_RTDETR
|
||||
### global config
|
||||
use_simple_ema: True
|
||||
ema_decay: 0.9996
|
||||
use_gpu: true
|
||||
|
||||
### reader config
|
||||
worker_num: 4
|
||||
|
||||
SemiTrainReader:
|
||||
sample_transforms:
|
||||
- Decode: {}
|
||||
- RandomDistort: {prob: 0.8}
|
||||
- RandomExpand: {fill_value: [0., 0., 0.]}
|
||||
- RandomCrop: {prob: 0.8}
|
||||
- RandomFlip: {}
|
||||
weak_aug:
|
||||
- RandomFlip: {prob: 0.0}
|
||||
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},
|
||||
]}
|
||||
sup_batch_transforms:
|
||||
- BatchRandomResizeForSSOD: {target_size: [480, 512, 544, 576, 608, 640, 640, 640, 672, 704, 736, 768, 800], random_size: True, random_interp: True, keep_ratio: False}
|
||||
- NormalizeImage: {mean: [0., 0., 0.], std: [1., 1., 1.], norm_type: none}
|
||||
- NormalizeBox: {}
|
||||
- BboxXYXY2XYWH: {}
|
||||
- Permute: {}
|
||||
unsup_batch_transforms:
|
||||
- BatchRandomResizeForSSOD: {target_size: [480, 512, 544, 576, 608, 640, 640, 640, 672, 704, 736, 768, 800], random_size: True, random_interp: True, keep_ratio: False}
|
||||
- NormalizeImage: {mean: [0., 0., 0.], std: [1., 1., 1.], norm_type: none}
|
||||
- NormalizeBox: {}
|
||||
- BboxXYXY2XYWH: {}
|
||||
- Permute: {}
|
||||
sup_batch_size: 2
|
||||
unsup_batch_size: 2
|
||||
shuffle: true
|
||||
drop_last: true
|
||||
collate_batch: false
|
||||
use_shared_memory: false
|
||||
|
||||
EvalReader:
|
||||
sample_transforms:
|
||||
- Decode: {}
|
||||
- Resize: {target_size: [640, 640], keep_ratio: False}
|
||||
- NormalizeImage: {mean: [0., 0., 0.], std: [1., 1., 1.], norm_type: none}
|
||||
- Permute: {}
|
||||
batch_size: 2
|
||||
shuffle: false
|
||||
drop_last: false
|
||||
|
||||
|
||||
TestReader:
|
||||
sample_transforms:
|
||||
- Decode: {}
|
||||
- Resize: { target_size: [640, 640], keep_ratio: False }
|
||||
- NormalizeImage: {mean: [0., 0., 0.], std: [1., 1., 1.], norm_type: none}
|
||||
- Permute: {}
|
||||
batch_size: 1
|
||||
shuffle: false
|
||||
drop_last: false
|
||||
|
||||
|
||||
pretrain_student_weights: https://paddledet.bj.bcebos.com/models/pretrained/ResNet50_vd_ssld_v2_pretrained.pdparams
|
||||
pretrain_teacher_weights: https://paddledet.bj.bcebos.com/models/pretrained/ResNet50_vd_ssld_v2_pretrained.pdparams
|
||||
|
||||
hidden_dim: 256
|
||||
use_focal_loss: True
|
||||
eval_size: [640, 640]
|
||||
|
||||
architecture: DETR
|
||||
DETR:
|
||||
backbone: ResNet
|
||||
neck: HybridEncoder
|
||||
transformer: RTDETRTransformer
|
||||
detr_head: DINOHead
|
||||
post_process: DETRPostProcess
|
||||
post_process_semi: DETRBBoxSemiPostProcess
|
||||
ResNet:
|
||||
# index 0 stands for res2
|
||||
depth: 50
|
||||
variant: d
|
||||
norm_type: bn
|
||||
freeze_at: 0
|
||||
return_idx: [1, 2, 3]
|
||||
lr_mult_list: [0.1, 0.1, 0.1, 0.1]
|
||||
num_stages: 4
|
||||
freeze_stem_only: True
|
||||
|
||||
HybridEncoder:
|
||||
hidden_dim: 256
|
||||
use_encoder_idx: [2]
|
||||
num_encoder_layers: 1
|
||||
encoder_layer:
|
||||
name: TransformerLayer
|
||||
d_model: 256
|
||||
nhead: 8
|
||||
dim_feedforward: 1024
|
||||
dropout: 0.
