configuration_options.md

Here are the key parameters grouped by their purpose:

General Settings

• --exp_name: Experiment name for logging purposes
• --seed: Random seed for reproducibility (default: 42)
• --device: GPU device(s) to use (default: [0])
• --num_workers: Number of data loading workers (default: 1)

Data Configuration

• --dataroot: Root directory for datasets (default: "/data/owcl_data")
• --datasets: Comma-separated list of datasets to use (default: "CIFAR100,SUN397,EuroSAT,OxfordIIITPet,Flowers102,FGVCAircraft,StanfordCars,Food101")
• --held_out_dataset: Dataset to use for held-out evaluation (default: "ImageNet,UCF101,DTD")
• --input_size: Input image size (default: 224)

Model Configuration

• --network_arc: Network architecture to use (default: "clip")
• --backbone: Backbone network (e.g., 'ViT-B/32' for CLIP, 'vitb14' for DINO)

Continual Learning Settings

• --incremental: Incremental learning scenario (default: "dataset", choices: ["dataset", "class", "task"]), dataset represents task incremental learning
• --num_classes: Number of classes per stage for class-incremental learning (default: 100)
• --randomize: Randomize class order (default: True)

Training Parameters

• --optimizer: Optimizer to use (default: "adamw")
• --batch_size: Batch size for training (default: 2048)
• --lr: Learning rate (default: 6e-4)
• --momentum: Momentum for optimizer (default: 0.9)
• --weight_decay: Weight decay for optimizer (default: 0.05)
• --n_epochs: Number of training epochs (default: 20)
• --criteria: Loss criteria to use (default: "osce", choices: ["cs", "osce", "osce_other"])
• --X_format: Input data format (default: "feature", choices: ["image", "feature", "embedding", "code"])

AnytimeCL Specific Options

• --learning_strategy: Learning strategy (default: "online", choices: ["online", "offline", "wake_sleep", "none"])
• --wake_bs: Batch size for wake training (default: 32)
• --wake_evaluation_iter_ratio: Ratio of iterations for wake evaluation (default: 0.25)
• --sampler_type: Type of sampler to use (default: "class_balanced", choices: ["weighted", "none", "fifo", "class_balanced", "uniform"])

Compression Options

• --need_compress: Enable feature compression (default: False)
• --CLS_weight: Use CLS token weight for compression (default: False)
• --per_instance: Perform per-instance compression (default: True)
• --int_quantize: Enable integer quantization for compression (default: False)
• --components: Number of components for compression (default: 5)
• --int_range: Integer range for quantization (default: 255)

Evaluation and Logging

• --results_dir: Directory to save results (default: "./results")
• --log_interval: Interval for logging during training (default: 10)
• --save_interval: Interval for saving model checkpoints (default: 5)
• --eval_interval: Interval for evaluation during training (default: 5)
• --eval_scenario: Evaluation scenario (default: "cumulative_cumulative")

Miscellaneous

• --include_the_other_class: Include "other" class in classification (action: store_true)
• --use_other_classifier: Use a separate classifier for the "other" class (action: store_true)
• --use_tuned_text_embedding: Use tuned text embedding (action: store_true)
• --accumulating_data_to_the_final_stage: Accumulate data to the final stage (action: store_true)
• --ema_exemplar_per_class_acc: Use EMA for exemplar per-class accuracy (action: store_true)
• --ema_exemplar_per_class_acc_decay: Decay rate for EMA exemplar per-class accuracy (default: 0.9)
• --fix_finetuned_model: Fix the fine-tuned model (action: store_true)

For a complete list of options and their descriptions, please refer to the options/ directory in the source code and the modify_commandline_options method in each module.