Skip to content

GitLab

"test/vscode:/vscode.git/clone" did not exist on "a578d300ba036221894113c7ec0d415a8dbb319e"
Commit 522a602f authored by wangkx1's avatar wangkx1
Browse files

siton bug

parent abb99c90
Hide whitespace changes
Inline Side-by-side
Showing with 2100 additions and 0 deletions
configs/semi_det/baseline/ppyoloe_plus_crn_s_80e_coco_sup005.yml 0 → 100644
View file @ 522a602f
_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
configs/semi_det/baseline/ppyoloe_plus_crn_s_80e_coco_sup010.yml 0 → 100644
View file @ 522a602f
_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
configs/semi_det/baseline/retinanet_r50_fpn_2x_coco_sup005.yml 0 → 100644
View file @ 522a602f
_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
configs/semi_det/baseline/retinanet_r50_fpn_2x_coco_sup010.yml 0 → 100644
View file @ 522a602f
_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
configs/semi_det/denseteacher/README.md 0 → 100644
View file @ 522a602f
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}
}
```
configs/semi_det/denseteacher/denseteacher_fcos_r50_fpn_coco_full.yml 0 → 100644
View file @ 522a602f
_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
configs/semi_det/denseteacher/denseteacher_fcos_r50_fpn_coco_semi005.yml 0 → 100644
View file @ 522a602f
_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
configs/semi_det/denseteacher/denseteacher_fcos_r50_fpn_coco_semi010.yml 0 → 100644
View file @ 522a602f
_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
configs/semi_det/denseteacher/denseteacher_fcos_r50_fpn_coco_semi010_lsj.yml 0 → 100644
View file @ 522a602f
_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
configs/semi_det/denseteacher/denseteacher_ppyoloe_plus_crn_l_coco_semi005.yml 0 → 100644
View file @ 522a602f
_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
configs/semi_det/denseteacher/denseteacher_ppyoloe_plus_crn_l_coco_semi010.yml 0 → 100644
View file @ 522a602f
_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
configs/semi_det/denseteacher/denseteacher_ppyoloe_plus_crn_s_coco_semi005.yml 0 → 100644
View file @ 522a602f
_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
configs/semi_det/denseteacher/denseteacher_ppyoloe_plus_crn_s_coco_semi010.yml 0 → 100644
View file @ 522a602f
_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
configs/slim/README.md 0 → 100755
View file @ 522a602f
# 模型压缩
在PaddleDetection中, 提供了基于[PaddleSlim](https://github.com/PaddlePaddle/PaddleSlim)进行模型压缩的完整教程和benchmark。目前支持的方法：
- [剪裁](prune)
- [量化](quant)
- [离线量化](post_quant)
- [蒸馏](distill)
- [联合策略](extensions)
推荐您使用剪裁和蒸馏联合训练，或者使用剪裁、量化训练和离线量化，进行检测模型压缩。 下面以YOLOv3为例，进行剪裁、蒸馏和量化实验。
## 实验环境
- Python 3.7+
- PaddlePaddle >= 2.1.0
- PaddleSlim >= 2.1.0
- CUDA 10.1+
- cuDNN >=7.6.5
**PaddleDetection、 PaddlePaddle与PaddleSlim 版本关系:**
| PaddleDetection版本 | PaddlePaddle版本 | PaddleSlim版本 | 备注 |
| :------------------: | :---------------: | :-------: |:---------------: |
| release/2.3 | >= 2.1 | 2.1 | 离线量化依赖Paddle 2.2及PaddleSlim 2.2 |
| release/2.1 | 2.2 | >= 2.1.0 | 2.1 | 量化模型导出依赖最新Paddle develop分支，可在[PaddlePaddle每日版本](https://www.paddlepaddle.org.cn/documentation/docs/zh/install/Tables.html#whl-dev)中下载安装 |
| release/2.0 | >= 2.0.1 | 2.0 | 量化依赖Paddle 2.1及PaddleSlim 2.1 |
#### 安装PaddleSlim
- 方法一：直接安装：
```
pip install paddleslim -i https://pypi.tuna.tsinghua.edu.cn/simple
```
- 方法二：编译安装：
```
git clone https://github.com/PaddlePaddle/PaddleSlim.git
cd PaddleSlim
python setup.py install
```
## 快速开始
### 训练
```shell
python tools/train.py -c configs/{MODEL.yml} --slim_config configs/slim/{SLIM_CONFIG.yml}
```
- `-c`: 指定模型配置文件。
- `--slim_config`: 指定压缩策略配置文件。
- 如果选择使用蒸馏，具体蒸馏方法和更多检测模型的蒸馏，请参考[蒸馏策略文档](distill/README.md)
### 评估
```shell
python tools/eval.py -c configs/{MODEL.yml} --slim_config configs/slim/{SLIM_CONFIG.yml} -o weights=output/{SLIM_CONFIG}/model_final
```
- `-c`: 指定模型配置文件。
- `--slim_config`: 指定压缩策略配置文件。
- `-o weights`: 指定压缩算法训好的模型路径。
### 测试
```shell
python tools/infer.py -c configs/{MODEL.yml} --slim_config configs/slim/{SLIM_CONFIG.yml} \
-o weights=output/{SLIM_CONFIG}/model_final
--infer_img={IMAGE_PATH}
```
- `-c`: 指定模型配置文件。
- `--slim_config`: 指定压缩策略配置文件。
- `-o weights`: 指定压缩算法训好的模型路径。
- `--infer_img`: 指定测试图像路径。
## 全链条部署
### 动转静导出模型
```shell
python tools/export_model.py -c configs/{MODEL.yml} --slim_config configs/slim/{SLIM_CONFIG.yml} -o weights=output/{SLIM_CONFIG}/model_final
```
- `-c`: 指定模型配置文件。
- `--slim_config`: 指定压缩策略配置文件。
- `-o weights`: 指定压缩算法训好的模型路径。
### 部署预测
- Paddle-Inference预测：
- [Python部署](../../deploy/python/README.md)
- [C++部署](../../deploy/cpp/README.md)
- [TensorRT预测部署教程](../../deploy/TENSOR_RT.md)
- 服务器端部署：使用[PaddleServing](../../deploy/serving/README.md)部署。
- 手机移动端部署：使用[Paddle-Lite](../../deploy/lite/README.md) 在手机移动端部署。
## Benchmark
### 剪裁
#### Pascal VOC上benchmark
| 模型 | 压缩策略 | GFLOPs | 模型体积(MB) | 输入尺寸 | 预测时延(SD855) | Box AP | 下载 | 模型配置文件 | 压缩算法配置文件 |
| :---------: | :-------: | :------------: |:-------------: | :------: | :-------------: | :------: | :-----------------------------------------------------: |:-------------: | :------: |
| YOLOv3-MobileNetV1 | baseline | 24.13 | 93 | 608 | 332.0ms | 75.1 | [下载链接](https://paddledet.bj.bcebos.com/models/yolov3_mobilenet_v1_270e_voc.pdparams) | [配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/yolov3/yolov3_mobilenet_v1_270e_voc.yml) | - |
| YOLOv3-MobileNetV1 | 剪裁-l1_norm(sensity) | 15.78(-34.49%) | 66(-29%) | 608 | - | 78.4(+3.3) | [下载链接](https://paddledet.bj.bcebos.com/models/slim/yolov3_mobilenet_v1_voc_prune_l1_norm.pdparams) | [配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/yolov3/yolov3_mobilenet_v1_270e_voc.yml) | [slim配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/slim/prune/yolov3_prune_l1_norm.yml) |
#### COCO上benchmark
| 模型 | 压缩策略 | GFLOPs | 模型体积(MB) | 输入尺寸 | 预测时延(SD855) | Box AP | 下载 | 模型配置文件 | 压缩算法配置文件 |
| :---------: | :-------: | :------------: |:-------------: | :------: | :-------------: | :------: | :-----------------------------------------------------: |:-------------: | :------: |
| PP-YOLO-MobileNetV3_large | baseline | -- | 18.5 | 608 | 25.1ms | 23.2 | [下载链接](https://paddledet.bj.bcebos.com/models/ppyolo_mbv3_large_coco.pdparams) | [配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/ppyolo/ppyolo_mbv3_large_coco.yml) | - |
| PP-YOLO-MobileNetV3_large | 剪裁-FPGM | -37% | 12.6 | 608 | - | 22.3 | [下载链接](https://paddledet.bj.bcebos.com/models/slim/ppyolo_mbv3_large_prune_fpgm.pdparams) | [配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/ppyolo/ppyolo_mbv3_large_coco.yml) | [slim配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/slim/prune/ppyolo_mbv3_large_prune_fpgm.yml) |
| YOLOv3-DarkNet53 | baseline | -- | 238.2 | 608 | - | 39.0 | [下载链接](https://paddledet.bj.bcebos.com/models/ppyolo_mbv3_large_coco.pdparams) | [配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/yolov3/yolov3_darknet53_270e_coco.yml) | - |