|
||||
activation: 'gelu'
|
||||
expansion: 1.0
|
||||
|
||||
|
||||
RTDETRTransformer:
|
||||
num_queries: 300
|
||||
position_embed_type: sine
|
||||
feat_strides: [8, 16, 32]
|
||||
num_levels: 3
|
||||
nhead: 8
|
||||
num_decoder_layers: 6
|
||||
dim_feedforward: 1024
|
||||
dropout: 0.0
|
||||
activation: relu
|
||||
num_denoising: 100
|
||||
label_noise_ratio: 0.5
|
||||
box_noise_scale: 1.0
|
||||
learnt_init_query: False
|
||||
|
||||
DINOHead:
|
||||
loss:
|
||||
name: DINOLoss
|
||||
loss_coeff: {class: 1, bbox: 5, giou: 2}
|
||||
aux_loss: True
|
||||
use_vfl: True
|
||||
matcher:
|
||||
name: HungarianMatcher
|
||||
matcher_coeff: {class: 2, bbox: 5, giou: 2}
|
||||
use_uni_match: True
|
||||
DETRPostProcess:
|
||||
num_top_queries: 300
|
||||
|
||||
|
||||
|
||||
SSOD: DETR_SSOD
|
||||
DETR_SSOD:
|
||||
teacher: DETR
|
||||
student: DETR
|
||||
train_cfg:
|
||||
sup_weight: 1.0
|
||||
unsup_weight: 1.0
|
||||
ema_start_iters: 10000
|
||||
pseudo_label_initial_score_thr: 0.7
|
||||
min_pseduo_box_size: 0
|
||||
concat_sup_data: True
|
||||
test_cfg:
|
||||
inference_on: teacher
|
||||
|
||||
|
||||
|
||||
metric: COCO
|
||||
num_classes: 80
|
||||
|
||||
# partial labeled COCO, use `SemiCOCODataSet` rather than `COCODataSet`
|
||||
TrainDataset:
|
||||
!SemiCOCODataSet
|
||||
image_dir: train2017
|
||||
anno_path: semi_annotations/instances_train2017.1@5.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@5-unlabeled.json
|
||||
dataset_dir: dataset/coco
|
||||
data_fields: ['image']
|
||||
supervised: False
|
||||
|
||||
EvalDataset:
|
||||
!COCODataSet
|
||||
image_dir: val2017
|
||||
anno_path: annotations/instances_val2017.json
|
||||
dataset_dir: dataset/coco
|
||||
allow_empty: true
|
||||
|
||||
TestDataset:
|
||||
!ImageFolder
|
||||
anno_path: annotations/instances_val2017.json # also support txt (like VOC's label_list.txt)
|
||||
dataset_dir: dataset/coco # if set, anno_path will be 'dataset_dir/anno_path'
|
||||
|
||||
epoch: 500 #epoch: 60
|
||||
|
||||
LearningRate:
|
||||
base_lr: 0.0002
|
||||
schedulers:
|
||||
- !PiecewiseDecay
|
||||
gamma: 1.0
|
||||
milestones: [500]
|
||||
use_warmup: false
|
||||
- !LinearWarmup
|
||||
start_factor: 0.001
|
||||
steps: 2000
|
||||
|
||||
OptimizerBuilder:
|
||||
clip_grad_by_norm: 0.1
|
||||
regularizer: false
|
||||
optimizer:
|
||||
type: AdamW
|
||||
weight_decay: 0.0001
|
||||
@@ -0,0 +1,212 @@
|
||||
_BASE_: [
|
||||
'../../runtime.yml',
|
||||
'../../rtdetr/_base_/rtdetr_r50vd.yml',
|
||||
'../../rtdetr/_base_/rtdetr_reader.yml',
|
||||
]
|
||||
eval_interval: 4000
|
||||
save_interval: 4000
|
||||
weights: output/rt_detr_ssod/model_final
|
||||
find_unused_parameters: True
|
||||
save_dir: output
|
||||
log_iter: 1
|
||||
ssod_method: Semi_RTDETR
|
||||
### global config
|
||||
use_simple_ema: True
|
||||
ema_decay: 0.9996
|
||||
use_gpu: true
|
||||
|
||||
### reader config
|
||||
worker_num: 4
