提交 ff28150a 编写于 作者: G Guanghua Yu 提交者: qingqing01

[PaddleDetection] add cascade deformable mdoels (#3082)

* Add cascade deformable mdoels
* Fix MODLE_ZOO.md
上级 eaf5324e
architecture: CascadeRCNN
train_feed: FasterRCNNTrainFeed
eval_feed: FasterRCNNEvalFeed
test_feed: FasterRCNNTestFeed
max_iters: 90000
snapshot_iter: 10000
use_gpu: true
log_smooth_window: 20
log_iter: 20
save_dir: output
pretrain_weights: https://paddle-imagenet-models-name.bj.bcebos.com/ResNet101_vd_pretrained.tar
weights: output/cascade_rcnn_dcn_r101_vd_fpn_1x/model_final
metric: COCO
num_classes: 81
CascadeRCNN:
backbone: ResNet
fpn: FPN
rpn_head: FPNRPNHead
roi_extractor: FPNRoIAlign
bbox_head: CascadeBBoxHead
bbox_assigner: CascadeBBoxAssigner
ResNet:
norm_type: bn
depth: 101
feature_maps: [2, 3, 4, 5]
freeze_at: 2
variant: d
dcn_stages: [3, 4, 5]
FPN:
min_level: 2
max_level: 6
num_chan: 256
spatial_scale: [0.03125, 0.0625, 0.125, 0.25]
FPNRPNHead:
anchor_generator:
anchor_sizes: [32, 64, 128, 256, 512]
aspect_ratios: [0.5, 1.0, 2.0]
stride: [16.0, 16.0]
variance: [1.0, 1.0, 1.0, 1.0]
anchor_start_size: 32
min_level: 2
max_level: 6
num_chan: 256
rpn_target_assign:
rpn_batch_size_per_im: 256
rpn_fg_fraction: 0.5
rpn_positive_overlap: 0.7
rpn_negative_overlap: 0.3
rpn_straddle_thresh: 0.0
train_proposal:
min_size: 0.0
nms_thresh: 0.7
pre_nms_top_n: 2000
post_nms_top_n: 2000
test_proposal:
min_size: 0.0
nms_thresh: 0.7
pre_nms_top_n: 1000
post_nms_top_n: 1000
FPNRoIAlign:
canconical_level: 4
canonical_size: 224
min_level: 2
max_level: 5
box_resolution: 7
sampling_ratio: 2
CascadeBBoxAssigner:
batch_size_per_im: 512
bbox_reg_weights: [10, 20, 30]
bg_thresh_lo: [0.0, 0.0, 0.0]
bg_thresh_hi: [0.5, 0.6, 0.7]
fg_thresh: [0.5, 0.6, 0.7]
fg_fraction: 0.25
CascadeBBoxHead:
head: FC6FC7Head
nms:
keep_top_k: 100
nms_threshold: 0.5
score_threshold: 0.05
FC6FC7Head:
num_chan: 1024
LearningRate:
base_lr: 0.02
schedulers:
- !PiecewiseDecay
gamma: 0.1
milestones: [60000, 80000]
- !LinearWarmup
start_factor: 0.1
steps: 1000
OptimizerBuilder:
optimizer:
momentum: 0.9
type: Momentum
regularizer:
factor: 0.0001
type: L2
FasterRCNNTrainFeed:
batch_size: 2
dataset:
dataset_dir: dataset/coco
annotation: annotations/instances_train2017.json
image_dir: train2017
batch_transforms:
- !PadBatch
pad_to_stride: 32
drop_last: false
num_workers: 2
FasterRCNNEvalFeed:
batch_size: 1
dataset:
dataset_dir: dataset/coco
annotation: annotations/instances_val2017.json
image_dir: val2017
batch_transforms:
- !PadBatch
pad_to_stride: 32
FasterRCNNTestFeed:
batch_size: 1
dataset:
