Domain-Adaptation

This project contains code for different domain adaptation methods on Office31 dataset.

Available methods include: DANN, CDAN, MCD, CORAL, MMD

This project can be easily extended to use on other datasets or perform other adaptaion methods. (Check Code Structure to find out where you need to change.)

Prepare Environment

Install the requirements using conda from requirements-conda.txt (or using pip from requirements.txt).

My setting: conda 22.11.1 with Python 3.8

If you have trouble with installing torch check this link to find the currect version for your device.

Run

1. Go to src directory.
2. Run main.py with appropriate arguments.

Examples:

• Perform MMD and CORAL adaptation on all 6 domain adaptations tasks from office31 dataset:

python main.py --model_names MMD CORAL --batch_size 32 

• Tuning parameters of DANN model

python main.py --max_epochs 10 --patience 3 --trials_count 1 --model_names DANN --num_workers 2 --batch_size 32 --source amazon --target webcam --hp_tune True 

Check “Available arguments” in “Code Structure” section for all the available arguments.

Code Structure

The main code for project is located in the src/ directory.

• main.py: The entry to program

  • Available arguments:

  • max_epochs: Maximum number of epochs to run the training

  • patience: Maximum number of epochs to continue if no improvement is seen (Early stopping parameter)

  • batch_size

  • num_workers

  • trials_count: Number of trials to run each of the tasks

  • initial_trial: The number to start indexing the trials from

  • download: Whether to download the dataset or not

  • root: Path to the root of project

  • data_root: Path to the data root

  • results_root: Path to the directory to store the results

  • model_names: Names of models to run separated by space - available options: DANN, CDAN, MMD, MCD, CORAL, SOURCE

  • lr: learning rate**

  • gamma: Gamma value for ExponentialLR**

  • hp_tune: Set true of you want to run for different hyperparameters, used for hyperparameter tuning

  • source: The source domain to run the training for, training will run for all the available domains if not specified - available options: amazon, dslr, webcam

  • target: The target domain to run the training for, training will run for all the available domains if not specified - available options: amazon, dslr, webcam

  • vishook_frequency: Number of epochs to wait before save a visualization

  • source_checkpoint_base_dir: Path to source-only trained model directory to use as base, set None to not use source-trained model***

  • source_checkpoint_trial_number: Trail number of source-only trained model to use

• models.py: Contains models for adaptation

• train.py: Contains base training iteration, dataset is also loaded here

• classifier_adapter.py: Contains ClassifierAdapter class which is used for training a source-only model without adaptation

• load_source.py: Load source-only trained model to use as base model for adaptation

• source.py: Contains source model

• train_source.py: Contains base source-only training iteration

• utils.py: Contains utility classes

• vis_hook.py: Contains VizHook class which is used for visualization

** Use can also set different lr and gammas for different models and tasks by changing hp_map in main.py directly.

*** For perfoming domain adaptation on source-trained model, one must should train the model for source using option --model_name SOURCE first

Acknowledgements

Pytorch Adapt