|
||||
|
||||
SemiTrainReader:
|
||||
sample_transforms:
|
||||
- Decode: {}
|
||||
- RandomDistort: {prob: 0.8}
|
||||
- RandomExpand: {fill_value: [0., 0., 0.]}
|
||||
- RandomCrop: {prob: 0.8}
|
||||
- RandomFlip: {}
|
||||
weak_aug:
|
||||
- RandomFlip: {prob: 0.0}
|
||||
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},
|
||||
]}
|
||||
sup_batch_transforms:
|
||||
- BatchRandomResizeForSSOD: {target_size: [480, 512, 544, 576, 608, 640, 640, 640, 672, 704, 736, 768, 800], random_size: True, random_interp: True, keep_ratio: False}
|
||||
- NormalizeImage: {mean: [0., 0., 0.], std: [1., 1., 1.], norm_type: none}
|
||||
- NormalizeBox: {}
|
||||
- BboxXYXY2XYWH: {}
|
||||
- Permute: {}
|
||||
unsup_batch_transforms:
|
||||
- BatchRandomResizeForSSOD: {target_size: [480, 512, 544, 576, 608, 640, 640, 640, 672, 704, 736, 768, 800], random_size: True, random_interp: True, keep_ratio: False}
|
||||
- NormalizeImage: {mean: [0., 0., 0.], std: [1., 1., 1.], norm_type: none}
|
||||
- NormalizeBox: {}
|
||||
- BboxXYXY2XYWH: {}
|
||||
- Permute: {}
|
||||
sup_batch_size: 2
|
||||
unsup_batch_size: 2
|
||||
shuffle: true
|
||||
drop_last: true
|
||||
collate_batch: false
|
||||
use_shared_memory: false
|
||||
|
||||
EvalReader:
|
||||
sample_transforms:
|
||||
- Decode: {}
|
||||
- Resize: {target_size: [640, 640], keep_ratio: False}
|
||||
- NormalizeImage: {mean: [0., 0., 0.], std: [1., 1., 1.], norm_type: none}
|
||||
- Permute: {}
|
||||
batch_size: 2
|
||||
shuffle: false
|
||||
drop_last: false
|
||||
|
||||
|
||||
TestReader:
|
||||
sample_transforms:
|
||||
- Decode: {}
|
||||
- Resize: { target_size: [640, 640], keep_ratio: False }
|
||||
- NormalizeImage: {mean: [0., 0., 0.], std: [1., 1., 1.], norm_type: none}
|
||||
- Permute: {}
|
||||
batch_size: 1
|
||||
shuffle: false
|
||||
drop_last: false
|
||||
|
||||
|
||||
pretrain_student_weights: https://bj.bcebos.com/v1/paddledet/data/semidet/rtdetr_ssod/baseline/rtdetr_r50vd_6x_coco_sup010.pdparams
|
||||
pretrain_teacher_weights: https://bj.bcebos.com/v1/paddledet/data/semidet/rtdetr_ssod/baseline/rtdetr_r50vd_6x_coco_sup010.pdparams
|
||||
|
||||
hidden_dim: 256
|
||||
use_focal_loss: True
|
||||
eval_size: [640, 640]
|
||||
|
||||
architecture: DETR
|
||||
DETR:
|
||||
backbone: ResNet
|
||||
neck: HybridEncoder
|
||||
transformer: RTDETRTransformer
|
||||
detr_head: DINOHead
|
||||
post_process: DETRPostProcess
|
||||
post_process_semi: DETRBBoxSemiPostProcess
|
||||
ResNet:
|
||||
depth: 50
|
||||
variant: d
|
||||
norm_type: bn
|
||||
freeze_at: 0
|
||||
return_idx: [1, 2, 3]
|
||||
lr_mult_list: [0.1, 0.1, 0.1, 0.1]
|
||||
num_stages: 4
|
||||
freeze_stem_only: True
|
||||
|
||||
HybridEncoder:
|
||||
hidden_dim: 256
|
||||
use_encoder_idx: [2]
|
||||
num_encoder_layers: 1
|
||||
encoder_layer:
|
||||
name: TransformerLayer
|
||||
d_model: 256
|
||||
nhead: 8
|
||||
dim_feedforward: 1024
|
||||
dropout: 0.
