Skip to content

Ark+

Vision-only chest-X-ray foundation model — a bare timm.models.swin_transformer.SwinTransformer (Swin-L/192 @ 768 px, window 12) pretrained on six CXR datasets via cyclic multi-task learning.

Embed dim Input size Returns Extra
1536 768×768 (transform, encoder) ark_plus + manual checkpoint + side-loaded timm==0.5.4

Install

uv pip install -e ".[ark_plus]"

Weights are not auto-downloadable. Three manual steps:

  1. Clone the upstream repo (only needed for reference; the recipe does not import from it):
git clone https://github.com/jlianglab/Ark third_party_models/Ark

  1. Request weights at https://forms.gle/qkoDGXNiKRPTDdCe8 (Google Form) or https://www.wjx.cn/vm/OvwfYFx.aspx (WeChat).

  2. Place Ark6_swinLarge768_ep50.pth.tar at third_party_models/Ark/Ark_Plus/Ark6_swinLarge768_ep50.pth.tar.

On first call the recipe side-installs timm==0.5.4 into third_party_models/Ark/timm-054/ via uv pip install --target --no-deps — the 0.5.4 Swin layout doesn't match modern timm (downsample is offset by one layer index in newer versions). The legacy timm is used only during model construction; the rest of the Python session keeps the venv's timm, so EVA-X and other timm-based backbones still work.

Override default paths

Both the checkpoint and the legacy-timm install dir can be redirected — useful when weights live on shared storage, or when you want a fast local cache for the side-install.

Kwarg Default Purpose
checkpoint_path= third_party_models/Ark/Ark_Plus/Ark6_swinLarge768_ep50.pth.tar Full path to the .pth.tar weight file.
legacy_timm_dir= third_party_models/Ark/timm-054/ Where timm==0.5.4 gets side-installed on first call. Point at a fast local cache (e.g. /tmp, /path/to/cache…) to avoid re-installing per user. 
from radharmony.evaluator.backbones import make_ark_plus

transform, encoder = make_ark_plus(
    checkpoint_path="/path/to/models/Ark_Plus/Ark6_swinLarge768_ep50.pth.tar",
    legacy_timm_dir="/path/to/cache/ark-timm-054",
    device="cuda:0",
)

Usage

from radharmony.evaluator.backbones import make_ark_plus
from radharmony.dataset import VinDrCXRTrainDataset

transform, encoder = make_ark_plus(device="cuda:0", output_keys={"img", "cls"})

ds = VinDrCXRTrainDataset(
    base_image_dir="/data/vindr/train/",
    transform=transform,
    cache_dir="/tmp/cache/ark_plus/",
    output_cls=True,
)

Segmentation mode

from radharmony.dataset import SIIMACRPTXTrainDataset

transform, encoder = make_ark_plus(device="cuda:0", output_keys={"img", "mask"})
# forward(x) -> Tensor[B, 1536, 24, 24]   (stride 32 -> 768/32 = 24)

ds = SIIMACRPTXTrainDataset(
    base_image_dir="/data/siim-acr-ptx/dicom-images-train/",
    csv_path="/data/siim-acr-ptx/train-rle.csv",
    transform=transform,
    mask_output_dir="/tmp/cache/siim_ptx_masks/",
    output_mask=True,
    cache_dir="/tmp/cache/ark_plus_seg/",
)

Captured from the final model.norm LayerNorm via a forward hook on the Swin hierarchical features.