Skip to content

Commit bf93244

Browse files
marshall-dteachmarshall-dteach
authored
【Seaformer&EfficientFormerV2】添加其他模型配置 (#3306)
1 parent 3fea7bb commit bf93244

File tree

9 files changed

+61
-9
lines changed

9 files changed

+61
-9
lines changed

‎configs/efficientformerv2/README.md

Lines changed: 1 addition & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -10,4 +10,5 @@
1010

1111
| Model | size | Head | Resolution | Training Iters | mIoU (slice) | Links |
1212
| :---------------: | :--: | :----: | :--------: | :------------: | :----------: | :-----------------------------: |
13+
| EfficientFormerV2 | s0 | FPNNet | 512x512 | 40000 | 33.43 | [model](https://bj.bcebos.com/paddleseg/dygraph/ade20k/efficientformerv2_s0_ade20k/model.pdparams) \| [log](https://bj.bcebos.com/paddleseg/dygraph/ade20k/efficientformerv2_s0_ade20k/train.log) \| [vdl](https://www.paddlepaddle.org.cn/paddle/visualdl/service/app/index?id=d55127711d121f8451e21679fc8a7b35) |
1314
| EfficientFormerV2 | s2 | FPNNet | 512x512 | 40000 | 42.16 | [model](https://bj.bcebos.com/paddleseg/dygraph/ade20k/efficientformerv2_s2_ade20k_512x512_40k/model.pdparams) \| [log](https://bj.bcebos.com/paddleseg/dygraph/ade20k/efficientformerv2_s2_ade20k_512x512_40k/train.log) \| [vdl](https://www.paddlepaddle.org.cn/paddle/visualdl/service/app/scalar?id=7397511b1f7607510aaafc218e4ec90d) |

‎configs/efficientformerv2/efficientformerv2_s0_ade20k_512x512_40k.yml

Lines changed: 6 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,6 @@
1+
_base_: 'efficientformerv2_s2_ade20k_512x512_40k.yml'
2+
3+
model:
4+
backbone:
5+
type: EfficientFormerv2_s0
6+
pretrained: https://bj.bcebos.com/paddleseg/dygraph/ade20k/efficientformerv2_s0/pretrain/eformer_s0_pretrained.pdparams

‎configs/efficientformerv2/efficientformerv2_s2_ade20k_512x512_40k.yml

Lines changed: 1 addition & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -10,7 +10,7 @@ model:
1010
pretrained: https://bj.bcebos.com/paddleseg/dygraph/ade20k/efficientformerv2_s2/pretrain/eformer_s2_pretrained.pdparams
1111
num_classes: 150
1212
backbone_indices: [0, 1, 2, 3]
13-
channels: 128
13+
head_channels: 128
1414

1515
val_dataset:
1616
transforms:

‎configs/seaformer/README.md

Lines changed: 3 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -10,4 +10,7 @@
1010