|
||||
activation: 'gelu'
|
||||
expansion: 1.0
|
||||
|
||||
|
||||
PPDETRTransformer:
|
||||
num_queries: 300
|
||||
position_embed_type: sine
|
||||
feat_strides: [8, 16, 32]
|
||||
num_levels: 3
|
||||
nhead: 8
|
||||
num_decoder_layers: 6
|
||||
dim_feedforward: 1024
|
||||
dropout: 0.0
|
||||
activation: relu
|
||||
num_denoising: 100
|
||||
label_noise_ratio: 0.5
|
||||
box_noise_scale: 1.0
|
||||
learnt_init_query: False
|
||||
|
||||
DINOHead:
|
||||
loss:
|
||||
name: DINOLoss
|
||||
loss_coeff: {class: 1, bbox: 5, giou: 2}
|
||||
aux_loss: True
|
||||
use_vfl: True
|
||||
matcher:
|
||||
name: HungarianMatcher
|
||||
matcher_coeff: {class: 2, bbox: 5, giou: 2}
|
||||
|
||||
DETRPostProcess:
|
||||
num_top_queries: 300
|
||||
|
||||
|
||||
|
||||
SSOD: DETR_SSOD
|
||||
DETR_SSOD:
|
||||
teacher: DETR
|
||||
student: DETR
|
||||
train_cfg:
|
||||
sup_weight: 1.0
|
||||
unsup_weight: 1.0
|
||||
ema_start_iters: -1
|
||||
pseudo_label_initial_score_thr: 0.7
|
||||
min_pseduo_box_size: 0
|
||||
concat_sup_data: True
|
||||
test_cfg:
|
||||
inference_on: teacher
|
||||
|
||||
|
||||
|
||||
metric: COCO
|
||||
num_classes: 80
|
||||
|
||||
# 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
|
||||
|
||||
EvalDataset:
|
||||
!COCODataSet
|
||||
image_dir: val2017
|
||||
anno_path: annotations/instances_val2017.json
|
||||
dataset_dir: dataset/coco
|
||||
allow_empty: true
|
||||
|
||||
TestDataset:
|
||||
!ImageFolder
|
||||
anno_path: annotations/instances_val2017.json # also support txt (like VOC's label_list.txt)
|
||||
dataset_dir: dataset/coco # if set, anno_path will be 'dataset_dir/anno_path'
|
||||
|
||||
epoch: 400 #epoch: 60
|
||||
|
||||
LearningRate:
|
||||
base_lr: 0.0002
|
||||
schedulers:
|
||||
- !PiecewiseDecay
|
||||
gamma: 1.0
|
||||
milestones: [400]
|
||||
use_warmup: false
|
||||
- !LinearWarmup
|
||||
start_factor: 0.001
|
||||
steps: 2000
|
||||
|
||||
OptimizerBuilder:
|
||||
clip_grad_by_norm: 0.1
|
||||
regularizer: false
|
||||
optimizer:
|
||||
type: AdamW
|
||||
weight_decay: 0.0001
|
||||
@@ -0,0 +1,215 @@
|
||||
_BASE_: [
|
||||
'../../runtime.yml',
|
||||
'../../rtdetr/_base_/rtdetr_r50vd.yml',
|
||||
'../../rtdetr/_base_/rtdetr_reader.yml',
|
||||
]
|
||||
|
||||
#for debug
|
||||
eval_interval: 4000
|
||||
save_interval: 4000
|
||||
weights: output/rt_detr_ssod/model_final
|
||||
find_unused_parameters: True
|
||||
save_dir: output
|
||||
log_iter: 50
|
||||
ssod_method: Semi_RTDETR
|
||||
### global config
|
||||
use_simple_ema: True
|
||||
ema_decay: 0.9996
|
||||
use_gpu: true
|
||||
|
||||
### reader config
|
||||
worker_num: 4
|
||||
|
||||
SemiTrainReader:
|
||||
sample_transforms:
|
||||
- Decode: {}
|
||||
- RandomDistort: {prob: 0.8}
|
||||
- RandomExpand: {fill_value: [0., 0., 0.]}
