Partial Domain Adaptation

Following is a Keras implementation of a Partial Domain Adaptation Model that assigns class labels to images in the Target domain by extracting domain-invariant features from the labelled Source and unlabelled Target domain images. This work caters towards a realistic scenario which relaxes the shared label space assumption (i.e., the target label space is a subspace of the source label space) by introducing a class-level weighting mechanism to remove the negative influence of outlier source classes. The architecture involves three sub-networks: (a) domain-invariant feature extractor, (b) classifier and (c) domain discriminator.

License

Dataset Download

The code is tested on the Office-31 dataset.

• Download it from this link: Office-31.
• Create a domain_adaptation_images directory under Data/Office/ and place the downloaded images inside.
• Image paths, along with the corresponding class labels, are listed in the .txt files present in the Office directory.
• The Data directory structure should look like this:

.
├── Data
|     └── Office
|          ├── domain_adaptation_images 
|          |           └── place the office-31 image folders here (amazon, dslr, webcam)
|          ├── amazon_10_list.txt
|          ├── amazon_31_list.txt
|          ├── dslr_10_list.txt
|          ├── dslr_31_list.txt
|          ├── webcam_31_list.txt
|          └── webcam_10_list.txt
└── ...

Requirements

This code is compatible with the mentioned versions of the following libraries. However, it might also be compatible with their prior versions.

pillow 6.0.0
scikit-learn 0.20.2
tensorflow 1.12.0
keras 2.2.4

Model Specifications

Domain-Invariant Feature Extractor:

• Model : Resnet50 (initialized with 'imagenet' weights)
• Loss : w₁ * classifier loss - w₂ * discriminator loss (w₁ = classifier loss weight and w₂ = discriminator loss weight)
  • The gradient-reversal process is done in a GAN fashion where the domain labels are flipped for the Source and Target (i.e. Source labels changed to 1 and Target labels changed to 0).
• Optimizer : Adam

Classifier:

• Model : Dense(400) --> Dense(100) --> Dense(number of classes)
• Loss : Categorical cross-entropy
• Optimizer : Adam

Domain Discriminator:

• Model : Dense(400) --> Dense(100) --> Dense(1)
• Loss : Binary cross-entropy
• Optimizer : Adam

Model Run

The model outputs the Target_labels and the Source and Target classification accuracies, which are saved under the Snapshot/Models directory.

• An example starter code is given below:

python driver.py --batch_size 32 --number_of_gpus 2 --lr_combined 0.00001 --num_iterations 5000

• Parameters:

Parameter Name	Default Value	Description
--number_of_gpus	default = '1'	"Number of gpus required to run"
--network_name	default = 'ResNet50'	"Name of the feature extractor network"
--dataset_name	default = 'Office'	"Name of the source dataset"
--dropout_classifier	default = 0.25	"Dropout ratio for classifier"
--dropout_discriminator	default = 0.25	"Dropout ratio for discriminator"
--source_path	default = 'amazon_10_list.txt'	"Path to source dataset"
--target_path	default = 'webcam_10_list.txt'	"Path to target dataset"
--lr_classifier	default = 0.0001	"Learning rate for classifier model"
--lr_discriminator	default = 0.0001	"Learning rate for discriminator model"
--lr_combined	default = 0.00001	"Learning rate for combined model"
--b1_classifier	default = 0.9	"Exponential decay rate of first moment for classifier model optimizer"
--b2_classifier	default = 0.999	"Exponential decay rate of second moment for classifier model optimizer"
--b1_discriminator	default = 0.9	"Exponential decay rate of first moment for discriminator model optimizer"
--b2_discriminator	default = 0.999	"Exponential decay rate of second moment for discriminator model optimizer"
--b1_combined	default = 0.9	"Exponential decay rate of first moment for combined model optimizer"
--b2_combined	default = 0.999	"Exponential decay rate of second moment for combined model optimizer"
--classifier_loss_weight	default = 1	"Classifier loss weight"
--discriminator_loss_weight	default = 4	"Discriminator loss weight"
--batch_size	default = 32	"Batch size for training"
--test_interval	default = 3	"Gap between two successive test phases"
--num_iterations	default = 12000	"Number of iterations"
--snapshot_interval	default = 500	"Minimum gap between saving output models"
--output_dir	default = 'models'	"Directory for saving output model"

Acknowledgements

The following code has drawn inspiration from the following papers:

• Zhangjie Cao, Mingsheng Long, Jianmin Wang and Michael I. Jordan. "Partial Transfer Learning with Selective Adversarial Networks." [Link]

• Jing Zhang, Zewei Ding, Wanqing Li and Philip Ogunbona. "Importance Weighted Adversarial Nets for Partial Domain Adaptation." [Link]

• Ganin, Yaroslav and Lempitsky. "Unsupervised domain adaptation by backpropagation." [Link]

• Li, Yanghao, Naiyan Wang, Jianping Shi, Jiaying Liu, and Xiaodi Hou. "Revisiting batch normalization for practical domain adaptation." [Link]

• Tzeng, Eric, Judy Hoffman, Kate Saenko, and Trevor Darrell. "Adversarial discriminative domain adaptation." [Link]