Modelpublic

HajihajihaJimmy/DeepSparse

sourceHugging Faceupdated 5mo agoView on Hugging Face

0likes

Model Card

Checkpoints

This directory holds all pretrained and finetuned model weights for DeepSparse, plus their training configs.

Note: This directory is excluded from the git repository. Checkpoints are hosted on HuggingFace: [HajihajihaJimmy/DeepSparse](https://huggingface.co/HajihajihaJimmy/DeepSparse)

Folder Layout

Each checkpoint folder contains two files:

logs/
  pretrain/
    ep_700.pth          Pretrained foundation model (AtlasAbdomenMini, epoch 700)
    config.yaml         Model + dataset config used at inference
  {dataset}+{n}v+s2/
    ep_400.pth          Finetuned model — dataset, n input views, Stage 2
    config.yaml         Model + dataset config used at inference

evaluate.py automatically reads ./logs/{name}/config.yaml when --cfg_path is not provided.

Checkpoint Map

Folder	Anatomy	Views	Original training run
`pretrain/`	— (pretraining)	6–24	`atlas_7k+mv+rv`
`abdomen+6v+s2/`	Abdomen (PANORAMA)	6	`atlas_7k+mv+rv+s1+ep7+abdomen+6v+s2`
`abdomen+8v+s2/`	Abdomen (PANORAMA)	8	`atlas_7k+mv+rv+s1+ep7+abdomen+8v+s2`
`abdomen+10v+s2/`	Abdomen (PANORAMA)	10	`atlas_7k+mv+rv+s1+ep7+abdomen+10v+s2`
`pelvis+6v+s2/`	Pelvis (PENGWIN)	6	`atlas_7k+mv+rv+s1+ep7+pelvis+6v+s2`
`pelvis+8v+s2/`	Pelvis (PENGWIN)	8	`atlas_7k+mv+rv+s1+ep7+pelvis+8v+s2`
`pelvis+10v+s2/`	Pelvis (PENGWIN)	10	`atlas_7k+mv+rv+s1+ep7+pelvis+10v+s2` (best at training ep 300)
`tooth+6v+s2/`	Tooth (ToothFairy)	6	`atlas_7k+mv+rv+s1+ep7+tooth+6v+s2`
`tooth+8v+s2/`	Tooth (ToothFairy)	8	`atlas_7k+mv+rv+s1+ep7+tooth+8v+s2`
`tooth+10v+s2/`	Tooth (ToothFairy)	10	`atlas_7k+mv+rv+s1+ep7+tooth+10v+s2`
`luna+6v+s2/`	Lung (LUNA16)	6	`atlas_7k+mv+rv+ft_s1+ep7+s2`
`luna+8v+s2/`	Lung (LUNA16)	8	`atlas_7k+mv+rv+s1+ep7+8v+s2`
`luna+10v+s2/`	Lung (LUNA16)	10	`atlas_7k+mv+rv+s1+ep7+10v+s2`

Usage

All evaluation scripts expect checkpoints at logs/{name}/ep_{epoch}.pth. The default epoch for inference is 400 (set via --epoch 400 in evaluate_s2.sh).