| YOLOv3-DarkNet53 | 剪裁-FPGM | -24% | - | 608 | - | 37.6 | [下载链接](https://paddledet.bj.bcebos.com/models/slim/yolov3_darknet_prune_fpgm.pdparams) | [配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/yolov3/yolov3_darknet53_270e_coco.yml) | [slim配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/slim/prune/yolov3_darknet_prune_fpgm.yml) |
| PP-YOLO_R50vd | baseline | -- | 183.3 | 608 | - | 44.8 | [下载链接](https://paddledet.bj.bcebos.com/models/ppyolo_r50vd_dcn_1x_coco.pdparams) | [配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/ppyolo/ppyolo_r50vd_dcn_1x_coco.yml) | - |
| PP-YOLO_R50vd | 剪裁-FPGM | -35% | - | 608 | - | 42.1 | [下载链接](https://paddledet.bj.bcebos.com/models/slim/ppyolo_r50vd_prune_fpgm.pdparams) | [配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/ppyolo/ppyolo_r50vd_dcn_1x_coco.yml) | [slim配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/slim/prune/ppyolo_r50vd_prune_fpgm.yml) |
说明：
- 目前剪裁除RCNN系列模型外，其余模型均已支持。
- SD855预测时延为使用PaddleLite部署，使用arm8架构并使用4线程(4 Threads)推理时延。
### 量化
#### COCO上benchmark
| 模型 | 压缩策略 | 输入尺寸 | 模型体积(MB) | 预测时延(V100) | 预测时延(SD855) | Box AP | 下载 | Inference模型下载 | 模型配置文件 | 压缩算法配置文件 |
| ------------------ | ------------ | -------- | :---------: | :---------: |:---------: | :---------: | :----------------------------------------------: | :----------------------------------------------: |:------------------------------------------: | :------------------------------------: |
| PP-YOLOE-l | baseline | 640 | - | 11.2ms(trt_fp32) &#124; 7.7ms(trt_fp16) | -- | 50.9 | [下载链接](https://paddledet.bj.bcebos.com/models/ppyoloe_crn_l_300e_coco.pdparams) | - | [配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/ppyoloe/ppyoloe_crn_l_300e_coco.yml) | - |
| PP-YOLOE-l | 普通在线量化 | 640 | - | 6.7ms(trt_int8) | -- | 48.8 | [下载链接](https://paddledet.bj.bcebos.com/models/slim/ppyoloe_l_coco_qat.pdparams) | - | [配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/ppyoloe/ppyoloe_crn_l_300e_coco.yml) | [配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/slim/quant/ppyoloe_l_qat.yml) |
| PP-YOLOv2_R50vd | baseline | 640 | 208.6 | 19.1ms | -- | 49.1 | [下载链接](https://paddledet.bj.bcebos.com/models/ppyolov2_r50vd_dcn_365e_coco.pdparams) | [下载链接](https://paddledet.bj.bcebos.com/models/slim/ppyolov2_r50vd_dcn_365e_coco.tar) | [配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/ppyolo/ppyolo_r50vd_dcn_1x_coco.yml) | - |
| PP-YOLOv2_R50vd | PACT在线量化 | 640 | -- | 17.3ms | -- | 48.1 | [下载链接](https://paddledet.bj.bcebos.com/models/slim/ppyolov2_r50vd_dcn_qat.pdparams) | [下载链接](https://paddledet.bj.bcebos.com/models/slim/ppyolov2_r50vd_dcn_qat.tar) | [配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/ppyolo/ppyolo_r50vd_dcn_1x_coco.yml) | [配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/slim/quant/ppyolov2_r50vd_dcn_qat.yml) |
| PP-YOLO_R50vd | baseline | 608 | 183.3 | 17.4ms | -- | 44.8 | [下载链接](https://paddledet.bj.bcebos.com/models/ppyolo_r50vd_dcn_1x_coco.pdparams) | [下载链接](https://paddledet.bj.bcebos.com/models/slim/ppyolo_r50vd_dcn_1x_coco.tar) | [配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/ppyolo/ppyolo_r50vd_dcn_1x_coco.yml) | - |
| PP-YOLO_R50vd | PACT在线量化 | 608 | 67.3 | 13.8ms | -- | 44.3 | [下载链接](https://paddledet.bj.bcebos.com/models/slim/ppyolo_r50vd_qat_pact.pdparams) | [下载链接](https://paddledet.bj.bcebos.com/models/slim/ppyolo_r50vd_qat_pact.tar) | [配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/ppyolo/ppyolo_r50vd_dcn_1x_coco.yml) | [配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/slim/quant/ppyolo_r50vd_qat_pact.yml) |
| PP-YOLO-MobileNetV3_large | baseline | 320 | 18.5 | 2.7ms | 27.9ms | 23.2 | [下载链接](https://paddledet.bj.bcebos.com/models/ppyolo_mbv3_large_coco.pdparams) | [下载链接](https://paddledet.bj.bcebos.com/models/slim/ppyolo_mbv3_large_coco.tar) | [配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/ppyolo/ppyolo_mbv3_large_coco.yml) | - |
| PP-YOLO-MobileNetV3_large | 普通在线量化 | 320 | 5.6 | -- | 25.1ms | 24.3 | [下载链接](https://paddledet.bj.bcebos.com/models/slim/ppyolo_mbv3_large_qat.pdparams) | [下载链接](https://paddledet.bj.bcebos.com/models/slim/ppyolo_mbv3_large_qat.tar) | [配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/ppyolo/ppyolo_mbv3_large_coco.yml) | [配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/slim/quant/ppyolo_mbv3_large_qat.yml) |
| YOLOv3-MobileNetV1 | baseline | 608 | 94.2 | 8.9ms | 332ms | 29.4 | [下载链接](https://paddledet.bj.bcebos.com/models/yolov3_mobilenet_v1_270e_coco.pdparams) | [下载链接](https://paddledet.bj.bcebos.com/models/slim/yolov3_mobilenet_v1_270e_coco.tar) | [配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/yolov3/yolov3_mobilenet_v1_270e_coco.yml) | - |
| YOLOv3-MobileNetV1 | 普通在线量化 | 608 | 25.4 | 6.6ms | 248ms | 30.5 | [下载链接](https://paddledet.bj.bcebos.com/models/slim/yolov3_mobilenet_v1_coco_qat.pdparams) | [下载链接](https://paddledet.bj.bcebos.com/models/slim/yolov3_mobilenet_v1_coco_qat.tar) | [配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/yolov3/yolov3_mobilenet_v1_270e_coco.yml) | [slim配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/slim/quant/yolov3_mobilenet_v1_qat.yml) |
| YOLOv3-MobileNetV3 | baseline | 608 | 90.3 | 9.4ms | 367.2ms | 31.4 | [下载链接](https://paddledet.bj.bcebos.com/models/yolov3_mobilenet_v3_large_270e_coco.pdparams) | [下载链接](https://paddledet.bj.bcebos.com/models/slim/yolov3_mobilenet_v3_large_270e_coco.tar) | [配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/yolov3/yolov3_mobilenet_v3_large_270e_coco.yml) | - |
| YOLOv3-MobileNetV3 | PACT在线量化 | 608 | 24.4 | 8.0ms | 280.0ms | 31.1 | [下载链接](https://paddledet.bj.bcebos.com/models/slim/yolov3_mobilenet_v3_coco_qat.pdparams) | [下载链接](https://paddledet.bj.bcebos.com/models/slim/yolov3_mobilenet_v3_coco_qat.tar) | [配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/yolov3/yolov3_mobilenet_v3_large_270e_coco.yml) | [slim配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/slim/quant/yolov3_mobilenet_v3_qat.yml) |
| YOLOv3-DarkNet53 | baseline | 608 | 238.2 | 16.0ms | -- | 39.0 | [下载链接](https://paddledet.bj.bcebos.com/models/yolov3_darknet53_270e_coco.pdparams) | [下载链接](https://paddledet.bj.bcebos.com/models/slim/yolov3_darknet53_270e_coco.tar) | [配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/yolov3/yolov3_darknet53_270e_coco.yml) | - |
| YOLOv3-DarkNet53 | 普通在线量化 | 608 | 78.8 | 12.4ms | -- | 38.8 | [下载链接](https://paddledet.bj.bcebos.com/models/slim/yolov3_darknet_coco_qat.pdparams) | [下载链接](https://paddledet.bj.bcebos.com/models/slim/yolov3_darknet_coco_qat.tar) | [配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/yolov3/yolov3_darknet53_270e_coco.yml) | [slim配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/slim/quant/yolov3_darknet_qat.yml) |
| SSD-MobileNet_v1 | baseline | 300 | 22.5 | 4.4ms | 26.6ms | 73.8 | [下载链接](https://paddledet.bj.bcebos.com/models/ssd_mobilenet_v1_300_120e_voc.pdparams) | [下载链接](https://paddledet.bj.bcebos.com/models/slim/ssd_mobilenet_v1_300_120e_voc.tar) |[配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/ssd/ssd_mobilenet_v1_300_120e_voc.yml) | - |