1111
| Model | size | Head | Resolution | Training Iters | mIoU (slice) | Links |
1212
| :-: | :-: | :-: | :-: | :-: | :-: | :-: |
13+
| SeaFormer | tiny | LightHead | 512x512 | 160000 | 34.58 | [model](https://bj.bcebos.com/paddleseg/dygraph/ade20k/seaformer_tiny_ade20k_512x512_160k/model.pdparams) \| [log](https://bj.bcebos.com/paddleseg/dygraph/ade20k/seaformer_tiny_ade20k_512x512_160k/train.log) \| [vdl](https://www.paddlepaddle.org.cn/paddle/visualdl/service/app/index?id=72cba29a94f195914b47554466a1e7d8) |
14+
| SeaFormer | small | LightHead | 512x512 | 160000 | 38.73 | [model](https://bj.bcebos.com/paddleseg/dygraph/ade20k/seaformer_small_ade20k_512x512_160k/model.pdparam) \| [log](https://bj.bcebos.com/paddleseg/dygraph/ade20k/seaformer_small_ade20k_512x512_160k/train.log) \| [vdl](https://www.paddlepaddle.org.cn/paddle/visualdl/service/app/scalar?id=1c26a450e838aa1b731ebc5154941ca4) |
1315
| SeaFormer | base | LightHead | 512x512 | 160000 | 40.92 | [model](https://paddleseg.bj.bcebos.com/dygraph/ade20k/seaformer_ade20k_512x512_160k/model.pdparams) \| [log](https://paddleseg.bj.bcebos.com/dygraph/ade20k/seaformer_ade20k_512x512_160k/train.log) \| [vdl](https://www.paddlepaddle.org.cn/paddle/visualdl/service/app/scalar?id=f43d0185995605fb99da6597121a96f1) |
16+
| SeaFormer | large | LightHead | 512x512 | 160000 | 43.66 | [model](https://bj.bcebos.com/paddleseg/dygraph/ade20k/seaformer_large_ade20k_512x512_160k/model.pdparams) \| [log](https://bj.bcebos.com/paddleseg/dygraph/ade20k/seaformer_large_ade20k_512x512_160k/train.log) \| [vdl](https://www.paddlepaddle.org.cn/paddle/visualdl/service/app/index?id=904c92bf201dd41c9fe10e7211340a5e) |

‎configs/seaformer/seaformer_large_ade20k_512x512_160k.yml

Lines changed: 11 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,11 @@
1+
_base_: 'seaformer_base_ade20k_512x512_160k.yml'
2+
3+
model:
4+
type: SeaFormerSeg
5+
backbone:
6+
type: SeaFormer_large
7+
pretrained: https://bj.bcebos.com/paddleseg/dygraph/ade20k/backbone/seaformer_large_imagenet_pretrained.pdparams
8+
num_classes: 150
9+
head_channels: 192
10+
embed_dims: [128, 160, 192]
11+
in_index: [0, 1, 2, 3]

‎configs/seaformer/seaformer_small_ade20k_512x512_160k.yml

Lines changed: 14 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,14 @@
1+
_base_: 'seaformer_base_ade20k_512x512_160k.yml'
2+
3+
model:
4+
type: SeaFormerSeg
5+
backbone:
6+
type: SeaFormer_small
7+
pretrained: https://bj.bcebos.com/paddleseg/dygraph/ade20k/backbone/seaformer_small_imagenet_pretrained.pdparams
8+
num_classes: 150
9+
head_channels: 128
10+
embed_dims: [96, 128]
11+
is_dw: True
12+
dropout_ratio: 0.1
13+
align_corners: False
14+
input_transform: 'multiple_select'

‎configs/seaformer/seaformer_tiny_ade20k_512x512_160k.yml

Lines changed: 14 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,14 @@
1+
_base_: 'seaformer_base_ade20k_512x512_160k.yml'
2+
3+
model:
4+
type: SeaFormerSeg
5+
backbone:
6+
type: SeaFormer_tiny
7+
pretrained: https://bj.bcebos.com/paddleseg/dygraph/ade20k/backbone/seaformer_tiny_imagenet_pretrained.pdparams
8+
num_classes: 150
9+
head_channels: 96
10+
embed_dims: [64, 96]
11+
is_dw: True
12+
dropout_ratio: 0.1
13+
align_corners: False
14+
input_transform: 'multiple_select'

