Skip to content

Commit aa73140

Browse files
nepeplwunepeplwu
authored
Add Segformer (PaddlePaddle#1293)
1 parent 4d8cbc4 commit aa73140

23 files changed

+1219
-51
lines changed

‎configs/README.md

+2-2
Original file line numberDiff line numberDiff line change
@@ -28,9 +28,9 @@
2828
### optimizer
2929
> 训练优化器
3030
> * 参数
31-
> * type : 优化器类型，目前只支持'sgd'和'adam'
32-
> * momentum : 动量
31+
> * type : 优化器类型，支持目前Paddle官方所有优化器
3332
> * weight_decay : L2正则化的值
33+
> * **others** : 请参考[Paddle官方Optimizer文档](https://www.paddlepaddle.org.cn/documentation/docs/zh/api/paddle/optimizer/Overview_cn.html)
3434
3535
----
3636
### lr_scheduler

‎configs/segformer/README.md

+18
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,18 @@
1+
# SegFormer: Simple and Efficient Design for Semantic Segmentation with Transformers
2+
3+
## Reference
4+
5+
> Xie, Enze, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M. Alvarez, and Ping Luo. "SegFormer: Simple and Efficient Design for Semantic Segmentation with Transformers." arXiv preprint arXiv:2105.15203 (2021).
6+
7+
## Performance
8+
9+
### Cityscapes
10+
11+
| Model | Backbone | Resolution | Training Iters | mIoU(slice) | mIoU (flip) | mIoU (ms+flip) | Links |
12+
|-|-|-|-|-|-|-|-|
13+
|SegFormer_B0|-|1024x1024|160000|76.73%|77.16%|-|[model](https://bj.bcebos.com/paddleseg/dygraph/cityscapes/segformer_b0_cityscapes_1024x1024_160k/model.pdparams) \| [log](https://bj.bcebos.com/paddleseg/dygraph/cityscapes/segformer_b0_cityscapes_1024x1024_160k/train.log) \| [vdl](https://paddlepaddle.org.cn/paddle/visualdl/service/app?id=227e067add44d44383c402ec5aead11b)|
14+
|SegFormer_B1|-|1024x1024|160000|78.35%|78.64%|-|[model](https://bj.bcebos.com/paddleseg/dygraph/cityscapes/segformer_b1_cityscapes_1024x1024_160k/model.pdparams) \| [log](https://bj.bcebos.com/paddleseg/dygraph/cityscapes/segformer_b1_cityscapes_1024x1024_160k/train.log) \| [vdl](https://paddlepaddle.org.cn/paddle/visualdl/service/app?id=a0f4e8eacf346826e3150989b6a9f849)|
15+
|SegFormer_B2|-|1024x1024|160000|81.60%|81.82%|-|[model](https://bj.bcebos.com/paddleseg/dygraph/cityscapes/segformer_b2_cityscapes_1024x1024_160k/model.pdparams) \| [log](https://bj.bcebos.com/paddleseg/dygraph/cityscapes/segformer_b2_cityscapes_1024x1024_160k/train.log) \| [vdl](https://paddlepaddle.org.cn/paddle/visualdl/service/app?id=734c0d99d858d0db7ff58f03d18289fe)|
16+
|SegFormer_B3|-|1024x1024|160000|82.47%|82.60%|-|[model](https://bj.bcebos.com/paddleseg/dygraph/cityscapes/segformer_b3_cityscapes_1024x1024_160k/model.pdparams) \| [log](https://bj.bcebos.com/paddleseg/dygraph/cityscapes/segformer_b3_cityscapes_1024x1024_160k/train.log) \| [vdl](https://paddlepaddle.org.cn/paddle/visualdl/service/app?id=406282a64c45d008bf4445c5669d6579)|
17+
|SegFormer_B4|-|1024x1024|160000|82.38%|82.59%|-|[model](https://bj.bcebos.com/paddleseg/dygraph/cityscapes/segformer_b4_cityscapes_1024x1024_160k/model.pdparams) \| [log](https://bj.bcebos.com/paddleseg/dygraph/cityscapes/segformer_b4_cityscapes_1024x1024_160k/train.log) \| [vdl](https://paddlepaddle.org.cn/paddle/visualdl/service/app?id=dc51a262eb8be9273970354ed445e760)|
18+
|SegFormer_B5|-|1024x1024|160000|82.58%|82.82%|-|[model](https://bj.bcebos.com/paddleseg/dygraph/cityscapes/segformer_b5_cityscapes_1024x1024_160k/model.pdparams) \| [log](https://bj.bcebos.com/paddleseg/dygraph/cityscapes/segformer_b5_cityscapes_1024x1024_160k/train.log) \| [vdl](https://paddlepaddle.org.cn/paddle/visualdl/service/app?id=306d042a8e4d82ccceabd988a478a2f8)|