annotation: dataset/coco/annotations/instances_val2017.json
batch_transforms:
- !PadBatch
pad_to_stride: 32
drop_last: false
num_workers: 2
architecture: CascadeRCNN
train_feed: FasterRCNNTrainFeed
eval_feed: FasterRCNNEvalFeed
test_feed: FasterRCNNTestFeed
max_iters: 90000
snapshot_iter: 10000
use_gpu: true
log_smooth_window: 20
log_iter: 20
save_dir: output
pretrain_weights: https://paddle-imagenet-models-name.bj.bcebos.com/ResNet50_cos_pretrained.tar
weights: output/cascade_rcnn_dcn_r50_fpn_1x/model_final
metric: COCO
num_classes: 81
CascadeRCNN:
backbone: ResNet
fpn: FPN
rpn_head: FPNRPNHead
roi_extractor: FPNRoIAlign
bbox_head: CascadeBBoxHead
bbox_assigner: CascadeBBoxAssigner
ResNet:
norm_type: bn
depth: 50
feature_maps: [2, 3, 4, 5]
freeze_at: 2
variant: b
dcn_stages: [3, 4, 5]
FPN:
min_level: 2
max_level: 6
num_chan: 256
spatial_scale: [0.03125, 0.0625, 0.125, 0.25]
FPNRPNHead:
anchor_generator:
anchor_sizes: [32, 64, 128, 256, 512]
aspect_ratios: [0.5, 1.0, 2.0]
stride: [16.0, 16.0]
variance: [1.0, 1.0, 1.0, 1.0]
anchor_start_size: 32
min_level: 2
max_level: 6
num_chan: 256
rpn_target_assign:
rpn_batch_size_per_im: 256
rpn_fg_fraction: 0.5
rpn_positive_overlap: 0.7
rpn_negative_overlap: 0.3
rpn_straddle_thresh: 0.0
train_proposal:
min_size: 0.0
nms_thresh: 0.7
pre_nms_top_n: 2000
post_nms_top_n: 2000
test_proposal:
min_size: 0.0
nms_thresh: 0.7
pre_nms_top_n: 1000
post_nms_top_n: 1000
FPNRoIAlign:
canconical_level: 4
canonical_size: 224
min_level: 2
max_level: 5
box_resolution: 7
sampling_ratio: 2
CascadeBBoxAssigner:
batch_size_per_im: 512
bbox_reg_weights: [10, 20, 30]
bg_thresh_lo: [0.0, 0.0, 0.0]
bg_thresh_hi: [0.5, 0.6, 0.7]
fg_thresh: [0.5, 0.6, 0.7]
fg_fraction: 0.25
CascadeBBoxHead:
head: FC6FC7Head
nms:
keep_top_k: 100
nms_threshold: 0.5
score_threshold: 0.05
FC6FC7Head:
num_chan: 1024
LearningRate:
base_lr: 0.02
schedulers:
- !PiecewiseDecay
gamma: 0.1
milestones: [60000, 80000]
- !LinearWarmup
start_factor: 0.1
steps: 1000
OptimizerBuilder:
optimizer:
momentum: 0.9
type: Momentum
regularizer:
factor: 0.0001
type: L2
FasterRCNNTrainFeed:
batch_size: 2
dataset:
dataset_dir: dataset/coco
annotation: annotations/instances_train2017.json
image_dir: train2017
batch_transforms:
- !PadBatch
pad_to_stride: 32
drop_last: false
num_workers: 2
FasterRCNNEvalFeed:
batch_size: 1
dataset:
dataset_dir: dataset/coco
annotation: annotations/instances_val2017.json
image_dir: val2017
batch_transforms:
- !PadBatch
pad_to_stride: 32
FasterRCNNTestFeed:
batch_size: 1
dataset:
annotation: dataset/coco/annotations/instances_val2017.json
batch_transforms:
- !PadBatch
pad_to_stride: 32
drop_last: false
num_workers: 2
architecture: CascadeRCNN
train_feed: FasterRCNNTrainFeed
eval_feed: FasterRCNNEvalFeed
test_feed: FasterRCNNTestFeed
max_iters: 90000
snapshot_iter: 10000
use_gpu: true
log_smooth_window: 20
log_iter: 20
save_dir: output
pretrain_weights: https://paddle-imagenet-models-name.bj.bcebos.com/ResNeXt101_vd_64x4d_pretrained.tar
weights: output/cascade_rcnn_dcn_x101_vd_64x4d_fpn_1x/model_final
metric: COCO
num_classes: 81
CascadeRCNN:
backbone: ResNeXt
fpn: FPN
rpn_head: FPNRPNHead
roi_extractor: FPNRoIAlign
bbox_head: CascadeBBoxHead
bbox_assigner: CascadeBBoxAssigner
ResNeXt:
norm_type: bn
depth: 101
feature_maps: [2, 3, 4, 5]
freeze_at: 2
group_width: 4
groups: 64
variant: d
dcn_stages: [3, 4, 5]
FPN:
min_level: 2
max_level: 6
num_chan: 256
spatial_scale: [0.03125, 0.0625, 0.125, 0.25]
FPNRPNHead:
anchor_generator:
anchor_sizes: [32, 64, 128, 256, 512]
aspect_ratios: [0.5, 1.0, 2.0]
stride: [16.0, 16.0]
variance: [1.0, 1.0, 1.0, 1.0]
anchor_start_size: 32
min_level: 2
max_level: 6
num_chan: 256
rpn_target_assign:
rpn_batch_size_per_im: 256
rpn_fg_fraction: 0.5
rpn_positive_overlap: 0.7
rpn_negative_overlap: 0.3
rpn_straddle_thresh: 0.0
train_proposal:
min_size: 0.0
nms_thresh: 0.7
pre_nms_top_n: 2000
post_nms_top_n: 2000
test_proposal:
min_size: 0.0
nms_thresh: 0.7
pre_nms_top_n: 1000
post_nms_top_n: 1000
FPNRoIAlign:
canconical_level: 4
canonical_size: 224
min_level: 2
max_level: 5
box_resolution: 7
sampling_ratio: 2
CascadeBBoxAssigner:
batch_size_per_im: 512
bbox_reg_weights: [10, 20, 30]
bg_thresh_lo: [0.0, 0.0, 0.0]
bg_thresh_hi: [0.5, 0.6, 0.7]
fg_thresh: [0.5, 0.6, 0.7]
fg_fraction: 0.25
CascadeBBoxHead:
head: FC6FC7Head
nms:
keep_top_k: 100
nms_threshold: 0.5
score_threshold: 0.05
FC6FC7Head:
num_chan: 1024
LearningRate:
base_lr: 0.02
schedulers:
- !PiecewiseDecay
gamma: 0.1
milestones: [60000, 80000]
- !LinearWarmup
start_factor: 0.1
steps: 1000
OptimizerBuilder:
optimizer:
momentum: 0.9
type: Momentum
regularizer:
factor: 0.0001
type: L2
FasterRCNNTrainFeed:
batch_size: 2
dataset:
dataset_dir: dataset/coco
annotation: annotations/instances_train2017.json
image_dir: train2017
batch_transforms:
- !PadBatch
pad_to_stride: 32
drop_last: false
num_workers: 2
FasterRCNNEvalFeed:
batch_size: 1
dataset:
dataset_dir: dataset/coco
annotation: annotations/instances_val2017.json
image_dir: val2017
batch_transforms:
- !PadBatch
pad_to_stride: 32
FasterRCNNTestFeed:
batch_size: 1
dataset:
annotation: dataset/coco/annotations/instances_val2017.json
batch_transforms:
- !PadBatch
pad_to_stride: 32
drop_last: false
num_workers: 2
architecture: RetinaNet
train_feed: FasterRCNNTrainFeed
eval_feed: FasterRCNNEvalFeed
test_feed: FasterRCNNTestFeed
max_iters: 180000
use_gpu: true
pretrain_weights: https://paddle-imagenet-models-name.bj.bcebos.com/ResNeXt101_vd_64x4d_pretrained.tar
weights: output/retinanet_x101_vd_64x4d_fpn_1x/model_final
log_smooth_window: 20
log_iter: 20
snapshot_iter: 30000
metric: COCO
save_dir: output
num_classes: 81
RetinaNet:
backbone: ResNeXt
fpn: FPN
retina_head: RetinaHead
ResNeXt:
depth: 101
feature_maps: [3, 4, 5]
freeze_at: 2
group_width: 4
groups: 64
norm_type: bn
variant: d
FPN:
max_level: 7
min_level: 3
num_chan: 256
spatial_scale: [0.03125, 0.0625, 0.125]
has_extra_convs: true
RetinaHead:
num_convs_per_octave: 4
num_chan: 256
max_level: 7
min_level: 3
prior_prob: 0.01
base_scale: 4
num_scales_per_octave: 3
anchor_generator:
aspect_ratios: [1.0, 2.0, 0.5]
variance: [1.0, 1.0, 1.0, 1.0]
target_assign:
positive_overlap: 0.5
negative_overlap: 0.4
gamma: 2.0
alpha: 0.25
sigma: 3.0151134457776365
output_decoder:
score_thresh: 0.05
nms_thresh: 0.5
pre_nms_top_n: 1000
detections_per_im: 100
nms_eta: 1.0
LearningRate:
base_lr: 0.005
schedulers:
- !PiecewiseDecay
gamma: 0.1
milestones: [120000, 160000]
- !LinearWarmup
start_factor: 0.1
steps: 1000
OptimizerBuilder:
optimizer:
momentum: 0.9
type: Momentum
regularizer:
factor: 0.0001
type: L2
FasterRCNNTrainFeed:
batch_size: 1
dataset:
dataset_dir: dataset/coco
annotation: annotations/instances_train2017.json
image_dir: train2017
batch_transforms:
- !PadBatch
pad_to_stride: 128
num_workers: 2
FasterRCNNEvalFeed:
batch_size: 1
dataset:
dataset_dir: dataset/coco
annotation: annotations/instances_val2017.json
image_dir: val2017
batch_transforms:
- !PadBatch
pad_to_stride: 128
num_workers: 2
FasterRCNNTestFeed:
batch_size: 1
dataset:
annotation: dataset/coco/annotations/instances_val2017.json
batch_transforms:
- !PadBatch
pad_to_stride: 128
num_workers: 2
......@@ -9,7 +9,7 @@
## Common settings
- All models below except SSD were trained on `coco_2017_train`, and tested on `coco_2017_val`.
- All models below were trained on `coco_2017_train`, and tested on `coco_2017_val`.
- Batch Normalization layers in backbones are replaced by Affine Channel layers.
- Unless otherwise noted, all ResNet backbones adopt the [ResNet-B](https://arxiv.org/pdf/1812.01187) variant..
- For RCNN and RetinaNet models, only horizontal flipping data augmentation was used in the training phase and no augmentations were used in the testing phase.