| SSD-MobileNet_v1 | 普通在线量化 | 300 | 7.1 | -- | 21.5ms | 72.9 | [下载链接](https://paddledet.bj.bcebos.com/models/slim/ssd_mobilenet_v1_300_voc_qat.pdparams) | [下载链接](https://paddledet.bj.bcebos.com/models/slim/ssd_mobilenet_v1_300_voc_qat.tar) | [配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/ssd/ssd_mobilenet_v1_300_120e_voc.yml) | [slim配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/slim/quant/ssd_mobilenet_v1_qat.yml) |
| Mask-ResNet50-FPN | baseline | (800, 1333) | 174.1 | 359.5ms | -- | 39.2/35.6 | [下载链接](https://paddledet.bj.bcebos.com/models/mask_rcnn_r50_fpn_1x_coco.pdparams) | [下载链接](https://paddledet.bj.bcebos.com/models/slim/mask_rcnn_r50_fpn_1x_coco.tar) |[配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/mask_rcnn/mask_rcnn_r50_fpn_1x_coco.yml) | - |
| Mask-ResNet50-FPN | 普通在线量化 | (800, 1333) | -- | -- | -- | 39.7(+0.5)/35.9(+0.3) | [下载链接](https://paddledet.bj.bcebos.com/models/slim/mask_rcnn_r50_fpn_1x_qat.pdparams) | [下载链接](https://paddledet.bj.bcebos.com/models/slim/mask_rcnn_r50_fpn_1x_qat.tar) | [配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/mask_rcnn/mask_rcnn_r50_fpn_1x_coco.yml) | [slim配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/slim/quant/mask_rcnn_r50_fpn_1x_qat.yml) |
说明：
- 上述V100预测时延非量化模型均是使用TensorRT-FP32测试，量化模型均使用TensorRT-INT8测试，并且都包含NMS耗时。
- SD855预测时延为使用PaddleLite部署，使用arm8架构并使用4线程(4 Threads)推理时延。
- 上述PP-YOLOE模型均在V100，开启TensorRT环境中测速，不包含NMS。（导出模型时指定：-o trt=True exclude_nms=True）
### 离线量化
需要准备val集，用来对离线量化模型进行校准，运行方式：
```shell
python tools/post_quant.py -c configs/{MODEL.yml} --slim_config configs/slim/post_quant/{SLIM_CONFIG.yml}
```
例如：
```shell
python3.7 tools/post_quant.py -c configs/ppyolo/ppyolo_mbv3_large_coco.yml --slim_config=configs/slim/post_quant/ppyolo_mbv3_large_ptq.yml
```
### 蒸馏
#### COCO上benchmark
| 模型 | 压缩策略 | 输入尺寸 | Box AP | 下载 | 模型配置文件 | 压缩算法配置文件 |
| ------------------ | ------------ | -------- | :---------: | :----------------------------------------------------------: | :----------------------------------------------------------: | :----------------------------------------------------------: |
| YOLOv3-MobileNetV1 | baseline | 608 | 29.4 | [下载链接](https://paddledet.bj.bcebos.com/models/yolov3_mobilenet_v1_270e_coco.pdparams) | [配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/yolov3/yolov3_mobilenet_v1_270e_coco.yml) | - |
| YOLOv3-MobileNetV1 | 蒸馏 | 608 | 31.0(+1.6) | [下载链接](https://paddledet.bj.bcebos.com/models/slim/yolov3_mobilenet_v1_coco_distill.pdparams) | [配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/yolov3/yolov3_mobilenet_v1_270e_coco.yml) | [slim配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/slim/distill/yolov3_mobilenet_v1_coco_distill.yml) |
- 具体蒸馏方法和更多检测模型的蒸馏，请参考[蒸馏策略文档](distill/README.md)
### 蒸馏剪裁联合策略
#### COCO上benchmark
| 模型 | 压缩策略 | 输入尺寸 | GFLOPs | 模型体积(MB) | 预测时延(SD855) | Box AP | 下载 | 模型配置文件 | 压缩算法配置文件 |
| ------------------ | ------------ | -------- | :---------: |:---------: |:---------: | :---------: |:----------------------------------------------------------: | :----------------------------------------------------------: | :----------------------------------------------------------: |
| YOLOv3-MobileNetV1 | baseline | 608 | 24.65 | 94.2 | 332.0ms | 29.4 | [下载链接](https://paddledet.bj.bcebos.com/models/yolov3_mobilenet_v1_270e_coco.pdparams) | [配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/yolov3/yolov3_mobilenet_v1_270e_coco.yml) | - |
| YOLOv3-MobileNetV1 | 蒸馏+剪裁 | 608 | 7.54(-69.4%) | 30.9(-67.2%) | 166.1ms | 28.4(-1.0) | [下载链接](https://paddledet.bj.bcebos.com/models/slim/yolov3_mobilenet_v1_coco_distill_prune.pdparams) | [配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/yolov3/yolov3_mobilenet_v1_270e_coco.yml) | [slim配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/slim/extensions/yolov3_mobilenet_v1_coco_distill_prune.yml) |
| YOLOv3-MobileNetV1 | 剪裁+量化 | 608 | - | - | - | - | - | [配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/yolov3/yolov3_mobilenet_v1_270e_voc.yml) | [slim配置文件](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/slim/extensions/yolov3_mobilenetv1_prune_qat.yml) |
configs/slim/README_en.md 0 → 100755
View file @ 522a602f
# Model Compression
In PaddleDetection, a complete tutorial and benchmarks for model compression based on [PaddleSlim](https://github.com/PaddlePaddle/PaddleSlim) are provided. Currently supported methods:
- [pruning](prune)
- [quantitative](quant)
- [distillation](distill)
- [The joint strategy](extensions)
It is recommended that you use a combination of pruning and distillation training, or use pruning and quantization for test model compression. The following takes YOLOv3 as an example to carry out cutting, distillation and quantization experiments.
## Experimental Environment
- Python 3.7+
- PaddlePaddle >= 2.1.0
- PaddleSlim >= 2.1.0
- CUDA 10.1+
- cuDNN >=7.6.5
**Version Dependency between PaddleDetection, Paddle and PaddleSlim Version**
| PaddleDetection Version | PaddlePaddle Version | PaddleSlim Version | Note |
| :---------------------: | :------------------: | :----------------: | :----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------: |
| release/2.1 | >= 2.1.0 | 2.1 | Quantitative model exports rely on the latest Paddle Develop branch, available in[PaddlePaddle Daily version](https://www.paddlepaddle.org.cn/documentation/docs/zh/install/Tables.html#whl-dev) |
| release/2.0 | >= 2.0.1 | 2.0 | Quantization depends on Paddle 2.1 and PaddleSlim 2.1 |
#### Install PaddleSlim
- Method 1: Install it directly：
```
pip install paddleslim -i https://pypi.tuna.tsinghua.edu.cn/simple
```
- Method 2: Compile and install：
```
git clone https://github.com/PaddlePaddle/PaddleSlim.git
cd PaddleSlim
python setup.py install
```
## Quick Start
### Train
```shell
python tools/train.py -c configs/{MODEL.yml} --slim_config configs/slim/{SLIM_CONFIG.yml}
```
- `-c`: Specify the model configuration file.
- `--slim_config`: Specify the compression policy profile.
- If you want to use distillation, please refer to [Distillation Doc](distill/README.md) for specific distillation methods and more distillation of detection models.
### Evaluation
```shell
python tools/eval.py -c configs/{MODEL.yml} --slim_config configs/slim/{SLIM_CONFIG.yml} -o weights=output/{SLIM_CONFIG}/model_final
```
- `-c`: Specify the model configuration file.
- `--slim_config`: Specify the compression policy profile.
- `-o weights`: Specifies the path of the model trained by the compression algorithm.
### Test
```shell
python tools/infer.py -c configs/{MODEL.yml} --slim_config configs/slim/{SLIM_CONFIG.yml} \
-o weights=output/{SLIM_CONFIG}/model_final
--infer_img={IMAGE_PATH}
```
- `-c`: Specify the model configuration file.
- `--slim_config`: Specify the compression policy profile.
- `-o weights`: Specifies the path of the model trained by the compression algorithm.
- `--infer_img`: Specifies the test image path.
## Full Chain Deployment
### the model is derived from moving to static
```shell
python tools/export_model.py -c configs/{MODEL.yml} --slim_config configs/slim/{SLIM_CONFIG.yml} -o weights=output/{SLIM_CONFIG}/model_final
```
- `-c`: Specify the model configuration file.