‎paddleseg/models/backbones/seaformer.py

Lines changed: 8 additions & 7 deletions
Original file line numberDiff line numberDiff line change
@@ -27,15 +27,16 @@ class SeaFormer(nn.Layer):
2727
The SeaFormer implementation based on PaddlePaddle.
2828
2929
The original article refers to
30-
Qiang Wang, et, al. "SEAFORMER: SQUEEZE-ENHANCED AXIAL TRANSFORMER FOR MOBILE SEMANTIC SEGMENTATION"
30+
Qiang Wan, et, al. "SEAFORMER: SQUEEZE-ENHANCED AXIAL TRANSFORMER FOR MOBILE SEMANTIC SEGMENTATION"
3131
(https://arxiv.org/pdf/2301.13156.pdf).
3232
3333
Args:
3434
in_channels (int, optional): The channels of input image. Default: 3.
3535
cfg (list[list], optional): Two values in the tuple indicate the indices of output of backbone. Defaulte:
3636
[[[3, 1, 16, 1], [3, 4, 32, 2], [3, 3, 32, 1]],
3737
[[5, 3, 64, 2], [5, 3, 64, 1]],
38-
[[3, 3, 128, 2], [3, 3, 128, 1]], [[5, 4, 192, 2]],
38+
[[3, 3, 128, 2], [3, 3, 128, 1]],
39+
[[5, 4, 192, 2]],
3940
[[3, 6, 256, 2]]]
4041
channels (list[int], optional): Number of channels hidden layer. Default: [16, 32, 64, 128, 192, 256].
4142
embed_dims (list, optional): The size of embedding dimensions. Default: [192, 256].
@@ -474,9 +475,9 @@ def SeaFormer_tiny(pretrained, **kwags):
474475
pretrained=pretrained,
475476
cfgs=[[[3, 1, 16, 1], [3, 4, 16, 2], [3, 3, 16, 1]],
476477
[[5, 3, 32, 2], [5, 3, 32, 1]], [[3, 3, 64, 2], [3, 3, 64, 1]],
477-
[[5, 3, 160, 2]], [[3, 6, 192, 2]]],
478+
[[5, 3, 128, 2]], [[3, 6, 160, 2]]],
478479
emb_dims=[128, 160],
479-
channels=[16, 24, 48, 96, 160, 192],
480+
channels=[16, 16, 32, 64, 128, 160],
480481
depths=[2, 2],
481482
num_heads=4,
482483
**kwags)
@@ -491,9 +492,9 @@ def SeaFormer_small(pretrained, **kwags):
491492
pretrained=pretrained,
492493
cfgs=[[[3, 1, 16, 1], [3, 4, 24, 2], [3, 3, 24, 1]],
493494
[[5, 3, 48, 2], [5, 3, 48, 1]], [[3, 3, 96, 2], [3, 3, 96, 1]],
494-
[[5, 3, 128, 2]], [[3, 6, 160, 2]]],
495+
[[5, 4, 160, 2]], [[3, 6, 192, 2]]],
495496
emb_dims=[160, 192],
496-
channels=[16, 16, 32, 64, 128, 160],
497+
channels=[16, 24, 48, 96, 160, 192],
497498
depths=[3, 3],
498499
num_heads=6,
499500
**kwags)
@@ -525,7 +526,7 @@ def SeaFormer_base(pretrained, **kwags):
525526
def SeaFormer_large(pretrained, **kwags):
526527
seaformer = SeaFormer(
527528
pretrained=pretrained,
528-
cfgs=[[[3, 3, 32, 1], [3, 4, 64, 2], [3, 3, 64, 1]],
529+
cfgs=[[[3, 3, 32, 1], [3, 4, 64, 2], [3, 4, 64, 1]],
529530
[[5, 4, 128, 2], [5, 4, 128, 1]],
530531
[[3, 4, 192, 2], [3, 4, 192, 1]], [[5, 4, 256, 2]],
531532
[[3, 6, 320, 2]]],

‎test_tipc/configs/efficientformerv2/train_infer_python.txt

Lines changed: 3 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -52,7 +52,9 @@ inference:deploy/python/infer.py
5252
--model_name:ocrnet_hrformer_small
5353
===========================infer_benchmark_params==========================
5454
random_infer_input:[{float32,[3,1024,2048]}]
55-
===========================disable_train_benchmark==========================
55+
===========================to_static_train_benchmark_params===========================
56+
to_static_train:--opts to_static_training=True
57+
===========================train_benchmark_params==========================
5658
batch_size:2|4
5759
fp_items:fp32|fp16
5860
epoch:500

0 commit comments

Comments
 (0)