|
||||
- RandomCrop: {prob: 0.8}
|
||||
- RandomFlip: {}
|
||||
weak_aug:
|
||||
- RandomFlip: {prob: 0.0}
|
||||
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},
|
||||
]}
|
||||
sup_batch_transforms:
|
||||
- BatchRandomResizeForSSOD: {target_size: [480, 512, 544, 576, 608, 640, 640, 640, 672, 704, 736, 768, 800], random_size: True, random_interp: True, keep_ratio: False}
|
||||
- NormalizeImage: {mean: [0., 0., 0.], std: [1., 1., 1.], norm_type: none}
|
||||
- NormalizeBox: {}
|
||||
- BboxXYXY2XYWH: {}
|
||||
- Permute: {}
|
||||
unsup_batch_transforms:
|
||||
- BatchRandomResizeForSSOD: {target_size: [480, 512, 544, 576, 608, 640, 640, 640, 672, 704, 736, 768, 800], random_size: True, random_interp: True, keep_ratio: False}
|
||||
- NormalizeImage: {mean: [0., 0., 0.], std: [1., 1., 1.], norm_type: none}
|
||||
- NormalizeBox: {}
|
||||
- BboxXYXY2XYWH: {}
|
||||
- Permute: {}
|
||||
sup_batch_size: 2
|
||||
unsup_batch_size: 2
|
||||
shuffle: true
|
||||
drop_last: true
|
||||
collate_batch: false
|
||||
use_shared_memory: false
|
||||
|
||||
EvalReader:
|
||||
sample_transforms:
|
||||
- Decode: {}
|
||||
- Resize: {target_size: [640, 640], keep_ratio: False}
|
||||
- NormalizeImage: {mean: [0., 0., 0.], std: [1., 1., 1.], norm_type: none}
|
||||
- Permute: {}
|
||||
batch_size: 2
|
||||
shuffle: false
|
||||
drop_last: false
|
||||
|
||||
|
||||
TestReader:
|
||||
sample_transforms:
|
||||
- Decode: {}
|
||||
- Resize: { target_size: [640, 640], keep_ratio: False }
|
||||
- NormalizeImage: {mean: [0., 0., 0.], std: [1., 1., 1.], norm_type: none}
|
||||
- Permute: {}
|
||||
batch_size: 1
|
||||
shuffle: false
|
||||
drop_last: false
|
||||
|
||||
|
||||
pretrain_student_weights: https://paddledet.bj.bcebos.com/models/pretrained/ResNet50_vd_ssld_v2_pretrained.pdparams
|
||||
pretrain_teacher_weights: https://paddledet.bj.bcebos.com/models/pretrained/ResNet50_vd_ssld_v2_pretrained.pdparams
|
||||
|
||||
hidden_dim: 256
|
||||
use_focal_loss: True
|
||||
eval_size: [640, 640]
|
||||
|
||||
architecture: DETR
|
||||
DETR:
|
||||
backbone: ResNet
|
||||
neck: HybridEncoder
|
||||
transformer: RTDETRTransformer
|
||||
detr_head: DINOHead
|
||||
post_process: DETRPostProcess
|
||||
post_process_semi: DETRBBoxSemiPostProcess
|
||||
ResNet:
|
||||
# index 0 stands for res2
|
||||
depth: 50
|
||||
variant: d
|
||||
norm_type: bn
|
||||
freeze_at: 0
|
||||
return_idx: [1, 2, 3]
|
||||
lr_mult_list: [0.1, 0.1, 0.1, 0.1]
|
||||
num_stages: 4
|
||||
freeze_stem_only: True
|
||||
|
||||
HybridEncoder:
|
||||
hidden_dim: 256
|
||||
use_encoder_idx: [2]
|
||||
num_encoder_layers: 1
|
||||
encoder_layer:
|
||||
name: TransformerLayer
|
||||
d_model: 256
|
||||
nhead: 8
|
||||
dim_feedforward: 1024
|
||||
dropout: 0.
