Init commit.

README.md

@@ -0,0 +1,98 @@
+# Semi-Supervised Domain Generalization in Real World: New Benchmark and Strong Baseline
+
+[paper](https://arxiv.org/pdf/2111.10221.pdf)
+
+## Requirements
+
+- numpy==1.19.2
+- Pillow==9.1.1
+- PyYAML==6.0
+- scikit_learn==1.1.1
+- six==1.15.0
+- torch==1.7.1
+- torchvision==0.8.2
+
+## Installation
+
+```
+git clone https://github.com/MetaVisionLab/SSDG.git
+cd SSDG
+pip install -r requirements.txt
+```
+
+## Data Preparation
+
+You can download the dataset to the folder  SSDG/data,which include three folders representing three datasets in our paper
+
+**PACS**
+
+Download the dataset [PACS](https://drive.google.com/file/d/1m4X4fROCCXMO0lRLrr6Zz9Vb3974NWhE/view) to data/pacs and unzip it(this dataset link directly provides images and splits)
+
+File structure:
+
+```
+pacs/
+|–– images/
+|–– splits/
+```
+
+**Digits-DG**
+
+Since we provide the dataset splits in this repo,you just need to download the dataset [Digits-DG](https://drive.google.com/file/d/15V7EsHfCcfbKgsDmzQKj_DfXt_XYp_P7/view) to data/digits_dg/images and unzip it
+
+File structure:
+
+```
+digits_dg/
+|–– images/
+|–– splits/
+```
+
+**Office-Home-DG**
+
+Since we provide the dataset splits in this repo,you just need to download the dataset [Digits-DG](https://drive.google.com/file/d/15V7EsHfCcfbKgsDmzQKj_DfXt_XYp_P7/view) to data/digits_dg/images and unzip it
+
+File structure:
+
+```
+office_home_dg/
+|–– images/
+|–– splits/
+```
+
+## Training
+
+```
+sh pacs.sh       #train pacs
+sh digits.sh     #train digits_dg
+sh officeHome.sh   #train office_home_dg
+```
+
+# Validation
+
+We evaluate our method on all the SSDG tasks for each dataset and report the average accuracy.For each task(for example A2C),we report its average accuracy on last 5 epoch.
+
+```
+cd tools/
+python parse_test_res_single.py log/UDAG_A2C.log --test-log
+```
+
+# Citation
+
+Please cite our paper:
+
+```
+@article{lin2021ssdg,
+  title={Semi-Supervised Domain Generalization in Real World: New Benchmark and Strong Baseline},
+  author={Lin, Luojun and Xie, Han and Yang, Zhifeng and Sun, Zhishu and Liu, Wenxi and Yu, Yuanlong and Chen, Weijie and Yang, Shicai and Xie, Di},
+  journal={arXiv preprint arXiv:2111.10221},
+  year={2021}
+}
+```
+# Contact us
+
+For any questions, please feel free to contact  [Han Xie](mailto:[email protected]) or Dr. [Luojun Lin](mailto:[email protected]).
+
+# Copyright
+
+This code is free to the academic community for research purpose only. For commercial purpose usage, please contact Dr. [Luojun Lin](mailto:[email protected]).

config/A2C.yaml

@@ -0,0 +1,32 @@
+data_transforms:
+  input_size: 224
+
+trainer:
+  n_step_f: 4 # MCD N_STEP_F
+  batch_size: 128
+#  max_epochs: 200
+  num_workers: 4
+  save_model_addr: 'checkpoints/A2C'
+  forget_rate: 0.5
+
+optimizer:
+  name: 'SGD'
+  params:
+    lr: 0.001
+    momentum: 0.9
+    weight_decay: 0.0005
+
+data:
+  type: 'pacs'
+  class_number: 7
+  root: 'data/pacs'
+  source_domain_x: 'art_painting'
+  source_domain_u_1: 'sketch'
+  source_domain_u_2: 'photo'
+  target_domain: 'cartoon'
+
+log:
+  save_addr: 'log'
+  save_name: 'UDAG_A2C'
+
+ratio: 0.4 # a ratio to split samples into clean set and noisy set

config/A2P.yaml

@@ -0,0 +1,32 @@
+data_transforms:
+  input_size: 224
+
+trainer:
+  n_step_f: 4 # MCD N_STEP_F
+  batch_size: 128
+#  max_epochs: 200
+  num_workers: 4
+  save_model_addr: 'checkpoints/A2P'
+  forget_rate: 0.5
+
+optimizer:
+  name: 'SGD'
+  params:
+    lr: 0.001
+    momentum: 0.9
+    weight_decay: 0.0005
+
+data:
+  type: 'pacs'
+  class_number: 7
+  root: 'data/pacs'
+  source_domain_x: 'art_painting'
+  source_domain_u_1: 'sketch'
+  source_domain_u_2: 'cartoon'
+  target_domain: 'photo'
+
+log:
+  save_addr: 'log'
+  save_name: 'UDAG_A2P'
+
+ratio: 0.4 # a ratio to split samples into clean set and noisy set

