Training Iterations

Training Iterations & Configurations

This document summarizes the different training configurations found in the configs/ directory. The primary models experimented with are DinoV2 and SigLIP.

DinoV2 Experiments

Base Model: facebook/dinov2-with-registers-base-imagenet1k-1-layer

Config File	Training Dataset	Unfrozen Layers	Augmentations	Notes
dino_linear.yaml	_dataset_train.json (3k)	[] (None, linear probe)	Basic (hflip)	Initial baseline with frozen backbone.
dino_linear_aug.yaml	_dataset_train.json (3k)	[] (None, linear probe)	Full list*	Tests impact of heavy augmentation.
dino_hft.yaml	_dataset_train.json (3k)	[11]	Basic (hflip)	Fine-tunes the last transformer block.
dino_hft_aug.yaml	_dataset_train.json (3k)	[11]	Full list*	Fine-tuning with heavy augmentation.
dino_hft_6k.yaml	_dataset_train_5k.json	[11]	Basic (hflip)	Scales up dataset size.
dino_hft_6k_aug.yaml	_dataset_train_5k.json	[11]	Full list*	Scales up dataset with heavy augmentation.
dino_hft_6k_crop.yaml	_dataset_train_5k.json	[11]	Basic (hflip)	Uses return_cropped: true in dataloader.
dino_hft_12k.yaml	_dataset_train_12k.json	[11]	Basic (hflip)	Further scales up dataset size.

* Full list of augmentations typically includes: resize_crop, rotation, hflip, jitter, blur, bw.

SigLIP Experiments

Base Model: google/siglip2-base-patch16-224

Config File	Training Dataset	Unfrozen Layers	Augmentations	Notes
siglip_linear.yaml	_dataset_train.json (3k)	[] (None, linear probe)	Basic	SigLIP baseline with frozen backbone.
siglip_head.yaml	_dataset_train.json (3k)	['head']	Basic	Fine-tunes only the final classification head.
siglip_head_aug.yaml	_dataset_train.json (3k)	['head']	Full list*	Fine-tuning head with heavy augmentation.
siglip_head_6k.yaml	_dataset_train_5k.json	['head']	Basic (hflip)	Scales up dataset size for SigLIP.

Other Models / Evaluation Configs

• aes_anatomy.yaml: Uses the pre-trained incantor/aes-pixai-1.2-anatomy-large-xgb model, likely for zero-shot evaluation, not for training within this repo.
• dino_hft_infer.yaml: A configuration file specifically designed for running inference (c_infer.py) on new, unlabeled datasets to perform auto-labeling.