Code implementation for Exploiting Labeled and Structured Data for Fine-Grained Domain Adaptation to Unseen Target Classes.

Train

Scripts to run the training are in PSZS/Scripts. The run.py script assumes working with a (set of) source and target domain(s). The runSource.py script can be used to train a model on only source domain data. For the source only training domain adaptation is not supported i.e. only --method erm works.

Datasets

Currently the following datasets are directly supported:

• CompCars
  • CompCarsModel
  • CompCarsHierarchy
• CUB-Paintings
  • CUBSpecies
  • CUBHierarchy These have to be specified as the --data argument. If a hierarchical dataset is specified but the head-type is non hierarchical the dataset will be automatically changed to the non hierarchical version. Whether a dataset is hierarchical can be specified via the multi_label attribute of the parent class CustomDataset (albeit it currently is hard coded in the run scripts).
    New datasets can be added as inherting from CustomDataset and should come with a dataset descriptor and annotation file.
    For hierarchical datasets the level structure is assumed to be from coarse to fine.

Supported DA

Currently implemented and supported methods that can be specified via the --method parameter are:

• DANN
• ADDA
• JAN
• CCSA
• MCC
• MCD
• MDD
• PAN
• UJDA Additional/Custom specification of feature or logit loss can be given via --model-kwargs as feature_loss_func and logit_loss_func alongside the specified accumulation scheme and other parameters.
  For details see the base class for all models PSZS/Models/CustomModel.

Classifiers

The type of classifier to be used is specified via --classification-type with available types:

• DefaultClassifier
• MaskedClassifier
• SeparatedClassifier The type of classification head used is specified via --head-type and needs to match the used dataset type. If not specified, a SimpleHead without hierarchy is used.
  All options are found under PSZS/Classifiers.

Evaluation

For evaluation the following options are available:

• Create a confusion matrix as an .html using the pycm library
• Create an excel report that tracks which classes are misclassified most commonly and what are the confusions
• Create an excel in addition to the .csv file with the final metrics
• Create a t-SNE visualization of the features
• Make a p-value significance test
• Measure A-Distance (WIP/currently buggy) The first three options are controllable via --create-excel, --create-class-summary and --create-report while the other options have no direct implementation but the code that achieves it is under PSZS/Utils/Eval.