未验证 提交 e639b354 编写于 作者: F Feng Ni 提交者: GitHub

Add retinanet_r50_2x (#7277)

* remove duplicate backbone variant, test=document_fix

* add retinanet_r50_2x
上级 78a6629b
......@@ -5,6 +5,7 @@
| Backbone | Model | imgs/GPU | lr schedule | FPS | Box AP | download | config |
| ------------ | --------- | -------- | ----------- | --- | ------ | ---------- | ----------- |
| ResNet50-FPN | RetinaNet | 2 | 1x | --- | 37.5 | [model](https://bj.bcebos.com/v1/paddledet/models/retinanet_r50_fpn_1x_coco.pdparams) | [config](./retinanet_r50_fpn_1x_coco.yml) |
| ResNet50-FPN | RetinaNet | 2 | 2x | --- | 39.1 | [model](https://bj.bcebos.com/v1/paddledet/models/retinanet_r50_fpn_2x_coco.pdparams) | [config](./retinanet_r50_fpn_2x_coco.yml) |
| ResNet101-FPN| RetinaNet | 2 | 2x | --- | 40.6 | [model](https://paddledet.bj.bcebos.com/models/retinanet_r101_fpn_2x_coco.pdparams) | [config](./retinanet_r101_fpn_2x_coco.yml) |
| ResNet50-FPN | RetinaNet + [FGD](../slim/distill/README.md) | 2 | 2x | --- | 40.8 | [model](https://bj.bcebos.com/v1/paddledet/models/retinanet_r101_distill_r50_2x_coco.pdparams) | [config](./retinanet_r50_fpn_2x_coco.yml)/[slim_config](../slim/distill/retinanet_resnet101_coco_distill.yml) |
......
architecture: RetinaNet
pretrain_weights: https://paddledet.bj.bcebos.com/models/pretrained/ResNet101_pretrained.pdparams
RetinaNet:
backbone: ResNet
neck: FPN
head: RetinaHead
ResNet:
depth: 101
variant: b
norm_type: bn
freeze_at: 0
return_idx: [1,2,3]
num_stages: 4
FPN:
out_channel: 256
spatial_scales: [0.125, 0.0625, 0.03125]
extra_stage: 2
has_extra_convs: true
use_c5: false
RetinaHead:
conv_feat:
name: RetinaFeat
feat_in: 256
feat_out: 256
num_convs: 4
norm_type: null
use_dcn: false
anchor_generator:
name: RetinaAnchorGenerator
octave_base_scale: 4
scales_per_octave: 3
aspect_ratios: [0.5, 1.0, 2.0]
strides: [8.0, 16.0, 32.0, 64.0, 128.0]
bbox_assigner:
name: MaxIoUAssigner
positive_overlap: 0.5
negative_overlap: 0.4
allow_low_quality: true
loss_class:
name: FocalLoss
gamma: 2.0
alpha: 0.25
loss_weight: 1.0
loss_bbox:
name: SmoothL1Loss
beta: 0.0
loss_weight: 1.0
nms:
name: MultiClassNMS
nms_top_k: 1000
keep_top_k: 100
score_threshold: 0.05
nms_threshold: 0.5
_BASE_: [
'../datasets/coco_detection.yml',
'../runtime.yml',
'_base_/retinanet_r101_fpn.yml',
'_base_/retinanet_r50_fpn.yml',
'_base_/optimizer_2x.yml',
'_base_/retinanet_reader.yml'
]
weights: https://paddledet.bj.bcebos.com/models/retinanet_r101_fpn_2x_coco.pdparams
weights: output/retinanet_r101_fpn_2x_coco/model_final
pretrain_weights: https://paddledet.bj.bcebos.com/models/pretrained/ResNet101_pretrained.pdparams
ResNet:
depth: 101
variant: b
norm_type: bn
freeze_at: 0
return_idx: [1, 2, 3]
num_stages: 4
_BASE_: [
'../datasets/coco_detection.yml',
'../runtime.yml',
'_base_/retinanet_r50_fpn.yml',
'_base_/optimizer_2x.yml',
'_base_/retinanet_reader.yml'
]
weights: output/retinanet_r50_fpn_2x_coco/model_final
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