- `--slim_config`: Specify the compression policy profile.
- `-o weights`: Specifies the path of the model trained by the compression algorithm.
### prediction and deployment
- Paddle-Inference Prediction：
- [Python Deployment](../../deploy/python/README.md)
- [C++ Deployment](../../deploy/cpp/README.md)
- [TensorRT Predictive Deployment Tutorial](../../deploy/TENSOR_RT.md)
- Server deployment: Used[PaddleServing](../../deploy/serving/README.md)
- Mobile deployment: Use[Paddle-Lite](../../deploy/lite/README.md) Deploy it on the mobile terminal.
## Benchmark
### Pruning
#### Pascal VOC Benchmark
| Model | Compression Strategy | GFLOPs | Model Volume(MB) | Input Size | Predict Delay(SD855) | Box AP | Download | Model Configuration File | Compression Algorithm Configuration File |
| :----------------: | :-------------------: | :------------: | :--------------: | :--------: | :------------------: | :--------: | :------------------------------------------------------------------------------------------------: | :--------------------------------------------------------------------------------------------------------------------------------: | :---------------------------------------------------------------------------------------------------------------------------------: |
| YOLOv3-MobileNetV1 | baseline | 24.13 | 93 | 608 | 332.0ms | 75.1 | [link](https://paddledet.bj.bcebos.com/models/yolov3_mobilenet_v1_270e_voc.pdparams) | [configuration file](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/yolov3/yolov3_mobilenet_v1_270e_voc.yml) | - |
| YOLOv3-MobileNetV1 | 剪裁-l1_norm(sensity) | 15.78(-34.49%) | 66(-29%) | 608 | - | 78.4(+3.3) | [link](https://paddledet.bj.bcebos.com/models/slim/yolov3_mobilenet_v1_voc_prune_l1_norm.pdparams) | [configuration file](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/yolov3/yolov3_mobilenet_v1_270e_voc.yml) | [slim configuration file](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/slim/prune/yolov3_prune_l1_norm.yml) |
#### COCO Benchmark
| Mode | Compression Strategy | GFLOPs | Model Volume(MB) | Input Size | Predict Delay(SD855) | Box AP | Download | Model Configuration File | Compression Algorithm Configuration File |
| :-----------------------: | :------------------: | :----: | :--------------: | :--------: | :------------------: | :----: | :---------------------------------------------------------------------------------------: | :------------------------------------------------------------------------------------------------------------------------------: | :-----------------------------------------------------------------------------------------------------------------------------------------: |
| PP-YOLO-MobileNetV3_large | baseline | -- | 18.5 | 608 | 25.1ms | 23.2 | [link](https://paddledet.bj.bcebos.com/models/ppyolo_mbv3_large_coco.pdparams) | [configuration file](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/ppyolo/ppyolo_mbv3_large_coco.yml) | - |
| PP-YOLO-MobileNetV3_large | 剪裁-FPGM | -37% | 12.6 | 608 | - | 22.3 | [link](https://paddledet.bj.bcebos.com/models/slim/ppyolo_mbv3_large_prune_fpgm.pdparams) | [configuration file](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/ppyolo/ppyolo_mbv3_large_coco.yml) | [slim configuration file](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/slim/prune/ppyolo_mbv3_large_prune_fpgm.yml) |
| YOLOv3-DarkNet53 | baseline | -- | 238.2 | 608 | - | 39.0 | [link](https://paddledet.bj.bcebos.com/models/ppyolo_mbv3_large_coco.pdparams) | [configuration file](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/yolov3/yolov3_darknet53_270e_coco.yml) | - |
| YOLOv3-DarkNet53 | 剪裁-FPGM | -24% | - | 608 | - | 37.6 | [link](https://paddledet.bj.bcebos.com/models/slim/yolov3_darknet_prune_fpgm.pdparams) | [configuration file](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/yolov3/yolov3_darknet53_270e_coco.yml) | [slim configuration file](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/slim/prune/yolov3_darknet_prune_fpgm.yml) |
| PP-YOLO_R50vd | baseline | -- | 183.3 | 608 | - | 44.8 | [link](https://paddledet.bj.bcebos.com/models/ppyolo_r50vd_dcn_1x_coco.pdparams) | [configuration file](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/ppyolo/ppyolo_r50vd_dcn_1x_coco.yml) | - |
| PP-YOLO_R50vd | 剪裁-FPGM | -35% | - | 608 | - | 42.1 | [link](https://paddledet.bj.bcebos.com/models/slim/ppyolo_r50vd_prune_fpgm.pdparams) | [configuration file](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/ppyolo/ppyolo_r50vd_dcn_1x_coco.yml) | [slim configuration file](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/slim/prune/ppyolo_r50vd_prune_fpgm.yml) |
Description:
- Currently, all models except RCNN series models are supported.
- The SD855 predicts the delay for deployment using Paddle Lite, using the ARM8 architecture and using 4 Threads (4 Threads) to reason the delay.
### Quantitative
#### COCO Benchmark
| Model | Compression Strategy | Input Size | Model Volume(MB) | Prediction Delay(V100) | Prediction Delay(SD855) | Box AP | Download | Download of Inference Model | Model Configuration File | Compression Algorithm Configuration File |
| ------------------------- | -------------------------- | ----------- | :--------------: | :--------------------: | :---------------------: | :-------------------: | :-----------------------------------------------------------------------------------------: | :-----------------------------------------------------------------------------------------: | :----------------------------------------------------------------------------------------------------------------------------------------: | :--------------------------------------------------------------------------------------------------------------------------------------: |
| PP-YOLOE-l | baseline | 640 | - | 11.2ms(trt_fp32) &#124; 7.7ms(trt_fp16) | -- | 50.9 | [link](https://paddledet.bj.bcebos.com/models/ppyoloe_crn_l_300e_coco.pdparams) | - | [Configuration File](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/ppyoloe/ppyoloe_crn_l_300e_coco.yml) | - |
| PP-YOLOE-l | Common Online quantitative | 640 | - | 6.7ms(trt_int8) | -- | 48.8 | [link](https://paddledet.bj.bcebos.com/models/slim/ppyoloe_l_coco_qat.pdparams) | - | [Configuration File](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/ppyoloe/ppyoloe_crn_l_300e_coco.yml) | [Configuration File](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/slim/quant/ppyoloe_l_qat.yml) |
| PP-YOLOv2_R50vd | baseline | 640 | 208.6 | 19.1ms | -- | 49.1 | [link](https://paddledet.bj.bcebos.com/models/ppyolov2_r50vd_dcn_365e_coco.pdparams) | [link](https://paddledet.bj.bcebos.com/models/slim/ppyolov2_r50vd_dcn_365e_coco.tar) | [Configuration File](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/ppyolo/ppyolo_r50vd_dcn_1x_coco.yml) | - |
| PP-YOLOv2_R50vd | PACT Online quantitative | 640 | -- | 17.3ms | -- | 48.1 | [link](https://paddledet.bj.bcebos.com/models/slim/ppyolov2_r50vd_dcn_qat.pdparams) | [link](https://paddledet.bj.bcebos.com/models/slim/ppyolov2_r50vd_dcn_qat.tar) | [Configuration File ](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/ppyolo/ppyolo_r50vd_dcn_1x_coco.yml) | [Configuration File ](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/slim/quant/ppyolov2_r50vd_dcn_qat.yml) |
| PP-YOLO_R50vd | baseline | 608 | 183.3 | 17.4ms | -- | 44.8 | [link](https://paddledet.bj.bcebos.com/models/ppyolo_r50vd_dcn_1x_coco.pdparams) | [link](https://paddledet.bj.bcebos.com/models/slim/ppyolo_r50vd_dcn_1x_coco.tar) | [Configuration File ](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/ppyolo/ppyolo_r50vd_dcn_1x_coco.yml) | - |