‎configs/segformer/segformer_b0_cityscapes_1024x1024_160k.yml

+31
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,31 @@
1+
_base_: '../_base_/cityscapes_1024x1024.yml'
2+
3+
batch_size: 2
4+
iters: 160000
5+
6+
model:
7+
type: SegFormer_B0
8+
num_classes: 19
9+
pretrained: https://bj.bcebos.com/paddleseg/dygraph/mix_vision_transformer_b0.tar.gz
10+
11+
optimizer:
12+
_inherited_: False
13+
type: AdamW
14+
beta1: 0.9
15+
beta2: 0.999
16+
weight_decay: 0.01
17+
18+
lr_scheduler:
19+
type: PolynomialDecay
20+
learning_rate: 0.00006
21+
power: 1
22+
23+
loss:
24+
types:
25+
- type: CrossEntropyLoss
26+
coef: [1]
27+
28+
test_config:
29+
is_slide: True
30+
crop_size: [1024, 1024]
31+
stride: [768, 768]

‎configs/segformer/segformer_b0_cityscapes_1024x512_160k.yml

+26
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,26 @@
1+
_base_: '../_base_/cityscapes.yml'
2+
3+
batch_size: 1
4+
iters: 160000
5+
6+
model:
7+
type: SegFormer_B0
8+
num_classes: 19
9+
pretrained: https://bj.bcebos.com/paddleseg/dygraph/mix_vision_transformer_b0.tar.gz
10+
11+
optimizer:
12+
_inherited_: False
13+
type: AdamW
14+
beta1: 0.9
15+
beta2: 0.999
16+
weight_decay: 0.01
17+
18+
lr_scheduler:
19+
type: PolynomialDecay
20+
learning_rate: 0.00006
21+
power: 1
22+
23+
loss:
24+
types:
25+
- type: CrossEntropyLoss
26+
coef: [1]

‎configs/segformer/segformer_b1_cityscapes_1024x1024_160k.yml

+31
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,31 @@
1+
_base_: '../_base_/cityscapes_1024x1024.yml'
2+
3+
batch_size: 2
4+
iters: 160000
5+
6+
model:
7+
type: SegFormer_B1
8+
num_classes: 19
9+
pretrained: https://bj.bcebos.com/paddleseg/dygraph/mix_vision_transformer_b1.tar.gz
10+
11+
optimizer:
12+
_inherited_: False
13+
type: AdamW
14+
beta1: 0.9
15+
beta2: 0.999
16+
weight_decay: 0.01
17+
18+
lr_scheduler:
19+
type: PolynomialDecay
20+
learning_rate: 0.00006
21+
power: 1
22+
23+
loss:
24+
types:
25+
- type: CrossEntropyLoss
26+
coef: [1]
27+
28+
test_config:
29+
is_slide: True
30+
crop_size: [1024, 1024]
31+
stride: [768, 768]

‎configs/segformer/segformer_b1_cityscapes_1024x512_160k.yml

+26
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,26 @@
1+
_base_: '../_base_/cityscapes.yml'
2+
3+
batch_size: 1
4+
iters: 160000
5+
6+
model:
7+
type: SegFormer_B1
8+
num_classes: 19
9+
pretrained: https://bj.bcebos.com/paddleseg/dygraph/mix_vision_transformer_b1.tar.gz
10+
11+
optimizer:
12+
_inherited_: False
13+
type: AdamW
14+
beta1: 0.9
15+
beta2: 0.999
16+
weight_decay: 0.01
17+
18+
lr_scheduler:
19+
type: PolynomialDecay
20+
learning_rate: 0.00006
21+
power: 1
22+
23+
loss:
24+
types:
25+
- type: CrossEntropyLoss
26+
coef: [1]