......@@ -70,6 +70,9 @@ The backbone models pretrained on ImageNet are available. All backbone models ar
| ResNet50-vd-FPN | Mask | c3-c5 | 1 | 2x | 42.9 | 38.0 | [model](https://paddlemodels.bj.bcebos.com/object_detection/mask_rcnn_dcn_r50_vd_fpn_2x.tar) |
| ResNet101-vd-FPN | Mask | c3-c5 | 1 | 1x | 44.6 | 39.2 | [model](https://paddlemodels.bj.bcebos.com/object_detection/mask_rcnn_dcn_r101_vd_fpn_1x.tar) |
| ResNeXt101-vd-FPN | Mask | c3-c5 | 1 | 1x | 46.2 | 40.4 | [model](https://paddlemodels.bj.bcebos.com/object_detection/mask_rcnn_dcn_x101_vd_64x4d_fpn_1x.tar) |
| ResNet50-FPN | Cascade Faster | c3-c5 | 2 | 1x | 44.2 | - | [model](https://paddlemodels.bj.bcebos.com/object_detection/cascade_rcnn_dcn_r50_fpn_1x.tar) |
| ResNet101-vd-FPN | Cascade Faster | c3-c5 | 2 | 1x | 46.4 | - | [model](https://paddlemodels.bj.bcebos.com/object_detection/cascade_rcnn_dcn_r101_vd_fpn_1x.tar) |
| ResNeXt101-vd-FPN | Cascade Faster | c3-c5 | 2 | 1x | 47.3 | - | [model](https://paddlemodels.bj.bcebos.com/object_detection/cascade_rcnn_dcn_x101_vd_64x4d_fpn_1x.tar) |
#### Notes:
- Deformable ConvNets v2(dcn_v2) reference from [Deformable ConvNets v2](https://arxiv.org/abs/1811.11168).
......@@ -111,16 +114,18 @@ results of image size 608/416/320 above.
### RetinaNet
| Backbone | Image/gpu | Lr schd | Box AP | Download |
| :----------- | :-----: | :-----: | :----: | :-------: |
| ResNet50-FPN | 2 | 1x | 36.0 | [model](https://paddlemodels.bj.bcebos.com/object_detection/retinanet_r50_fpn_1x.tar) |
| ResNet101-FPN | 2 | 1x | 37.3 | [model](https://paddlemodels.bj.bcebos.com/object_detection/retinanet_r101_fpn_1x.tar) |
| Backbone | Image/gpu | Lr schd | Box AP | Download |
| :---------------: | :-----: | :-----: | :----: | :-------: |
| ResNet50-FPN | 2 | 1x | 36.0 | [model](https://paddlemodels.bj.bcebos.com/object_detection/retinanet_r50_fpn_1x.tar) |
| ResNet101-FPN | 2 | 1x | 37.3 | [model](https://paddlemodels.bj.bcebos.com/object_detection/retinanet_r101_fpn_1x.tar) |
| ResNeXt101-vd-FPN | 1 | 1x | 40.5 | [model](https://paddlemodels.bj.bcebos.com/object_detection/retinanet_x101_vd_64x4d_fpn_1x.tar) |
**Notes:** In RetinaNet, the base LR is changed to 0.01 for minibatch size 16.
### SSD
| Backbone | Size | Image/gpu | Lr schd | Box AP | Download |
| :----------: | :--: | :-------: | :-----: | :----: | :-------: |
| VGG16 | 300 | 8 | 40w | 25.1 | [model](https://paddlemodels.bj.bcebos.com/object_detection/ssd_vgg16_300.tar) |
| VGG16 | 512 | 8 | 40w | 29.1 | [model](https://paddlemodels.bj.bcebos.com/object_detection/ssd_vgg16_512.tar) |
......
