EVA-X¶

ViT pretrained with masked image modeling on a merged 520k-CXR corpus (hustvl/EVA-X). Model code lives in the vendored submodule at third_party_models/EVA-X/; weights are pulled from the MapleF/eva_x HuggingFace repo on first use.

Variant	Embed dim	Params
`tiny`	192	~6M
`small`	384	~22M
`base` (default)	768	~86M

Input size	Returns	Extra
224×224 (CXR norm, not ImageNet)	`(transform, encoder)`	`eva_x` + git submodule

Preprocessing: resize 256, center-crop 224 (BICUBIC), normalize with EVA-X's own CXR-pretraining stats (mean=0.49185243 / std=0.28509309 per channel) — the canonical timm-style crop_pct=224/256 eval pipeline.

Install¶

uv pip install -e ".[eva_x]"
git submodule update --init third_party_models/EVA-X

Weights are auto-downloaded from HuggingFace on first call.

Override default paths¶

Kwarg	Default	Purpose
`repo_path=`	`third_party_models/EVA-X/`	Path to the cloned EVA-X submodule. The recipe adds this to `sys.path` to import the model code.

from radharmony.evaluator.backbones import make_eva_x

transform, encoder = make_eva_x(
    variant="base",
    repo_path="/shared/repos/EVA-X",
    device="cuda:0",
)

Usage¶

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

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

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

Segmentation mode¶

from radharmony.dataset import SIIMACRPTXTrainDataset

transform, encoder = make_eva_x(device="cuda:0", output_keys={"img", "mask"})
# forward(x) -> Tensor[B, D, 14, 14]   (224 / 16 = 14)

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/eva_x_seg/",
)