| PP-YOLO_R50vd | PACT Online quantitative | 608 | 67.3 | 13.8ms | -- | 44.3 | [link](https://paddledet.bj.bcebos.com/models/slim/ppyolo_r50vd_qat_pact.pdparams) | [link](https://paddledet.bj.bcebos.com/models/slim/ppyolo_r50vd_qat_pact.tar) | [Configuration File ](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/ppyolo/ppyolo_r50vd_dcn_1x_coco.yml) | [Configuration File ](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/slim/quant/ppyolo_r50vd_qat_pact.yml) |
| PP-YOLO-MobileNetV3_large | baseline | 320 | 18.5 | 2.7ms | 27.9ms | 23.2 | [link](https://paddledet.bj.bcebos.com/models/ppyolo_mbv3_large_coco.pdparams) | [link](https://paddledet.bj.bcebos.com/models/slim/ppyolo_mbv3_large_coco.tar) | [Configuration File ](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/ppyolo/ppyolo_mbv3_large_coco.yml) | - |
| PP-YOLO-MobileNetV3_large | Common Online quantitative | 320 | 5.6 | -- | 25.1ms | 24.3 | [link](https://paddledet.bj.bcebos.com/models/slim/ppyolo_mbv3_large_qat.pdparams) | [link](https://paddledet.bj.bcebos.com/models/slim/ppyolo_mbv3_large_qat.tar) | [Configuration File ](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/ppyolo/ppyolo_mbv3_large_coco.yml) | [Configuration File ](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/slim/quant/ppyolo_mbv3_large_qat.yml) |
| YOLOv3-MobileNetV1 | baseline | 608 | 94.2 | 8.9ms | 332ms | 29.4 | [link](https://paddledet.bj.bcebos.com/models/yolov3_mobilenet_v1_270e_coco.pdparams) | [link](https://paddledet.bj.bcebos.com/models/slim/yolov3_mobilenet_v1_270e_coco.tar) | [Configuration File ](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/yolov3/yolov3_mobilenet_v1_270e_coco.yml) | - |
| YOLOv3-MobileNetV1 | Common Online quantitative | 608 | 25.4 | 6.6ms | 248ms | 30.5 | [link](https://paddledet.bj.bcebos.com/models/slim/yolov3_mobilenet_v1_coco_qat.pdparams) | [link](https://paddledet.bj.bcebos.com/models/slim/yolov3_mobilenet_v1_coco_qat.tar) | [Configuration File ](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/yolov3/yolov3_mobilenet_v1_270e_coco.yml) | [slim Configuration File ](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/slim/quant/yolov3_mobilenet_v1_qat.yml) |
| YOLOv3-MobileNetV3 | baseline | 608 | 90.3 | 9.4ms | 367.2ms | 31.4 | [link](https://paddledet.bj.bcebos.com/models/yolov3_mobilenet_v3_large_270e_coco.pdparams) | [link](https://paddledet.bj.bcebos.com/models/slim/yolov3_mobilenet_v3_large_270e_coco.tar) | [Configuration File ](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/yolov3/yolov3_mobilenet_v3_large_270e_coco.yml) | - |
| YOLOv3-MobileNetV3 | PACT Online quantitative | 608 | 24.4 | 8.0ms | 280.0ms | 31.1 | [link](https://paddledet.bj.bcebos.com/models/slim/yolov3_mobilenet_v3_coco_qat.pdparams) | [link](https://paddledet.bj.bcebos.com/models/slim/yolov3_mobilenet_v3_coco_qat.tar) | [Configuration File ](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/yolov3/yolov3_mobilenet_v3_large_270e_coco.yml) | [slim Configuration File ](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/slim/quant/yolov3_mobilenet_v3_qat.yml) |
| YOLOv3-DarkNet53 | baseline | 608 | 238.2 | 16.0ms | -- | 39.0 | [link](https://paddledet.bj.bcebos.com/models/yolov3_darknet53_270e_coco.pdparams) | [link](https://paddledet.bj.bcebos.com/models/slim/yolov3_darknet53_270e_coco.tar) | [Configuration File ](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/yolov3/yolov3_darknet53_270e_coco.yml) | - |
| YOLOv3-DarkNet53 | Common Online quantitative | 608 | 78.8 | 12.4ms | -- | 38.8 | [link](https://paddledet.bj.bcebos.com/models/slim/yolov3_darknet_coco_qat.pdparams) | [link](https://paddledet.bj.bcebos.com/models/slim/yolov3_darknet_coco_qat.tar) | [Configuration File ](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/yolov3/yolov3_darknet53_270e_coco.yml) | [slim Configuration File ](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/slim/quant/yolov3_darknet_qat.yml) |
| SSD-MobileNet_v1 | baseline | 300 | 22.5 | 4.4ms | 26.6ms | 73.8 | [link](https://paddledet.bj.bcebos.com/models/ssd_mobilenet_v1_300_120e_voc.pdparams) | [link](https://paddledet.bj.bcebos.com/models/slim/ssd_mobilenet_v1_300_120e_voc.tar) | [Configuration File ](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/ssd/ssd_mobilenet_v1_300_120e_voc.yml) | - |
| SSD-MobileNet_v1 | Common Online quantitative | 300 | 7.1 | -- | 21.5ms | 72.9 | [link](https://paddledet.bj.bcebos.com/models/slim/ssd_mobilenet_v1_300_voc_qat.pdparams) | [link](https://paddledet.bj.bcebos.com/models/slim/ssd_mobilenet_v1_300_voc_qat.tar) | [Configuration File ](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/ssd/ssd_mobilenet_v1_300_120e_voc.yml) | [slim Configuration File ](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/slim/quant/ssd_mobilenet_v1_qat.yml) |
| Mask-ResNet50-FPN | baseline | (800, 1333) | 174.1 | 359.5ms | -- | 39.2/35.6 | [link](https://paddledet.bj.bcebos.com/models/mask_rcnn_r50_fpn_1x_coco.pdparams) | [link](https://paddledet.bj.bcebos.com/models/slim/mask_rcnn_r50_fpn_1x_coco.tar) | [Configuration File ](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/mask_rcnn/mask_rcnn_r50_fpn_1x_coco.yml) | - |
| Mask-ResNet50-FPN | Common Online quantitative | (800, 1333) | -- | -- | -- | 39.7(+0.5)/35.9(+0.3) | [link](https://paddledet.bj.bcebos.com/models/slim/mask_rcnn_r50_fpn_1x_qat.pdparams) | [link](https://paddledet.bj.bcebos.com/models/slim/mask_rcnn_r50_fpn_1x_qat.tar) | [Configuration File ](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/mask_rcnn/mask_rcnn_r50_fpn_1x_coco.yml) | [slim Configuration File ](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/slim/quant/mask_rcnn_r50_fpn_1x_qat.yml) |
Description:
- The above V100 prediction delay non-quantified model is tested by TensorRT FP32, and the quantified model is tested by TensorRT INT8, and both of them include NMS time.
- The SD855 predicts the delay for deployment using PaddleLite, using the ARM8 architecture and using 4 Threads (4 Threads) to reason the delay.
### Distillation
#### COCO Benchmark
| Model | Compression Strategy | Input Size | Box AP | Download | Model Configuration File | Compression Strategy Configuration File |
| ------------------ | -------------------- | ---------- | :--------: | :-------------------------------------------------------------------------------------------: | :----------------------------------------------------------------------------------------------------------------------------------: | :-----------------------------------------------------------------------------------------------------------------------------------------------: |
| YOLOv3-MobileNetV1 | baseline | 608 | 29.4 | [link](https://paddledet.bj.bcebos.com/models/yolov3_mobilenet_v1_270e_coco.pdparams) | [Configuration File ](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/yolov3/yolov3_mobilenet_v1_270e_coco.yml) | - |
| YOLOv3-MobileNetV1 | Distillation | 608 | 31.0(+1.6) | [link](https://paddledet.bj.bcebos.com/models/slim/yolov3_mobilenet_v1_coco_distill.pdparams) | [Configuration File ](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/yolov3/yolov3_mobilenet_v1_270e_coco.yml) | [slimConfiguration File ](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/slim/distill/yolov3_mobilenet_v1_coco_distill.yml) |
- For the specific distillation method and more distillation detection models, please refer to [distill](distill/README.md).
