复现论文：DPStyler: Dynamic PromptStyler for Source-Free Domain Generalization

论文：[2403.16697] DPStyler: Dynamic PromptStyler for Source-Free Domain Generalization

github: 

TYLfromSEU/DPStyler: DPStyler: Dynamic PromptStyler for Source-Free Domain Generalization

论文:

这篇论文还是在PromptStyler:Prompt-driven Style Generation for Source-free Domain Generalization基础上对一些架构进行修改的。

Contribution

Style Generation Module:

dynamically refresh style by regenerating all style word vectors including Random and StyleMix between epochs.(randomly select one method)

Style Removal Module:

propose a domain uncertainty loss after encoder, to reduce specific style information and learn domain-invariant features

Model Ensemble:inference stage

for each text template, train a model and each model is involved in the inference process

Because introducing more randomness into the training process

大致代码流程

 

复现结果

	PACS	VLCS	OfficeHome	DomainNet	avg
ResNet-50	92.1__-1.5	81.4__-2.1	70.5__-2.0	45.5__-2.5	72.4__-2.0
ViT-B/16	96.6__-0.5	80.8__-3.2		56.9__-2.0

PACS:

	Art Painting	Cartoon	Photo	Sketch	avg
ResNet-50	92.7__-1.3	96.4__+1.1	99.6__0	79.6__-5.9	92.1__-1.5
ViT-B/16	97.6__-0.1	99.2__+0.3	99.9__0	89.8__-2.1	96.6__-0.5

ResNet-50：

92.4 96.5 99.6 78.4

93.4 96.2 99.5 81.7

92.2 96.5 99.6 78.6

ViT-B/16：

97.6 99.3 99.9 89.0

97.6 99.1 99.9 90.8

97.6 99.2 99.9 89.6

VLCS:

VLCS	Caletch	LableMe	SUN09	VOC2007	avg
ResNet-50	100__0	65.0__-8.3	73.9__+2.1	86.5__-2.5	81.4__-2.1
ViT-B/16	100__0	59.7__-9.4	80.4__+4.1	82.9__-7.6	80.8__-3.2

ResNet-50：

100 67.9 87.2 74.2 82.325

100 62.2 85.7 73.0 80.225

100 64.4 86.7 74.4 81.375

81.304

ViT-B/16：

100 62.4 88.5 77.6 82.125

100 61.1 83.1 75.5 79.925

100 55.7 77 88 80.175

80.7417

Office-Home:

Office-Home	Art	Clipart	Product	Real World	avg
ResNet-50	72.3__-1.3	49.1__-2.1	78.8__-3.3	81.6__-1.3	70.5__-2.0
ViT-B/16

ResNet-50：

72.2 49 78.5 81.5

72.4 48.9 78.6 81.5

72.4 49.3 79.3 81.7

ViT-B/16：

DomainNet:

DomainNet	Clipart	Infograph	Painting	Quickdraw	Real	Sketch	avg
ResNet-50	51.6__-3.8	40.6__-0.9	53.4__-2.6	6.7__-0.4	73.4__-4.1	47.3__-2.4	45.5__-2.5
ViT-B/16	68.6__-2.7	49.5__-1.1	65.3__-1.5	15.8__-0.4	80.7__-3.0	61.7__-2.7	56.9__-2.0

ResNet-50：

51.5 40.7 53.5 6.7 73.6 47.3

51.7 40.4 53.4 6.7 73.3 47.3

51.5 40.7 53.4 6.8 73.2 47.4

ViT-B/16

68.7 49.6 65.3 15.9 80.8 61.8

68.4 49.4 65.2 15.9 80.6 61.6

68.8 49.5 65.3 15.7 80.8 61.7

报错:

1、ERROR: Ignored the following versions that require a different python version:

23.10.0 Requires-Python >=3.8; 23.10.1 Requires-Python >=3.8; 23.11.0 Requires-Python >=3.8; 23.12.0 Requires-Python >=3.8; 23.12.1 Requires-Python >=3.8; 23.7.0 Requires-Python >=3.8; 23.9.0 Requires-Python >=3.8; 23.9.1 Requires-Python >=3.8; 24.1.0 Requires-Python >=3.8; 24.1.1 Requires-Python >=3.8; 24.10.0 Requires-Python >=3.9; 24.1a1 Requires-Python >=3.8; 24.2.0 Requires-Python >=3.8; 24.3.0 Requires-Python >=3.8; 24.4.0 Requires-Python >=3.8; 24.4.1 Requires-Python >=3.8; 24.4.2 Requires-Python >=3.8; 24.8.0 Requires-Python >=3.8 ERROR: Could not find a version that satisfies the requirement clip==1.0 (from versions: 0.0.1, 0.1.0, 0.2.0) ERROR: No matching distribution found for clip==1.0 failed CondaEnvException: Pip failed

原因：直接使用作者github上面的environment.yml文件发生报错。里面的配置有的用python3.7，有的用3.8很混乱。搞不懂作者是怎么配置的。

但是使用我之前配置的虚拟环境也可以这场运行这篇论文的代码。

链接：zhanghr2001/PromptTA: Source-free Domain Generalization

这是我之前复现这篇论文的环境，根据readme，配置到requirement.txt就可以。 

2、连接超时Connection timed out'

requests.exceptions.ConnectionError: HTTPSConnectionPool(host='drive.google.com', port=443): Max retries exceeded with url: /uc?id=1gkbf_KaxoBws-GWT3XIPZ7BnkqbAxIFa (Caused by NewConnectionError('<urllib3.connection.HTTPSConnection object at 0x7f20626e8990>: Failed to establish a new connection: [Errno 110] Connection timed out'))

原因：要在作者给的github上的链接上面下载数据集，我用的之前的数据集，有一些问题。

3、ModuleNotFoundError: No module named ‘clip‘

命令：

pip install clip
pip install openai-clip

相关设置

1、数据集和模型名称设置

数据集：

PACS    VLCS     office_home_dg    domainnet：6个域文件+split文件夹（每个域的_train.txt和_test.txt）

模型名称：NET=resnet50_clip
vitb16_clip
vitl14_clip

2、运行命令

复现DPStyler命令：
sh run_clip_pacs.sh
sh run_clip_vlcs.sh
sh run_clip_office_home.sh
sh run_clip_domainnet.sh

这个作者还提供了复现PromptStyler的代码：

这个作者复现PromptStyler命令：
sh run_clip_pacs_ps.sh
sh run_clip_vlcs_ps.sh
sh run_clip_office_home_ps.sh
sh run_clip_domainnet_ps.sh

3、tensorboard命令查看

tensorboard --logdir ./output/vlcs/NS80_lr5_arcface_5_0.5_E100_resnet50_clip_R_wohead/seed3/tensorboard  --load_fast true  --port 6007

若显示端口被占用再加上 --load_fast true  --port 6007