Parameter Introduction
September 25, 2024 ยท View on GitHub
Data-Related Parameters
--name: Experiment name, used for saving models and parameters undercheckpoints/, facilitating management and tracking of different experimental results.--data_dir: Directory path for training data.--num_worker: Number of worker threads used for data loading, affecting the parallelism and efficiency of data preprocessing.--pad: Amount of padding for input data. Please distinguish this from the--padin Position Shifting.--h, --w: Height and width of the input images.--rr: Random rotation applied to one or more views to enhance data diversity.--ra: Random affine transformation applied to one or more views to enhance data diversity.--re: Random occlusion applied to one or more views to enhance data diversity.--cj: Color jitter applied to one or more views to enhance data diversity.--erasing_p: Probability of random occlusion, controlling the proportion of randomly occluded areas in the images.
Training-Related Parameters
--warm_epoch: Warm-up phase, setting the learning rate to gradually increase over the firstKepochs.--lr: Learning rate.--DA: Whether to use color data augmentation.--droprate: Dropout rate.--autocast: Whether to use mixed precision training.--load_from: Path to the pre-loaded checkpoint for restoring the model from a previous training state.--gpu_ids: Specification of the GPU devices used, supporting multi-GPU configurations for flexible training environments.--batchsize: Number of samples per training step.
Model-Related Parameters
--block: Number of ClassBlocks in the model.--cls_loss: Type of loss function for Representation Learning. Various preset or custom losses can be used, withCELossas the default.--feature_loss: Type of loss function for Metric Learning. Various preset or custom losses can be used, with no loss applied by default.--kl_loss: Type of loss function for Mutual Learning. Various preset or custom losses can be used, with no loss applied by default.--num_bottleneck: Dimensionality of feature embeddings.--backbone: Backbone architecture used. Various preset or custom backbones can be selected, withcvt13as the default.--head: Head architecture used. Various preset or custom heads can be selected, withFSRA_CNNas the default.--head_pool: Type of pooling used in the head, with various preset or custom pooling methods available, defaulting tomax pooling.