### Distillation Pruning Combined Strategy
#### COCO Benchmark
| Model | Compression Strategy | Input Size | GFLOPs | Model Volume(MB) | Prediction Delay(SD855) | Box AP | Download | Model Configuration File | Compression Algorithm Configuration File |
| ------------------ | ------------------------ | ---------- | :----------: | :--------------: | :---------------------: | :--------: | :-------------------------------------------------------------------------------------------------: | :----------------------------------------------------------------------------------------------------------------------------------: | :--------------------------------------------------------------------------------------------------------------------------------------------------------: |
| YOLOv3-MobileNetV1 | baseline | 608 | 24.65 | 94.2 | 332.0ms | 29.4 | [link](https://paddledet.bj.bcebos.com/models/yolov3_mobilenet_v1_270e_coco.pdparams) | [Configuration File ](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/yolov3/yolov3_mobilenet_v1_270e_coco.yml) | - |
| YOLOv3-MobileNetV1 | Distillation + Tailoring | 608 | 7.54(-69.4%) | 30.9(-67.2%) | 166.1ms | 28.4(-1.0) | [link](https://paddledet.bj.bcebos.com/models/slim/yolov3_mobilenet_v1_coco_distill_prune.pdparams) | [Configuration File ](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/yolov3/yolov3_mobilenet_v1_270e_coco.yml) | [slimConfiguration File ](https://github.com/PaddlePaddle/PaddleDetection/tree/develop/configs/slim/extensions/yolov3_mobilenet_v1_coco_distill_prune.yml) |
configs/slim/distill/README.md 0 → 100644
View file @ 522a602f
# Distillation(蒸馏)
## 内容
- [YOLOv3模型蒸馏](#YOLOv3模型蒸馏)
- [FGD模型蒸馏](#FGD模型蒸馏)
- [CWD模型蒸馏](#CWD模型蒸馏)
- [LD模型蒸馏](#LD模型蒸馏)
- [PPYOLOE模型蒸馏](#PPYOLOE模型蒸馏)
- [引用](#引用)
## YOLOv3模型蒸馏
以YOLOv3-MobileNetV1为例，使用YOLOv3-ResNet34作为蒸馏训练的teacher网络, 对YOLOv3-MobileNetV1结构的student网络进行蒸馏。
COCO数据集作为目标检测任务的训练目标难度更大，意味着teacher网络会预测出更多的背景bbox，如果直接用teacher的预测输出作为student学习的`soft label`会有严重的类别不均衡问题。解决这个问题需要引入新的方法，详细背景请参考论文:[Object detection at 200 Frames Per Second](https://arxiv.org/abs/1805.06361)。
为了确定蒸馏的对象，我们首先需要找到student和teacher网络得到的`x,y,w,h,cls,objectness`等Tensor，用teacher得到的结果指导student训练。具体实现可参考[代码](../../../ppdet/slim/distill_loss.py)
| 模型 | 方案 | 输入尺寸 | epochs | Box mAP | 配置文件 | 下载链接 |
| :---------------: | :---------: | :----: | :----: |:-----------: | :--------------: | :------------: |
| YOLOv3-ResNet34 | teacher | 608 | 270e | 36.2 | [config](../../yolov3/yolov3_r34_270e_coco.yml) | [download](https://paddledet.bj.bcebos.com/models/yolov3_r34_270e_coco.pdparams) |
| YOLOv3-MobileNetV1 | student | 608 | 270e | 29.4 | [config](../../yolov3/yolov3_mobilenet_v1_270e_coco.yml) | [download](https://paddledet.bj.bcebos.com/models/yolov3_mobilenet_v1_270e_coco.pdparams) |
| YOLOv3-MobileNetV1 | distill | 608 | 270e | 31.0(+1.6) | [config](../../yolov3/yolov3_mobilenet_v1_270e_coco.yml),[slim_config](./yolov3_mobilenet_v1_coco_distill.yml) | [download](https://paddledet.bj.bcebos.com/models/slim/yolov3_mobilenet_v1_coco_distill.pdparams) |
<details>
<summary> 快速开始 </summary>
```shell
# 单卡训练(不推荐)
python tools/train.py -c configs/yolov3/yolov3_mobilenet_v1_270e_coco.yml --slim_config configs/slim/distill/yolov3_mobilenet_v1_coco_distill.yml
# 多卡训练
python -m paddle.distributed.launch --log_dir=logs/ --gpus 0,1,2,3,4,5,6,7 tools/train.py -c configs/yolov3/yolov3_mobilenet_v1_270e_coco.yml --slim_config configs/slim/distill/yolov3_mobilenet_v1_coco_distill.yml
# 评估
python tools/eval.py -c configs/yolov3/yolov3_mobilenet_v1_270e_coco.yml -o weights=https://paddledet.bj.bcebos.com/models/slim/yolov3_mobilenet_v1_coco_distill.pdparams
# 预测
python tools/infer.py -c configs/yolov3/yolov3_mobilenet_v1_270e_coco.yml -o weights=https://paddledet.bj.bcebos.com/models/slim/yolov3_mobilenet_v1_coco_distill.pdparams --infer_img=demo/000000014439_640x640.jpg
```
- `-c`: 指定模型配置文件，也是student配置文件。
- `--slim_config`: 指定压缩策略配置文件，也是teacher配置文件。
</details>
## FGD模型蒸馏
FGD全称为[Focal and Global Knowledge Distillation for Detectors](https://arxiv.org/abs/2111.11837v1)，是目标检测任务的一种蒸馏方法，FGD蒸馏分为两个部分`Focal`和`Global`。`Focal`蒸馏分离图像的前景和背景，让学生模型分别关注教师模型的前景和背景部分特征的关键像素；`Global`蒸馏部分重建不同像素之间的关系并将其从教师转移到学生，以补偿`Focal`蒸馏中丢失的全局信息。试验结果表明，FGD蒸馏算法在基于anchor和anchor free的方法上能有效提升模型精度。
在PaddleDetection中，我们实现了FGD算法，并基于RetinaNet算法进行验证，实验结果如下：
| 模型 | 方案 | 输入尺寸 | epochs | Box mAP | 配置文件 | 下载链接 |
| ----------------- | ----------- | ------ | :----: | :-----------: | :--------------: | :------------: |
| RetinaNet-ResNet101| teacher | 1333x800 | 2x | 40.6 | [config](../../retinanet/retinanet_r101_fpn_2x_coco.yml) | [download](https://paddledet.bj.bcebos.com/models/retinanet_r101_fpn_2x_coco.pdparams) |
| RetinaNet-ResNet50 | student | 1333x800 | 2x | 39.1 | [config](../../retinanet/retinanet_r50_fpn_2x_coco.yml) | [download](https://paddledet.bj.bcebos.com/models/retinanet_r50_fpn_2x_coco.pdparams) |
| RetinaNet-ResNet50 | FGD | 1333x800 | 2x | 40.8(+1.7) | [config](../../retinanet/retinanet_r50_fpn_2x_coco.yml),[slim_config](./retinanet_resnet101_coco_distill.yml) | [download](https://paddledet.bj.bcebos.com/models/retinanet_r101_distill_r50_2x_coco.pdparams) |
<details>
<summary> 快速开始 </summary>
```shell
# 单卡训练(不推荐)
python tools/train.py -c configs/retinanet/retinanet_r50_fpn_2x_coco.yml --slim_config configs/slim/distill/retinanet_resnet101_coco_distill.yml
# 多卡训练
python -m paddle.distributed.launch --log_dir=logs/ --gpus 0,1,2,3,4,5,6,7 tools/train.py -c configs/retinanet/retinanet_r50_fpn_2x_coco.yml --slim_config configs/slim/distill/retinanet_resnet101_coco_distill.yml
# 评估
python tools/eval.py -c configs/retinanet/retinanet_r50_fpn_2x_coco.yml -o weights=https://paddledet.bj.bcebos.com/models/retinanet_r101_distill_r50_2x_coco.pdparams
# 预测
python tools/infer.py -c configs/retinanet/retinanet_r50_fpn_2x_coco.yml -o weights=https://paddledet.bj.bcebos.com/models/retinanet_r101_distill_r50_2x_coco.pdparams --infer_img=demo/000000014439_640x640.jpg
```
- `-c`: 指定模型配置文件，也是student配置文件。
- `--slim_config`: 指定压缩策略配置文件，也是teacher配置文件。
</details>
## CWD模型蒸馏
CWD全称为[Channel-wise Knowledge Distillation for Dense Prediction*](https://arxiv.org/pdf/2011.13256.pdf)，通过最小化教师网络与学生网络的通道概率图之间的 Kullback-Leibler (KL) 散度，使得在蒸馏过程更加关注每个通道的最显著的区域，进而提升文本检测与图像分割任务的精度。在PaddleDetection中，我们实现了CWD算法，并基于GFL和RetinaNet模型进行验证，实验结果如下：
| 模型 | 方案 | 输入尺寸 | epochs | Box mAP | 配置文件 | 下载链接 |
| ----------------- | ----------- | ------ | :----: | :-----------: | :--------------: | :------------: |
| RetinaNet-ResNet101| teacher | 1333x800 | 2x | 40.6 | [config](../../retinanet/retinanet_r101_fpn_2x_coco.yml) | [download](https://paddledet.bj.bcebos.com/models/retinanet_r101_fpn_2x_coco.pdparams) |