......@@ -10,7 +10,7 @@
## 通用设置
- SSD模型在VOC数据集中训练和测试,其余所有模型均在COCO17数据集中训练和测试。
- 所有模型均在COCO17数据集中训练和测试。
- 除非特殊说明,所有ResNet骨干网络采用[ResNet-B](https://arxiv.org/pdf/1812.01187)结构。
- 对于RCNN和RetinaNet系列模型,训练阶段仅使用水平翻转作为数据增强,测试阶段不使用数据增强。
......@@ -70,6 +70,9 @@ Paddle提供基于ImageNet的骨架网络预训练模型。所有预训练模型
| ResNet50-vd-FPN | Mask | c3-c5 | 1 | 2x | 42.9 | 38.0 | [下载链接](https://paddlemodels.bj.bcebos.com/object_detection/mask_rcnn_dcn_r50_vd_fpn_2x.tar) |
| ResNet101-vd-FPN | Mask | c3-c5 | 1 | 1x | 44.6 | 39.2 | [下载链接](https://paddlemodels.bj.bcebos.com/object_detection/mask_rcnn_dcn_r101_vd_fpn_1x.tar) |
| ResNeXt101-vd-FPN | Mask | c3-c5 | 1 | 1x | 46.2 | 40.4 | [下载链接](https://paddlemodels.bj.bcebos.com/object_detection/mask_rcnn_dcn_x101_vd_64x4d_fpn_1x.tar) |
| ResNet50-FPN | Cascade Faster | c3-c5 | 2 | 1x | 44.2 | - | [下载链接](https://paddlemodels.bj.bcebos.com/object_detection/cascade_rcnn_dcn_r50_fpn_1x.tar) |
| ResNet101-vd-FPN | Cascade Faster | c3-c5 | 2 | 1x | 46.4 | - | [下载链接](https://paddlemodels.bj.bcebos.com/object_detection/cascade_rcnn_dcn_r101_vd_fpn_1x.tar) |
| ResNeXt101-vd-FPN | Cascade Faster | c3-c5 | 2 | 1x | 47.3 | - | [下载链接](https://paddlemodels.bj.bcebos.com/object_detection/cascade_rcnn_dcn_x101_vd_64x4d_fpn_1x.tar) |
#### 注意事项:
- Deformable卷积网络v2(dcn_v2)参考自论文[Deformable ConvNets v2](https://arxiv.org/abs/1811.11168).
......@@ -108,16 +111,18 @@ Paddle提供基于ImageNet的骨架网络预训练模型。所有预训练模型
### RetinaNet
| 骨架网络 | 每张GPU图片个数 | 学习率策略 | Box AP | 下载 |
| :----------- | :-----: | :-----: | :----: | :-------: |
| ResNet50-FPN | 2 | 1x | 36.0 | [下载链接](https://paddlemodels.bj.bcebos.com/object_detection/retinanet_r50_fpn_1x.tar) |
| ResNet101-FPN | 2 | 1x | 37.3 | [下载链接](https://paddlemodels.bj.bcebos.com/object_detection/retinanet_r101_fpn_1x.tar) |
| 骨架网络 | 每张GPU图片个数 | 学习率策略 | Box AP | 下载 |
| :---------------: | :-----: | :-----: | :----: | :-------: |
| ResNet50-FPN | 2 | 1x | 36.0 | [下载链接](https://paddlemodels.bj.bcebos.com/object_detection/retinanet_r50_fpn_1x.tar) |
| ResNet101-FPN | 2 | 1x | 37.3 | [下载链接](https://paddlemodels.bj.bcebos.com/object_detection/retinanet_r101_fpn_1x.tar) |
| ResNeXt101-vd-FPN | 1 | 1x | 40.5 | [下载链接](https://paddlemodels.bj.bcebos.com/object_detection/retinanet_x101_vd_64x4d_fpn_1x.tar) |
**注意事项:** RetinaNet系列模型中,在总batch size为16下情况下,初始学习率改为0.01。
### SSD
| 骨架网络 | 输入尺寸 | 每张GPU图片个数 | 学习率策略 | Box AP | 下载 |
| :----------: | :--: | :-------: | :-----: | :----: | :-------: |
| VGG16 | 300 | 8 | 40万 | 25.1 | [下载链接](https://paddlemodels.bj.bcebos.com/object_detection/ssd_vgg16_300.tar) |
| VGG16 | 512 | 8 | 40万 | 29.1 | [下载链接](https://paddlemodels.bj.bcebos.com/object_detection/ssd_vgg16_512.tar) |
......
Markdown is supported
0% .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑!
想要评论请 注册