README-config.md

Model configurations

This README provides guidelines on how to configure PromptCCD hyperparameters. The *.yaml can be accesed at config/%DATASET%/.

Important configs

• run_ccd: set to true when running the model for training and testing.
• ccd_model: indicate which CCD model to choose.
• manual_seed: seed for training.
• save_path: indicate where to store the experiment results.
• eval_version: evaluation metric to evaluate the model, e.g., ccd or gcd.
• transductive_evaluation: please refer to supplementary material sec. S3.

Data

• dataset: dataset to choose for experiment.
• ccd_split_ratio: how the dataset is splitted into n_stage(s)
• random_split_ratio: split ratio between train and val datasets.

Optimization

• epoch: number of training for each stage.
• optim: optimizer algorithm.
• base_lr: initial learning rate.
• lr_scheduler: learning rate scheduler.
• temperature: temperature parameter for info ncd logits loss function.
• sup_con_weight: weight for the supervised contrastive loss function for each stage. (default: [0.35, 0., 0., 0.])

Model & Prompt module

• use_dinov2: if set true, then DINOv2 model is used for training. (default: DINO)
• grad_from_block: DINO's transformer starting block to optimize.
• prompt_pool: set to true for prompt learning.
• top_k: top k mean components for prompting .
• fit_gmm_every_n_epoch: GMM optimiztion schedule.
• num_gmm_samples: number of gmm samples per class to be stored after training.
• covariance_type: GMM covariance types.