‎configs/segformer/segformer_b2_cityscapes_1024x1024_160k.yml

+31
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,31 @@
1+
_base_: '../_base_/cityscapes_1024x1024.yml'
2+
3+
batch_size: 2
4+
iters: 160000
5+
6+
model:
7+
type: SegFormer_B2
8+
num_classes: 19
9+
pretrained: https://bj.bcebos.com/paddleseg/dygraph/mix_vision_transformer_b2.tar.gz
10+
11+
optimizer:
12+
_inherited_: False
13+
type: AdamW
14+
beta1: 0.9
15+
beta2: 0.999
16+
weight_decay: 0.01
17+
18+
lr_scheduler:
19+
type: PolynomialDecay
20+
learning_rate: 0.00006
21+
power: 1
22+
23+
loss:
24+
types:
25+
- type: CrossEntropyLoss
26+
coef: [1]
27+
28+
test_config:
29+
is_slide: True
30+
crop_size: [1024, 1024]
31+
stride: [768, 768]

‎configs/segformer/segformer_b2_cityscapes_1024x512_160k.yml

+26
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,26 @@
1+
_base_: '../_base_/cityscapes.yml'
2+
3+
batch_size: 1
4+
iters: 160000
5+
6+
model:
7+
type: SegFormer_B2
8+
num_classes: 19
9+
pretrained: https://bj.bcebos.com/paddleseg/dygraph/mix_vision_transformer_b2.tar.gz
10+
11+
optimizer:
12+
_inherited_: False
13+
type: AdamW
14+
beta1: 0.9
15+
beta2: 0.999
16+
weight_decay: 0.01
17+
18+
lr_scheduler:
19+
type: PolynomialDecay
20+
learning_rate: 0.00006
21+
power: 1
22+
23+
loss:
24+
types:
25+
- type: CrossEntropyLoss
26+
coef: [1]

‎configs/segformer/segformer_b3_cityscapes_1024x1024_160k.yml

+31
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,31 @@
1+
_base_: '../_base_/cityscapes_1024x1024.yml'
2+
3+
batch_size: 2
4+
iters: 160000
5+
6+
model:
7+
type: SegFormer_B3
8+
num_classes: 19
9+
pretrained: https://bj.bcebos.com/paddleseg/dygraph/mix_vision_transformer_b3.tar.gz
10+
11+
optimizer:
12+
_inherited_: False
13+
type: AdamW
14+
beta1: 0.9
15+
beta2: 0.999
16+
weight_decay: 0.01
17+
18+
lr_scheduler:
19+
type: PolynomialDecay
20+
learning_rate: 0.00006
21+
power: 1
22+
23+
loss:
24+
types:
25+
- type: CrossEntropyLoss
26+
coef: [1]
27+
28+
test_config:
29+
is_slide: True
30+
crop_size: [1024, 1024]
31+
stride: [768, 768]

‎configs/segformer/segformer_b3_cityscapes_1024x512_160k.yml

+26
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,26 @@
1+
_base_: '../_base_/cityscapes.yml'
2+
3+
batch_size: 1
4+
iters: 160000
5+
6+
model:
7+
type: SegFormer_B3
8+
num_classes: 19
9+
pretrained: https://bj.bcebos.com/paddleseg/dygraph/mix_vision_transformer_b3.tar.gz
10+
11+
optimizer:
12+
_inherited_: False
13+
type: AdamW
14+
beta1: 0.9
15+
beta2: 0.999
16+
weight_decay: 0.01
17+
18+
lr_scheduler:
19+
type: PolynomialDecay
20+
learning_rate: 0.00006
21+
power: 1
22+
23+
loss:
24+
types:
25+
- type: CrossEntropyLoss
26+
coef: [1]