| RetinaNet-ResNet50 | student | 1333x800 | 2x | 39.1 | [config](../../retinanet/retinanet_r50_fpn_2x_coco.yml) | [download](https://paddledet.bj.bcebos.com/models/retinanet_r50_fpn_2x_coco.pdparams) |
| RetinaNet-ResNet50 | CWD | 1333x800 | 2x | 40.5(+1.4) | [config](../../retinanet/retinanet_r50_fpn_2x_coco.yml),[slim_config](./retinanet_resnet101_coco_distill_cwd.yml) | [download](https://paddledet.bj.bcebos.com/models/retinanet_r50_fpn_2x_coco_cwd.pdparams) |
| GFL_ResNet101-vd| teacher | 1333x800 | 2x | 46.8 | [config](../../gfl/gfl_r101vd_fpn_mstrain_2x_coco.yml) | [download](https://paddledet.bj.bcebos.com/models/gfl_r101vd_fpn_mstrain_2x_coco.pdparams) |
| GFL_ResNet50 | student | 1333x800 | 1x | 41.0 | [config](../../gfl/gfl_r50_fpn_1x_coco.yml) | [download](https://paddledet.bj.bcebos.com/models/gfl_r50_fpn_1x_coco.pdparams) |
| GFL_ResNet50 | CWD | 1333x800 | 2x | 44.0(+3.0) | [config](../../gfl/gfl_r50_fpn_1x_coco.yml),[slim_config](./gfl_r101vd_fpn_coco_distill_cwd.yml) | [download](https://bj.bcebos.com/v1/paddledet/models/gfl_r50_fpn_2x_coco_cwd.pdparams) |
<details>
<summary> 快速开始 </summary>
```shell
# 单卡训练(不推荐)
python tools/train.py -c configs/retinanet/retinanet_r50_fpn_2x_coco.yml --slim_config configs/slim/distill/retinanet_resnet101_coco_distill_cwd.yml
# 多卡训练
python -m paddle.distributed.launch --log_dir=logs/ --gpus 0,1,2,3,4,5,6,7 tools/train.py -c configs/retinanet/retinanet_r50_fpn_2x_coco.yml --slim_config configs/slim/distill/retinanet_resnet101_coco_distill_cwd.yml
# 评估
python tools/eval.py -c configs/retinanet/retinanet_r50_fpn_2x_coco.yml -o weights=https://paddledet.bj.bcebos.com/models/retinanet_r50_fpn_2x_coco_cwd.pdparams
# 预测
python tools/infer.py -c configs/retinanet/retinanet_r50_fpn_2x_coco.yml -o weights=https://paddledet.bj.bcebos.com/models/retinanet_r50_fpn_2x_coco_cwd.pdparams --infer_img=demo/000000014439_640x640.jpg
# 单卡训练(不推荐)
python tools/train.py -c configs/gfl/gfl_r50_fpn_1x_coco.yml --slim_config configs/slim/distill/gfl_r101vd_fpn_coco_distill_cwd.yml
# 多卡训练
python -m paddle.distributed.launch --log_dir=logs/ --gpus 0,1,2,3,4,5,6,7 tools/train.py -c configs/gfl/gfl_r50_fpn_1x_coco.yml --slim_config configs/slim/distill/gfl_r101vd_fpn_coco_distill_cwd.yml
# 评估
python tools/eval.py -c configs/gfl/gfl_r50_fpn_1x_coco.yml -o weights=https://paddledet.bj.bcebos.com/models/gfl_r50_fpn_2x_coco_cwd.pdparams
# 预测
python tools/infer.py -c configs/gfl/gfl_r50_fpn_1x_coco.yml -o weights=https://paddledet.bj.bcebos.com/models/gfl_r50_fpn_2x_coco_cwd.pdparams --infer_img=demo/000000014439_640x640.jpg
```
- `-c`: 指定模型配置文件，也是student配置文件。
- `--slim_config`: 指定压缩策略配置文件，也是teacher配置文件。
</details>
## LD模型蒸馏
LD全称为[Localization Distillation for Dense Object Detection](https://arxiv.org/abs/2102.12252)，将回归框表示为概率分布，把分类任务的KD用在定位任务上，并且使用因地制宜、分而治之的策略，在不同的区域分别学习分类知识与定位知识。在PaddleDetection中，我们实现了LD算法，并基于GFL模型进行验证，实验结果如下：
| 模型 | 方案 | 输入尺寸 | epochs | Box mAP | 配置文件 | 下载链接 |
| ----------------- | ----------- | ------ | :----: | :-----------: | :--------------: | :------------: |
| GFL_ResNet101-vd| teacher | 1333x800 | 2x | 46.8 | [config](../../gfl/gfl_r101vd_fpn_mstrain_2x_coco.yml) | [download](https://paddledet.bj.bcebos.com/models/gfl_r101vd_fpn_mstrain_2x_coco.pdparams) |
| GFL_ResNet18-vd | student | 1333x800 | 1x | 36.6 | [config](../../gfl/gfl_r18vd_1x_coco.yml) | [download](https://paddledet.bj.bcebos.com/models/gfl_r18vd_1x_coco.pdparams) |
| GFL_ResNet18-vd | LD | 1333x800 | 1x | 38.2(+1.6) | [config](../../gfl/gfl_slim_ld_r18vd_1x_coco.yml),[slim_config](./gfl_ld_distill.yml) | [download](https://bj.bcebos.com/v1/paddledet/models/gfl_slim_ld_r18vd_1x_coco.pdparams) |
<details>
<summary> 快速开始 </summary>
```shell
# 单卡训练(不推荐)
python tools/train.py -c configs/gfl/gfl_slim_ld_r18vd_1x_coco.yml --slim_config configs/slim/distill/gfl_ld_distill.yml
# 多卡训练
python -m paddle.distributed.launch --log_dir=logs/ --gpus 0,1,2,3,4,5,6,7 tools/train.py -c configs/gfl/gfl_slim_ld_r18vd_1x_coco.yml --slim_config configs/slim/distill/gfl_ld_distill.yml
# 评估
python tools/eval.py -c configs/gfl/gfl_slim_ld_r18vd_1x_coco.yml -o weights=https://paddledet.bj.bcebos.com/models/gfl_slim_ld_r18vd_1x_coco.pdparams
# 预测
python tools/infer.py -c configs/gfl/gfl_slim_ld_r18vd_1x_coco.yml -o weights=https://paddledet.bj.bcebos.com/models/gfl_slim_ld_r18vd_1x_coco.pdparams --infer_img=demo/000000014439_640x640.jpg
```
- `-c`: 指定模型配置文件，也是student配置文件。
- `--slim_config`: 指定压缩策略配置文件，也是teacher配置文件。
</details>
## PPYOLOE模型蒸馏
PaddleDetection提供了对PPYOLOE+ 进行模型蒸馏的方案，结合了logits蒸馏和feature蒸馏。
| 模型 | 方案 | 输入尺寸 | epochs | Box mAP | 配置文件 | 下载链接 |
| ----------------- | ----------- | ------ | :----: | :-----------: | :--------------: | :------------: |
| PP-YOLOE+_x | teacher | 640 | 80e | 54.7 | [config](../../ppyoloe/ppyoloe_plus_crn_x_80e_coco.yml) | [model](https://bj.bcebos.com/v1/paddledet/models/ppyoloe_plus_crn_x_80e_coco.pdparams) |
| PP-YOLOE+_l | student | 640 | 80e | 52.9 | [config](../../ppyoloe/ppyoloe_plus_crn_l_80e_coco.yml) | [model](https://bj.bcebos.com/v1/paddledet/models/ppyoloe_plus_crn_l_80e_coco.pdparams) |
| PP-YOLOE+_l | distill | 640 | 80e | **54.0(+1.1)** | [config](../../ppyoloe/distill/ppyoloe_plus_crn_l_80e_coco_distill.yml),[slim_config](./ppyoloe_plus_distill_x_distill_l.yml) | [model](https://bj.bcebos.com/v1/paddledet/models/ppyoloe_plus_crn_l_80e_coco_distill.pdparams) |
| PP-YOLOE+_l | teacher | 640 | 80e | 52.9 | [config](../../ppyoloe/ppyoloe_plus_crn_l_80e_coco.yml) | [model](https://bj.bcebos.com/v1/paddledet/models/ppyoloe_plus_crn_l_80e_coco.pdparams) |
| PP-YOLOE+_m | student | 640 | 80e | 49.8 | [config](../../ppyoloe/ppyoloe_plus_crn_m_80e_coco.yml) | [model](https://bj.bcebos.com/v1/paddledet/models/ppyoloe_plus_crn_m_80e_coco.pdparams) |
| PP-YOLOE+_m | distill | 640 | 80e | **51.0(+1.2)** | [config](../../ppyoloe/distill/ppyoloe_plus_crn_m_80e_coco_distill.yml),[slim_config](./ppyoloe_plus_distill_l_distill_m.yml) | [model](https://bj.bcebos.com/v1/paddledet/models/ppyoloe_plus_crn_m_80e_coco_distill.pdparams) |
<details>
<summary> 快速开始 </summary>
```shell
# 单卡训练(不推荐)
python tools/train.py -c configs/ppyoloe/distill/ppyoloe_plus_crn_l_80e_coco_distill.yml --slim_config configs/slim/distill/ppyoloe_plus_distill_x_distill_l.yml
# 多卡训练
python -m paddle.distributed.launch --log_dir=logs/ --gpus 0,1,2,3,4,5,6,7 tools/train.py -c configs/ppyoloe/distill/ppyoloe_plus_crn_l_80e_coco_distill.yml --slim_config configs/slim/distill/ppyoloe_plus_distill_x_distill_l.yml