|
||||
activation: 'gelu'
|
||||
expansion: 1.0
|
||||
|
||||
|
||||
RTDETRTransformer:
|
||||
num_queries: 300
|
||||
position_embed_type: sine
|
||||
feat_strides: [8, 16, 32]
|
||||
num_levels: 3
|
||||
nhead: 8
|
||||
num_decoder_layers: 6
|
||||
dim_feedforward: 1024
|
||||
dropout: 0.0
|
||||
activation: relu
|
||||
num_denoising: 100
|
||||
label_noise_ratio: 0.5
|
||||
box_noise_scale: 1.0
|
||||
learnt_init_query: False
|
||||
|
||||
DINOHead:
|
||||
loss:
|
||||
name: DINOLoss
|
||||
loss_coeff: {class: 1, bbox: 5, giou: 2}
|
||||
aux_loss: True
|
||||
use_vfl: True
|
||||
matcher:
|
||||
name: HungarianMatcher
|
||||
matcher_coeff: {class: 2, bbox: 5, giou: 2}
|
||||
use_uni_match: True
|
||||
DETRPostProcess:
|
||||
num_top_queries: 300
|
||||
|
||||
|
||||
|
||||
SSOD: DETR_SSOD
|
||||
DETR_SSOD:
|
||||
teacher: DETR
|
||||
student: DETR
|
||||
train_cfg:
|
||||
sup_weight: 1.0
|
||||
unsup_weight: 1.0
|
||||
ema_start_iters: 10000
|
||||
pseudo_label_initial_score_thr: 0.7
|
||||
min_pseduo_box_size: 0
|
||||
concat_sup_data: True
|
||||
test_cfg:
|
||||
inference_on: teacher
|
||||
|
||||
|
||||
|
||||
metric: COCO
|
||||
num_classes: 80
|
||||
|
||||
# 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
|
||||
|
||||
EvalDataset:
|
||||
!COCODataSet
|
||||
image_dir: val2017
|
||||
anno_path: annotations/instances_val2017.json
|
||||
dataset_dir: dataset/coco
|
||||
allow_empty: true
|
||||
|
||||
TestDataset:
|
||||
!ImageFolder
|
||||
anno_path: annotations/instances_val2017.json # also support txt (like VOC's label_list.txt)
|
||||
dataset_dir: dataset/coco # if set, anno_path will be 'dataset_dir/anno_path'
|
||||
|
||||
epoch: 500 #epoch: 60
|
||||
|
||||
LearningRate:
|
||||
base_lr: 0.0002
|
||||
schedulers:
|
||||
- !PiecewiseDecay
|
||||
gamma: 1.0
|
||||
milestones: [500]
|
||||
use_warmup: false
|
||||
- !LinearWarmup
|
||||
start_factor: 0.001
|
||||
steps: 2000
|
||||
|
||||
OptimizerBuilder:
|
||||
clip_grad_by_norm: 0.1
|
||||
regularizer: false
|
||||
optimizer:
|
||||
type: AdamW
|
||||
weight_decay: 0.0001
|
||||
1
paddle_detection/configs/semi_det/semi_detr/README.md
Normal file
1
paddle_detection/configs/semi_det/semi_detr/README.md
Normal file
@@ -0,0 +1 @@
|
||||
## To be released
|
||||
Reference in New Issue
Block a user