config/A2S.yaml

@@ -0,0 +1,32 @@
+data_transforms:
+  input_size: 224
+
+trainer:
+  n_step_f: 4 # MCD N_STEP_F
+  batch_size: 128
+#  max_epochs: 200
+  num_workers: 4
+  save_model_addr: 'checkpoints/A2S'
+  forget_rate: 0.5
+
+optimizer:
+  name: 'SGD'
+  params:
+    lr: 0.001
+    momentum: 0.9
+    weight_decay: 0.0005
+
+data:
+  type: 'pacs'
+  class_number: 7
+  root: 'data/pacs'
+  source_domain_x: 'art_painting'
+  source_domain_u_1: 'photo'
+  source_domain_u_2: 'cartoon'
+  target_domain: 'sketch'
+
+log:
+  save_addr: 'log'
+  save_name: 'UDAG_A2S'
+
+ratio: 0.4 # a ratio to split samples into clean set and noisy set

config/C2A.yaml

@@ -0,0 +1,32 @@
+data_transforms:
+  input_size: 224
+
+trainer:
+  n_step_f: 4 # MCD N_STEP_F
+  batch_size: 128
+#  max_epochs: 200
+  num_workers: 4
+  save_model_addr: 'checkpoints/C2A'
+  forget_rate: 0.5
+
+optimizer:
+  name: 'SGD'
+  params:
+    lr: 0.001
+    momentum: 0.9
+    weight_decay: 0.0005
+
+data:
+  type: 'pacs'
+  class_number: 7
+  root: 'data/pacs'
+  source_domain_x: 'cartoon'
+  source_domain_u_1: 'sketch'
+  source_domain_u_2: 'photo'
+  target_domain: 'art_painting'
+
+log:
+  save_addr: 'log'
+  save_name: 'UDAG_C2A'
+
+ratio: 0.4 # a ratio to split samples into clean set and noisy set

config/C2P.yaml

@@ -0,0 +1,32 @@
+data_transforms:
+  input_size: 224
+
+trainer:
+  n_step_f: 4 # MCD N_STEP_F
+  batch_size: 128
+#  max_epochs: 200
+  num_workers: 4
+  save_model_addr: 'checkpoints/C2P'
+  forget_rate: 0.5
+
+optimizer:
+  name: 'SGD'
+  params:
+    lr: 0.001
+    momentum: 0.9
+    weight_decay: 0.0005
+
+data:
+  type: 'pacs'
+  class_number: 7
+  root: 'data/pacs'
+  source_domain_x: 'cartoon'
+  source_domain_u_1: 'sketch'
+  source_domain_u_2: 'art_painting'
+  target_domain: 'photo'
+
+log:
+  save_addr: 'log'
+  save_name: 'UDAG_C2P'
+
+ratio: 0.4 # a ratio to split samples into clean set and noisy set

config/C2S.yaml

@@ -0,0 +1,32 @@
+data_transforms:
+  input_size: 224
+
+trainer:
+  n_step_f: 4 # MCD N_STEP_F
+  batch_size: 128
+#  max_epochs: 200
+  num_workers: 4
+  save_model_addr: 'checkpoints/C2S'
+  forget_rate: 0.5
+
+optimizer:
+  name: 'SGD'
+  params:
+    lr: 0.001
+    momentum: 0.9
+    weight_decay: 0.0005
+
+data:
+  type: 'pacs'
+  class_number: 7
+  root: 'data/pacs'
+  source_domain_x: 'cartoon'
+  source_domain_u_1: 'art_painting'
+  source_domain_u_2: 'photo'
+  target_domain: 'sketch'
+
+log:
+  save_addr: 'log'
+  save_name: 'UDAG_C2S'
+
+ratio: 0.4 # a ratio to split samples into clean set and noisy set

config/P2A.yaml

@@ -0,0 +1,36 @@
+data_transforms:
+  input_size: 224
+
+trainer:
+  n_step_f: 4 # MCD N_STEP_F
+  batch_size: 128
+#  max_epochs: 200
+  num_workers: 4
+  save_model_addr: 'checkpoints/P2A'
+  forget_rate: 0.5
+
+optimizer:
+  name: 'SGD'
+  params:
+    lr: 0.001
+    momentum: 0.9
+    weight_decay: 0.0005
+
+data:
+  type: 'pacs'
+  class_number: 7
+  root: 'data/pacs'
+  source_domain_x: 'photo'
+  source_domain_u_1: 'sketch'
+  source_domain_u_2: 'cartoon'
+  target_domain: 'art_painting'
+
+log:
+  save_addr: 'log'
+  save_name: 'UDAG_P2A'
+
+ratio: 0.4 # a ratio to split samples into clean set and noisy set
+
+
+
+

config/P2C.yaml

@@ -0,0 +1,36 @@
+data_transforms:
+  input_size: 224
+
+trainer:
+  n_step_f: 4 # MCD N_STEP_F
+  batch_size: 128
+#  max_epochs: 200
+  num_workers: 4
+  save_model_addr: 'checkpoints/P2C'
+  forget_rate: 0.5
+
+optimizer:
+  name: 'SGD'
+  params:
+    lr: 0.001
+    momentum: 0.9
+    weight_decay: 0.0005
+
+data:
+  type: 'pacs'
+  class_number: 7
+  root: 'data/pacs'
+  source_domain_x: 'photo'
+  source_domain_u_1: 'sketch'
+  source_domain_u_2: 'art_painting'
+  target_domain: 'cartoon'
+
+log:
+  save_addr: 'log'
+  save_name: 'UDAG_P2C'
+
+ratio: 0.4 # a ratio to split samples into clean set and noisy set
+
+
+
+