‎configs/segformer/segformer_b4_cityscapes_1024x1024_160k.yml

+31
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,31 @@
1+
_base_: '../_base_/cityscapes_1024x1024.yml'
2+
3+
batch_size: 2
4+
iters: 160000
5+
6+
model:
7+
type: SegFormer_B4
8+
num_classes: 19
9+
pretrained: https://bj.bcebos.com/paddleseg/dygraph/mix_vision_transformer_b4.tar.gz
10+
11+
optimizer:
12+
_inherited_: False
13+
type: AdamW
14+
beta1: 0.9
15+
beta2: 0.999
16+
weight_decay: 0.01
17+
18+
lr_scheduler:
19+
type: PolynomialDecay
20+
learning_rate: 0.00006
21+
power: 1
22+
23+
loss:
24+
types:
25+
- type: CrossEntropyLoss
26+
coef: [1]
27+
28+
test_config:
29+
is_slide: True
30+
crop_size: [1024, 1024]
31+
stride: [768, 768]

‎configs/segformer/segformer_b4_cityscapes_1024x512_160k.yml

+26
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,26 @@
1+
_base_: '../_base_/cityscapes.yml'
2+
3+
batch_size: 1
4+
iters: 160000
5+
6+
model:
7+
type: SegFormer_B4
8+
num_classes: 19
9+
pretrained: https://bj.bcebos.com/paddleseg/dygraph/mix_vision_transformer_b4.tar.gz
10+
11+
optimizer:
12+
_inherited_: False
13+
type: AdamW
14+
beta1: 0.9
15+
beta2: 0.999
16+
weight_decay: 0.01
17+
18+
lr_scheduler:
19+
type: PolynomialDecay
20+
learning_rate: 0.00006
21+
power: 1
22+
23+
loss:
24+
types:
25+
- type: CrossEntropyLoss
26+
coef: [1]

‎configs/segformer/segformer_b5_cityscapes_1024x1024_160k.yml

+31
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,31 @@
1+
_base_: '../_base_/cityscapes_1024x1024.yml'
2+
3+
batch_size: 1
4+
iters: 160000
5+
6+
model:
7+
type: SegFormer_B5
8+
num_classes: 19
9+
pretrained: https://bj.bcebos.com/paddleseg/dygraph/mix_vision_transformer_b5.tar.gz
10+
11+
optimizer:
12+
_inherited_: False
13+
type: AdamW
14+
beta1: 0.9
15+
beta2: 0.999
16+
weight_decay: 0.01
17+
18+
lr_scheduler:
19+
type: PolynomialDecay
20+
learning_rate: 0.00006
21+
power: 1
22+
23+
loss:
24+
types:
25+
- type: CrossEntropyLoss
26+
coef: [1]
27+
28+
test_config:
29+
is_slide: True
30+
crop_size: [1024, 1024]
31+
stride: [768, 768]

‎configs/segformer/segformer_b5_cityscapes_1024x512_160k.yml

+26
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,26 @@
1+
_base_: '../_base_/cityscapes.yml'
2+
3+
batch_size: 1
4+
iters: 160000
5+
6+
model:
7+
type: SegFormer_B5
8+
num_classes: 19
9+
10+
optimizer:
11+
_inherited_: False
12+
type: AdamW
13+
beta1: 0.9
14+
beta2: 0.999
15+
weight_decay: 0.01
16+
17+
lr_scheduler:
18+
type: PolynomialDecay
19+
learning_rate: 0.00006
20+
power: 1
21+
22+
23+
loss:
24+
types:
25+
- type: CrossEntropyLoss
26+
coef: [1]

‎docs/design/use/use.md

+2-2
Original file line numberDiff line numberDiff line change
@@ -28,9 +28,9 @@
2828
### optimizer
2929
> 训练优化器
3030
> * 参数
31-
> * type : 优化器类型，目前只支持'sgd'和'adam'
32-
> * momentum : 动量
31+
> * type : 优化器类型，支持目前Paddle官方所有优化器
3332
> * weight_decay : L2正则化的值
33+
> * **others** : 请参考[Paddle官方Optimizer文档](https://www.paddlepaddle.org.cn/documentation/docs/zh/api/paddle/optimizer/Overview_cn.html)
3434
3535
----
3636
### lr_scheduler

0 commit comments

Comments
 (0)