/deep-transfer-learning

A collection of implementations of deep domain adaptation algorithms

Primary LanguagePython

Deep Transfer Learning on PyTorch

This is a PyTorch library for deep transfer learning. We divide the code into two aspects: Single-source Unsupervised Domain Adaptation (SUDA) and Multi-source Unsupervised Domain Adaptation (MUDA). There are many SUDA methods, however I find there is a few MUDA methods with deep learning. Besides, MUDA with deep learning might be a more promising direction for domain adaptation.

Here I have implemented some deep transfer methods as follows:

  • UDA
    • DDC:Deep Domain Confusion Maximizing for Domain Invariance
    • DAN: Learning Transferable Features with Deep Adaptation Networks (ICML2015)
    • Deep Coral: Deep CORAL Correlation Alignment for Deep Domain Adaptation (ECCV2016)
    • Revgrad: Unsupervised Domain Adaptation by Backpropagation (ICML2015)
    • MRAN: Multi-representation adaptation network for cross-domain image classification (Neural Network 2019)
    • DSAN: Deep Subdomain Adaptation Network for Image Classification (IEEE Transactions on Neural Networks and Learning Systems 2020)
  • MUDA
    • Aligning Domain-specific Distribution and Classifier for Cross-domain Classification from Multiple Sources (AAAI2019)
  • Application
    • Cross-domain Fraud Detection: Modeling Users’ Behavior Sequences with Hierarchical Explainable Network for Cross-domain Fraud Detection (WWW2020)
  • Survey

Results on Office31(UDA)

Method A - W D - W W - D A - D D - A W - A Average
ResNet 68.4±0.5 96.7±0.5 99.3±0.1 68.9±0.2 62.5±0.3 60.7±0.3 76.1
DDC 75.8±0.2 95.0±0.2 98.2±0.1 77.5±0.3 67.4±0.4 64.0±0.5 79.7
DDC* 78.3±0.4 97.1±0.1 100.0±0.0 81.7±0.9 65.2±0.6 65.1±0.4 81.2
DAN 83.8±0.4 96.8±0.2 99.5±0.1 78.4±0.2 66.7±0.3 62.7±0.2 81.3
DAN* 82.6±0.7 97.7±0.1 100.0±0.0 83.1±0.9 66.8±0.3 66.6±0.4 82.8
DCORAL* 79.0±0.5 98.0±0.2 100.0±0.0 82.7±0.1 65.3±0.3 64.5±0.3 81.6
Revgrad 82.0±0.4 96.9±0.2 99.1±0.1 79.7±0.4 68.2±0.4 67.4±0.5 82.2
Revgrad* 82.6±0.9 97.8±0.2 100.0±0.0 83.3±0.9 66.8±0.1 66.1±0.5 82.8
MRAN 91.4±0.1 96.9±0.3 99.8±0.2 86.4±0.6 68.3±0.5 70.9±0.6 85.6
DSAN 93.6±0.2 98.4±0.1 100.0±0.0 90.2±0.7 73.5±0.5 74.8±0.4 88.4

Note that the results without '*' comes from paper. The results with '*' are run by myself with the code.

Results on Office31(MUDA)

Standards Method A,W - D A,D - W D,W - A Average
ResNet 99.3 96.7 62.5 86.2
DAN 99.5 96.8 66.7 87.7
Single Best DCORAL 99.7 98.0 65.3 87.7
RevGrad 99.1 96.9 68.2 88.1
DAN 99.6 97.8 67.6 88.3
Source Combine DCORAL 99.3 98.0 67.1 88.1
RevGrad 99.7 98.1 67.6 88.5
Multi-Source MFSAN 99.5 98.5 72.7 90.2

Results on OfficeHome(MUDA)

Standards Method C,P,R - A A,P,R - C A,C,R - P A,C,P - R Average
ResNet 65.3 49.6 79.7 75.4 67.5
DAN 64.1 50.8 78.2 75.0 67.0
Single Best DCORAL 68.2 56.5 80.3 75.9 70.2
RevGrad 67.9 55.9 80.4 75.8 70.0
DAN 68.5 59.4 79.0 82.5 72.4
Source Combine DCORAL 68.1 58.6 79.5 82.7 72.2
RevGrad 68.4 59.1 79.5 82.7 72.4
Multi-Source MFSAN 72.1 62.0 80.3 81.8 74.1

Note that (1) Source combine: all source domains are combined together into a traditional single-source v.s. target setting. (2) Single best: among the multiple source domains, we report the best single source transfer results. (3) Multi-source: the results of MUDA methods.

Note

If you find that your accuracy is 100%, the problem might be the dataset folder. Please note that the folder structure required for the data provider to work is:

-dataset
    -amazon
    -webcam
    -dslr

Contact

If you have any problem about this library, please create an Issue or send us an Email at: