import torch
from huggingface_hub import snapshot_download
from safetensors.torch import load_file
from copy import deepcopy
import torchvision.transforms as tfm
from vismatch import THIRD_PARTY_DIR, BaseMatcher
from vismatch.utils import resize_to_divisible, add_to_path
add_to_path(THIRD_PARTY_DIR.joinpath("EfficientLoFTR"))
from src.loftr import LoFTR, full_default_cfg, opt_default_cfg, reparameter
[docs]
class EfficientLoFTRMatcher(BaseMatcher):
divisible_size = 32
def __init__(self, device="cpu", cfg="full", **kwargs):
super().__init__(device, **kwargs)
self.precision = kwargs.get("precision", self.get_precision())
model_path = f"{snapshot_download('vismatch/eloftr')}/eloftr_outdoors.safetensors"
self.matcher = LoFTR(config=deepcopy(full_default_cfg if cfg == "full" else opt_default_cfg))
state_dict = load_file(model_path)
self.matcher.load_state_dict(state_dict)
self.matcher = reparameter(self.matcher).to(self.device).eval()
[docs]
def get_precision(self):
return "fp16"
[docs]
def preprocess(self, img):
_, h, w = img.shape
orig_shape = h, w
img = resize_to_divisible(img, self.divisible_size)
return tfm.Grayscale()(img).unsqueeze(0), orig_shape
def _forward(self, img0, img1):
img0, img0_orig_shape = self.preprocess(img0)
img1, img1_orig_shape = self.preprocess(img1)
batch = {"image0": img0, "image1": img1}
if self.precision == "mp" and self.device == "cuda":
with torch.autocast(enabled=True, device_type="cuda"):
self.matcher(batch)
else:
self.matcher(batch)
mkpts0 = batch["mkpts0_f"]
mkpts1 = batch["mkpts1_f"]
H0, W0, H1, W1 = *img0.shape[-2:], *img1.shape[-2:]
mkpts0 = self.rescale_coords(mkpts0, *img0_orig_shape, H0, W0)
mkpts1 = self.rescale_coords(mkpts1, *img1_orig_shape, H1, W1)
return mkpts0, mkpts1, None, None, None, None