from pathlib import Path from types import SimpleNamespace import warnings import numpy as np import torch from torch import nn import torch.nn.functional as F from typing import Optional, List, Callable try: from flash_attn.modules.mha import FlashCrossAttention except ModuleNotFoundError: FlashCrossAttention = None if FlashCrossAttention or hasattr(F, "scaled_dot_product_attention"): FLASH_AVAILABLE = True else: FLASH_AVAILABLE = False torch.backends.cudnn.deterministic = True @torch.cuda.amp.custom_fwd(cast_inputs=torch.float32) def normalize_keypoints(kpts: torch.Tensor, size: torch.Tensor) -> torch.Tensor: if isinstance(size, torch.Size): size = torch.tensor(size)[None] shift = size.float().to(kpts) / 2 scale = size.max(1).values.float().to(kpts) / 2 kpts = (kpts - shift[:, None]) / scale[:, None, None] return kpts def rotate_half(x: torch.Tensor) -> torch.Tensor: x = x.unflatten(-1, (-1, 2)) x1, x2 = x.unbind(dim=-1) return torch.stack((-x2, x1), dim=-1).flatten(start_dim=-2) def apply_cached_rotary_emb(freqs: torch.Tensor, t: torch.Tensor) -> torch.Tensor: return (t * freqs[0]) + (rotate_half(t) * freqs[1]) class LearnableFourierPositionalEncoding(nn.Module): def __init__(self, M: int, dim: int, F_dim: int = None, gamma: float = 1.0) -> None: super().__init__() F_dim = F_dim if F_dim is not None else dim self.gamma = gamma self.Wr = nn.Linear(M, F_dim // 2, bias=False) nn.init.normal_(self.Wr.weight.data, mean=0, std=self.gamma**-2) def forward(self, x: torch.Tensor) -> torch.Tensor: """encode position vector""" projected = self.Wr(x) cosines, sines = torch.cos(projected), torch.sin(projected) emb = torch.stack([cosines, sines], 0).unsqueeze(-3) return emb.repeat_interleave(2, dim=-1) class TokenConfidence(nn.Module): def __init__(self, dim: int) -> None: super().__init__() self.token = nn.Sequential(nn.Linear(dim, 1), nn.Sigmoid()) def forward(self, desc0: torch.Tensor, desc1: torch.Tensor): """get confidence tokens""" return ( self.token(desc0.detach().float()).squeeze(-1), self.token(desc1.detach().float()).squeeze(-1), ) class Attention(nn.Module): def __init__(self, allow_flash: bool) -> None: super().__init__() if allow_flash and not FLASH_AVAILABLE: warnings.warn( "FlashAttention is not available. For optimal speed, " "consider installing torch >= 2.0 or flash-attn.", stacklevel=2, ) self.enable_flash = allow_flash and FLASH_AVAILABLE if allow_flash and FlashCrossAttention: self.flash_ = FlashCrossAttention() def forward(self, q, k, v) -> torch.Tensor: if self.enable_flash and q.device.type == "cuda": if FlashCrossAttention: q, k, v = [x.transpose(-2, -3) for x in [q, k, v]] m = self.flash_(q.half(), torch.stack([k, v], 2).half()) return m.transpose(-2, -3).to(q.dtype) else: # use torch 2.0 scaled_dot_product_attention with flash args = [x.half().contiguous() for x in [q, k, v]] with torch.backends.cuda.sdp_kernel(enable_flash=True): return F.scaled_dot_product_attention(*args).to(q.dtype) elif hasattr(F, "scaled_dot_product_attention"): args = [x.contiguous() for x in [q, k, v]] return F.scaled_dot_product_attention(*args).to(q.dtype) else: s = q.shape[-1] ** -0.5 attn = F.softmax(torch.einsum("...id,...jd->...ij", q, k) * s, -1) return torch.einsum("...ij,...jd->...id", attn, v) class Transformer(nn.Module): def __init__( self, embed_dim: int, num_heads: int, flash: bool = False, bias: bool = True ) -> None: super().__init__() self.embed_dim = embed_dim self.num_heads = num_heads assert self.embed_dim % num_heads == 0 self.head_dim = self.embed_dim // num_heads self.Wqkv = nn.Linear(embed_dim, 3 * embed_dim, bias=bias) self.inner_attn = Attention(flash) self.out_proj = nn.Linear(embed_dim, embed_dim, bias=bias) self.ffn = nn.Sequential( nn.Linear(2 * embed_dim, 2 * embed_dim), nn.LayerNorm(2 * embed_dim, elementwise_affine=True), nn.GELU(), nn.Linear(2 * embed_dim, embed_dim), ) def _forward(self, x: torch.Tensor, encoding: Optional[torch.Tensor] = None): qkv = self.Wqkv(x) qkv = qkv.unflatten(-1, (self.num_heads, -1, 3)).transpose(1, 2) q, k, v = qkv[..., 0], qkv[..., 1], qkv[..., 2] if encoding is not None: q = apply_cached_rotary_emb(encoding, q) k = apply_cached_rotary_emb(encoding, k) context = self.inner_attn(q, k, v) message = self.out_proj(context.transpose(1, 2).flatten(start_dim=-2)) return x + self.ffn(torch.cat([x, message], -1)) def forward(self, x0, x1, encoding0=None, encoding1=None): return self._forward(x0, encoding0), self._forward(x1, encoding1) class CrossTransformer(nn.Module): def __init__( self, embed_dim: int, num_heads: int, flash: bool = False, bias: bool = True ) -> None: super().__init__() self.heads = num_heads dim_head = embed_dim // num_heads self.scale = dim_head**-0.5 inner_dim = dim_head * num_heads self.to_qk = nn.Linear(embed_dim, inner_dim, bias=bias) self.to_v = nn.Linear(embed_dim, inner_dim, bias=bias) self.to_out = nn.Linear(inner_dim, embed_dim, bias=bias) self.ffn = nn.Sequential( nn.Linear(2 * embed_dim, 2 * embed_dim), nn.LayerNorm(2 * embed_dim, elementwise_affine=True), nn.GELU(), nn.Linear(2 * embed_dim, embed_dim), ) if flash and FLASH_AVAILABLE: self.flash = Attention(True) else: self.flash = None def map_(self, func: Callable, x0: torch.Tensor, x1: torch.Tensor): return func(x0), func(x1) def forward(self, x0: torch.Tensor, x1: torch.Tensor) -> List[torch.Tensor]: qk0, qk1 = self.map_(self.to_qk, x0, x1) v0, v1 = self.map_(self.to_v, x0, x1) qk0, qk1, v0, v1 = map( lambda t: t.unflatten(-1, (self.heads, -1)).transpose(1, 2), (qk0, qk1, v0, v1), ) if self.flash is not None: m0 = self.flash(qk0, qk1, v1) m1 = self.flash(qk1, qk0, v0) else: qk0, qk1 = qk0 * self.scale**0.5, qk1 * self.scale**0.5 sim = torch.einsum("b h i d, b h j d -> b h i j", qk0, qk1) attn01 = F.softmax(sim, dim=-1) attn10 = F.softmax(sim.transpose(-2, -1).contiguous(), dim=-1) m0 = torch.einsum("bhij, bhjd -> bhid", attn01, v1) m1 = torch.einsum("bhji, bhjd -> bhid", attn10.transpose(-2, -1), v0) m0, m1 = self.map_(lambda t: t.transpose(1, 2).flatten(start_dim=-2), m0, m1) m0, m1 = self.map_(self.to_out, m0, m1) x0 = x0 + self.ffn(torch.cat([x0, m0], -1)) x1 = x1 + self.ffn(torch.cat([x1, m1], -1)) return x0, x1 def sigmoid_log_double_softmax( sim: torch.Tensor, z0: torch.Tensor, z1: torch.Tensor ) -> torch.Tensor: """create the log assignment matrix from logits and similarity""" b, m, n = sim.shape certainties = F.logsigmoid(z0) + F.logsigmoid(z1).transpose(1, 2) scores0 = F.log_softmax(sim, 2) scores1 = F.log_softmax(sim.transpose(-1, -2).contiguous(), 2).transpose(-1, -2) scores = sim.new_full((b, m + 1, n + 1), 0) scores[:, :m, :n] = scores0 + scores1 + certainties scores[:, :-1, -1] = F.logsigmoid(-z0.squeeze(-1)) scores[:, -1, :-1] = F.logsigmoid(-z1.squeeze(-1)) return scores class MatchAssignment(nn.Module): def __init__(self, dim: int) -> None: super().__init__() self.dim = dim self.matchability = nn.Linear(dim, 1, bias=True) self.final_proj = nn.Linear(dim, dim, bias=True) def forward(self, desc0: torch.Tensor, desc1: torch.Tensor): """build assignment matrix from descriptors""" mdesc0, mdesc1 = self.final_proj(desc0), self.final_proj(desc1) _, _, d = mdesc0.shape mdesc0, mdesc1 = mdesc0 / d**0.25, mdesc1 / d**0.25 sim = torch.einsum("bmd,bnd->bmn", mdesc0, mdesc1) z0 = self.matchability(desc0) z1 = self.matchability(desc1) scores = sigmoid_log_double_softmax(sim, z0, z1) return scores, sim def scores(self, desc0: torch.Tensor, desc1: torch.Tensor): m0 = torch.sigmoid(self.matchability(desc0)).squeeze(-1) m1 = torch.sigmoid(self.matchability(desc1)).squeeze(-1) return m0, m1 def filter_matches(scores: torch.Tensor, th: float): """obtain matches from a log assignment matrix [Bx M+1 x N+1]""" max0, max1 = scores[:, :-1, :-1].max(2), scores[:, :-1, :-1].max(1) m0, m1 = max0.indices, max1.indices mutual0 = torch.arange(m0.shape[1]).to(m0)[None] == m1.gather(1, m0) mutual1 = torch.arange(m1.shape[1]).to(m1)[None] == m0.gather(1, m1) max0_exp = max0.values.exp() zero = max0_exp.new_tensor(0) mscores0 = torch.where(mutual0, max0_exp, zero) mscores1 = torch.where(mutual1, mscores0.gather(1, m1), zero) if th is not None: valid0 = mutual0 & (mscores0 > th) else: valid0 = mutual0 valid1 = mutual1 & valid0.gather(1, m1) m0 = torch.where(valid0, m0, m0.new_tensor(-1)) m1 = torch.where(valid1, m1, m1.new_tensor(-1)) return m0, m1, mscores0, mscores1 class LightGlue(nn.Module): default_conf = { "name": "lightglue", # just for interfacing "input_dim": 256, # input descriptor dimension (autoselected from weights) "descriptor_dim": 256, "n_layers": 9, "num_heads": 4, "flash": True, # enable FlashAttention if available. "mp": False, # enable mixed precision "depth_confidence": 0.95, # early stopping, disable with -1 "width_confidence": 0.99, # point pruning, disable with -1 "filter_threshold": 0.1, # match threshold "weights": None, } required_data_keys = ["image0", "image1"] version = "v0.1_arxiv" url = "https://github.com/cvg/LightGlue/releases/download/{}/{}_lightglue.pth" features = { "superpoint": ("superpoint_lightglue", 256), "disk": ("disk_lightglue", 128), } def __init__(self, features="superpoint", **conf) -> None: super().__init__() self.conf = {**self.default_conf, **conf} if features is not None: assert features in list(self.features.keys()) self.conf["weights"], self.conf["input_dim"] = self.features[features] self.conf = conf = SimpleNamespace(**self.conf) if conf.input_dim != conf.descriptor_dim: self.input_proj = nn.Linear(conf.input_dim, conf.descriptor_dim, bias=True) else: self.input_proj = nn.Identity() head_dim = conf.descriptor_dim // conf.num_heads self.posenc = LearnableFourierPositionalEncoding(2, head_dim, head_dim) h, n, d = conf.num_heads, conf.n_layers, conf.descriptor_dim self.self_attn = nn.ModuleList( [Transformer(d, h, conf.flash) for _ in range(n)] ) self.cross_attn = nn.ModuleList( [CrossTransformer(d, h, conf.flash) for _ in range(n)] ) self.log_assignment = nn.ModuleList([MatchAssignment(d) for _ in range(n)]) self.token_confidence = nn.ModuleList( [TokenConfidence(d) for _ in range(n - 1)] ) if features is not None: fname = f"{conf.weights}_{self.version}.pth".replace(".", "-") state_dict = torch.hub.load_state_dict_from_url( self.url.format(self.version, features), file_name=fname ) self.load_state_dict(state_dict, strict=False) elif conf.weights is not None: path = Path(__file__).parent path = path / "weights/{}.pth".format(self.conf.weights) state_dict = torch.load(str(path), map_location="cpu") self.load_state_dict(state_dict, strict=False) print("Loaded LightGlue model") def forward(self, data: dict) -> dict: """ Match keypoints and descriptors between two images Input (dict): image0: dict keypoints: [B x M x 2] descriptors: [B x M x D] image: [B x C x H x W] or image_size: [B x 2] image1: dict keypoints: [B x N x 2] descriptors: [B x N x D] image: [B x C x H x W] or image_size: [B x 2] Output (dict): log_assignment: [B x M+1 x N+1] matches0: [B x M] matching_scores0: [B x M] matches1: [B x N] matching_scores1: [B x N] matches: List[[Si x 2]], scores: List[[Si]] """ with torch.autocast(enabled=self.conf.mp, device_type="cuda"): return self._forward(data) def _forward(self, data: dict) -> dict: for key in self.required_data_keys: assert key in data, f"Missing key {key} in data" data0, data1 = data["image0"], data["image1"] kpts0_, kpts1_ = data0["keypoints"], data1["keypoints"] b, m, _ = kpts0_.shape b, n, _ = kpts1_.shape size0, size1 = data0.get("image_size"), data1.get("image_size") size0 = size0 if size0 is not None else data0["image"].shape[-2:][::-1] size1 = size1 if size1 is not None else data1["image"].shape[-2:][::-1] kpts0 = normalize_keypoints(kpts0_, size=size0) kpts1 = normalize_keypoints(kpts1_, size=size1) assert torch.all(kpts0 >= -1) and torch.all(kpts0 <= 1) assert torch.all(kpts1 >= -1) and torch.all(kpts1 <= 1) desc0 = data0["descriptors"].detach() desc1 = data1["descriptors"].detach() assert desc0.shape[-1] == self.conf.input_dim assert desc1.shape[-1] == self.conf.input_dim if torch.is_autocast_enabled(): desc0 = desc0.half() desc1 = desc1.half() desc0 = self.input_proj(desc0) desc1 = self.input_proj(desc1) # cache positional embeddings encoding0 = self.posenc(kpts0) encoding1 = self.posenc(kpts1) # GNN + final_proj + assignment ind0 = torch.arange(0, m).to(device=kpts0.device)[None] ind1 = torch.arange(0, n).to(device=kpts0.device)[None] prune0 = torch.ones_like(ind0) # store layer where pruning is detected prune1 = torch.ones_like(ind1) dec, wic = self.conf.depth_confidence, self.conf.width_confidence token0, token1 = None, None for i in range(self.conf.n_layers): # self+cross attention desc0, desc1 = self.self_attn[i](desc0, desc1, encoding0, encoding1) desc0, desc1 = self.cross_attn[i](desc0, desc1) if i == self.conf.n_layers - 1: continue # no early stopping or adaptive width at last layer if dec > 0: # early stopping token0, token1 = self.token_confidence[i](desc0, desc1) if self.stop(token0, token1, self.conf_th(i), dec, m + n): break if wic > 0: # point pruning match0, match1 = self.log_assignment[i].scores(desc0, desc1) mask0 = self.get_mask(token0, match0, self.conf_th(i), 1 - wic) mask1 = self.get_mask(token1, match1, self.conf_th(i), 1 - wic) ind0, ind1 = ind0[mask0][None], ind1[mask1][None] desc0, desc1 = desc0[mask0][None], desc1[mask1][None] if desc0.shape[-2] == 0 or desc1.shape[-2] == 0: break encoding0 = encoding0[:, :, mask0][:, None] encoding1 = encoding1[:, :, mask1][:, None] prune0[:, ind0] += 1 prune1[:, ind1] += 1 if wic > 0: # scatter with indices after pruning scores_, _ = self.log_assignment[i](desc0, desc1) dt, dev = scores_.dtype, scores_.device scores = torch.zeros(b, m + 1, n + 1, dtype=dt, device=dev) scores[:, :-1, :-1] = -torch.inf scores[:, ind0[0], -1] = scores_[:, :-1, -1] scores[:, -1, ind1[0]] = scores_[:, -1, :-1] x, y = torch.meshgrid(ind0[0], ind1[0], indexing="ij") scores[:, x, y] = scores_[:, :-1, :-1] else: scores, _ = self.log_assignment[i](desc0, desc1) m0, m1, mscores0, mscores1 = filter_matches(scores, self.conf.filter_threshold) matches, mscores = [], [] for k in range(b): valid = m0[k] > -1 matches.append(torch.stack([torch.where(valid)[0], m0[k][valid]], -1)) mscores.append(mscores0[k][valid]) return { "log_assignment": scores, "matches0": m0, "matches1": m1, "matching_scores0": mscores0, "matching_scores1": mscores1, "stop": i + 1, "prune0": prune0, "prune1": prune1, "matches": matches, "scores": mscores, } def conf_th(self, i: int) -> float: """scaled confidence threshold""" return np.clip(0.8 + 0.1 * np.exp(-4.0 * i / self.conf.n_layers), 0, 1) def get_mask( self, confidence: torch.Tensor, match: torch.Tensor, conf_th: float, match_th: float, ) -> torch.Tensor: """mask points which should be removed""" if conf_th and confidence is not None: mask = ( torch.where(confidence > conf_th, match, match.new_tensor(1.0)) > match_th ) else: mask = match > match_th return mask def stop( self, token0: torch.Tensor, token1: torch.Tensor, conf_th: float, inl_th: float, seql: int, ) -> torch.Tensor: """evaluate stopping condition""" tokens = torch.cat([token0, token1], -1) if conf_th: pos = 1.0 - (tokens < conf_th).float().sum() / seql return pos > inl_th else: return tokens.mean() > inl_th