# 评估
python tools/eval.py -c configs/ppyoloe/distill/ppyoloe_plus_crn_l_80e_coco_distill.yml -o weights=https://paddledet.bj.bcebos.com/models/ppyoloe_plus_crn_l_80e_coco_distill.pdparams
# 预测
python tools/infer.py -c configs/ppyoloe/distill/ppyoloe_plus_crn_l_80e_coco_distill.yml -o weights=https://paddledet.bj.bcebos.com/models/ppyoloe_plus_crn_l_80e_coco_distill.pdparams --infer_img=demo/000000014439_640x640.jpg
```
- `-c`: 指定模型配置文件，也是student配置文件。
- `--slim_config`: 指定压缩策略配置文件，也是teacher配置文件。
</details>
## 引用
```
@article{mehta2018object,
title={Object detection at 200 Frames Per Second},
author={Rakesh Mehta and Cemalettin Ozturk},
year={2018},
eprint={1805.06361},
archivePrefix={arXiv},
primaryClass={cs.CV}
}
@inproceedings{yang2022focal,
title={Focal and global knowledge distillation for detectors},
author={Yang, Zhendong and Li, Zhe and Jiang, Xiaohu and Gong, Yuan and Yuan, Zehuan and Zhao, Danpei and Yuan, Chun},
booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
pages={4643--4652},
year={2022}
}
@Inproceedings{zheng2022LD,
title={Localization Distillation for Dense Object Detection},
author= {Zheng, Zhaohui and Ye, Rongguang and Wang, Ping and Ren, Dongwei and Zuo, Wangmeng and Hou, Qibin and Cheng, Mingming},
booktitle={CVPR},
year={2022}
}
@inproceedings{shu2021channel,
title={Channel-wise knowledge distillation for dense prediction},
author={Shu, Changyong and Liu, Yifan and Gao, Jianfei and Yan, Zheng and Shen, Chunhua},
booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision},
pages={5311--5320},
year={2021}
}
```
configs/slim/distill/ppyoloe_plus_distill_l_distill_m.yml 0 → 100644
View file @ 522a602f
# teacher and slim config
_BASE_: [
'../../ppyoloe/ppyoloe_plus_crn_l_80e_coco.yml',
]
depth_mult: 1.0
width_mult: 1.0
for_distill: True
architecture: PPYOLOE
PPYOLOE:
backbone: CSPResNet
neck: CustomCSPPAN
yolo_head: PPYOLOEHead
post_process: ~
pretrain_weights: https://paddledet.bj.bcebos.com/models/ppyoloe_plus_crn_l_80e_coco.pdparams
find_unused_parameters: True
worker_num: 4
TrainReader:
sample_transforms:
- Decode: {}
- RandomDistort: {}
- RandomExpand: {fill_value: [123.675, 116.28, 103.53]}
- RandomCrop: {}
- RandomFlip: {}
batch_transforms:
- BatchRandomResize: {target_size: [640], random_size: True, random_interp: True, keep_ratio: False}
- NormalizeImage: {mean: [0., 0., 0.], std: [1., 1., 1.], norm_type: none}
- Permute: {}
- PadGT: {}
batch_size: 8
shuffle: True
drop_last: True
use_shared_memory: True
collate_batch: True
slim: Distill
slim_method: PPYOLOEDistill
distill_loss: DistillPPYOLOELoss
DistillPPYOLOELoss: # L -> M
loss_weight: {'logits': 4.0, 'feat': 1.0}
logits_distill: True
logits_loss_weight: {'class': 1.0, 'iou': 2.5, 'dfl': 0.5}
logits_ld_distill: True
logits_ld_params: {'weight': 20000, 'T': 10}
feat_distill: True
feat_distiller: 'fgd' # ['cwd', 'fgd', 'pkd', 'mgd', 'mimic']
feat_distill_place: 'neck_feats'
teacher_width_mult: 1.0 # L
student_width_mult: 0.75 # M
feat_out_channels: [768, 384, 192] # The actual channel will multiply width_mult
configs/slim/distill/ppyoloe_plus_distill_m_distill_s.yml 0 → 100644
View file @ 522a602f
# teacher and slim config
_BASE_: [
'../../ppyoloe/ppyoloe_plus_crn_m_80e_coco.yml',
]
depth_mult: 0.67
width_mult: 0.75
for_distill: True
architecture: PPYOLOE
PPYOLOE:
backbone: CSPResNet
neck: CustomCSPPAN
yolo_head: PPYOLOEHead
post_process: ~
pretrain_weights: https://paddledet.bj.bcebos.com/models/ppyoloe_plus_crn_m_80e_coco.pdparams
find_unused_parameters: True
worker_num: 4
TrainReader:
sample_transforms:
- Decode: {}
- RandomDistort: {}
- RandomExpand: {fill_value: [123.675, 116.28, 103.53]}
- RandomCrop: {}
- RandomFlip: {}
batch_transforms:
- BatchRandomResize: {target_size: [640], random_size: True, random_interp: True, keep_ratio: False}
- NormalizeImage: {mean: [0., 0., 0.], std: [1., 1., 1.], norm_type: none}
- Permute: {}
- PadGT: {}
batch_size: 8
shuffle: True
drop_last: True
use_shared_memory: True
collate_batch: True
slim: Distill
slim_method: PPYOLOEDistill
distill_loss: DistillPPYOLOELoss
DistillPPYOLOELoss: # M -> S
loss_weight: {'logits': 4.0, 'feat': 1.0}
logits_distill: True
logits_loss_weight: {'class': 1.0, 'iou': 2.5, 'dfl': 0.5}
logits_ld_distill: True
logits_ld_params: {'weight': 20000, 'T': 10}
feat_distill: True
feat_distiller: 'fgd' # ['cwd', 'fgd', 'pkd', 'mgd', 'mimic']
feat_distill_place: 'neck_feats'
teacher_width_mult: 0.75 # M
student_width_mult: 0.5 # S
feat_out_channels: [768, 384, 192] # The actual channel will multiply width_mult
configs/slim/distill/ppyoloe_plus_distill_x_distill_l.yml 0 → 100644
View file @ 522a602f
# teacher and slim config
_BASE_: [
'../../ppyoloe/ppyoloe_plus_crn_x_80e_coco.yml',
]
depth_mult: 1.33
width_mult: 1.25
for_distill: True
architecture: PPYOLOE
PPYOLOE:
backbone: CSPResNet
neck: CustomCSPPAN
yolo_head: PPYOLOEHead
post_process: ~
pretrain_weights: https://paddledet.bj.bcebos.com/models/ppyoloe_plus_crn_x_80e_coco.pdparams
find_unused_parameters: True
worker_num: 4
TrainReader:
sample_transforms:
- Decode: {}
- RandomDistort: {}
- RandomExpand: {fill_value: [123.675, 116.28, 103.53]}
- RandomCrop: {}
- RandomFlip: {}
batch_transforms:
- BatchRandomResize: {target_size: [640], random_size: True, random_interp: True, keep_ratio: False}
- NormalizeImage: {mean: [0., 0., 0.], std: [1., 1., 1.], norm_type: none}
- Permute: {}
- PadGT: {}
batch_size: 8
shuffle: True
drop_last: True
use_shared_memory: True
collate_batch: True
slim: Distill
slim_method: PPYOLOEDistill
distill_loss: DistillPPYOLOELoss
DistillPPYOLOELoss: # X -> L
loss_weight: {'logits': 4.0, 'feat': 1.0}
logits_distill: True
logits_loss_weight: {'class': 1.0, 'iou': 2.5, 'dfl': 0.5}
logits_ld_distill: True
logits_ld_params: {'weight': 20000, 'T': 10}
feat_distill: True
feat_distiller: 'fgd' # ['cwd', 'fgd', 'pkd', 'mgd', 'mimic']
feat_distill_place: 'neck_feats'
teacher_width_mult: 1.25 # X
student_width_mult: 1.0 # L
feat_out_channels: [768, 384, 192] # The actual channel will multiply width_mult
configs/slim/distill/retinanet_resnet101_coco_distill.yml 0 → 100644
View file @ 522a602f
_BASE_: [
'../../retinanet/retinanet_r101_fpn_2x_coco.yml',
]
pretrain_weights: https://paddledet.bj.bcebos.com/models/retinanet_r101_fpn_2x_coco.pdparams
slim: Distill
slim_method: FGD
distill_loss: FGDFeatureLoss
distill_loss_name: ['neck_f_4', 'neck_f_3', 'neck_f_2', 'neck_f_1', 'neck_f_0']
FGDFeatureLoss:
student_channels: 256
teacher_channels: 256
temp: 0.5
alpha_fgd: 0.001
beta_fgd: 0.0005
gamma_fgd: 0.0005
lambda_fgd: 0.000005
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment