add some change twice

losses.py

@@ -0,0 +1,138 @@
+"""Loss utilities for RoRD training."""
+from __future__ import annotations
+
+import math
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+
+def _augment_homography_matrix(h_2x3: torch.Tensor) -> torch.Tensor:
+    """Append the third row [0, 0, 1] to build a full 3x3 homography."""
+    if h_2x3.dim() != 3 or h_2x3.size(1) != 2 or h_2x3.size(2) != 3:
+        raise ValueError("Expected homography with shape (B, 2, 3)")
+
+    batch_size = h_2x3.size(0)
+    device = h_2x3.device
+    bottom_row = torch.tensor([0.0, 0.0, 1.0], device=device, dtype=h_2x3.dtype)
+    bottom_row = bottom_row.view(1, 1, 3).expand(batch_size, -1, -1)
+    return torch.cat([h_2x3, bottom_row], dim=1)
+
+
+def warp_feature_map(feature_map: torch.Tensor, h_inv: torch.Tensor) -> torch.Tensor:
+    """Warp feature map according to inverse homography."""
+    return F.grid_sample(
+        feature_map,
+        F.affine_grid(h_inv, feature_map.size(), align_corners=False),
+        align_corners=False,
+    )
+
+
+def compute_detection_loss(
+    det_original: torch.Tensor,
+    det_rotated: torch.Tensor,
+    h: torch.Tensor,
+) -> torch.Tensor:
+    """Binary cross-entropy + smooth L1 detection loss."""
+    h_full = _augment_homography_matrix(h)
+    h_inv = torch.inverse(h_full)[:, :2, :]
+    warped_det = warp_feature_map(det_rotated, h_inv)
+
+    bce_loss = F.binary_cross_entropy(det_original, warped_det)
+    smooth_l1_loss = F.smooth_l1_loss(det_original, warped_det)
+    return bce_loss + 0.1 * smooth_l1_loss
+
+
+def compute_description_loss(
+    desc_original: torch.Tensor,
+    desc_rotated: torch.Tensor,
+    h: torch.Tensor,
+    margin: float = 1.0,
+) -> torch.Tensor:
+    """Triplet-style descriptor loss with Manhattan-aware sampling."""
+    batch_size, channels, height, width = desc_original.size()
+    num_samples = 200
+
+    grid_side = int(math.sqrt(num_samples))
+    h_coords = torch.linspace(-1, 1, grid_side, device=desc_original.device)
+    w_coords = torch.linspace(-1, 1, grid_side, device=desc_original.device)
+
+    manhattan_h = torch.cat([h_coords, torch.zeros_like(h_coords)])
+    manhattan_w = torch.cat([torch.zeros_like(w_coords), w_coords])
+    manhattan_coords = torch.stack([manhattan_h, manhattan_w], dim=1)
+    manhattan_coords = manhattan_coords.unsqueeze(0).repeat(batch_size, 1, 1)
+
+    anchor = F.grid_sample(
+        desc_original,
+        manhattan_coords.unsqueeze(1),
+        align_corners=False,
+    ).squeeze(2).transpose(1, 2)
+
+    coords_hom = torch.cat(
+        [manhattan_coords, torch.ones(batch_size, manhattan_coords.size(1), 1, device=desc_original.device)],
+        dim=2,
+    )
+
+    h_full = _augment_homography_matrix(h)
+    h_inv = torch.inverse(h_full)
+    coords_transformed = (coords_hom @ h_inv.transpose(1, 2))[:, :, :2]
+
+    positive = F.grid_sample(
+        desc_rotated,
+        coords_transformed.unsqueeze(1),
+        align_corners=False,
+    ).squeeze(2).transpose(1, 2)
+
+    negative_list = []
+    if manhattan_coords.size(1) > 0:
+        angles = [0, 90, 180, 270]
+        for angle in angles:
+            if angle == 0:
+                continue
+            theta = torch.tensor(angle * math.pi / 180.0, device=desc_original.device)
+            cos_t = torch.cos(theta)
+            sin_t = torch.sin(theta)
+            rot = torch.stack(
+                [
+                    torch.stack([cos_t, -sin_t]),
+                    torch.stack([sin_t, cos_t]),
+                ]
+            )
+            rotated_coords = manhattan_coords @ rot.T
+            negative_list.append(rotated_coords)
+
+    if negative_list:
+        neg_coords = torch.stack(negative_list, dim=1).reshape(batch_size, -1, 2)
+        negative_candidates = F.grid_sample(
+            desc_rotated,
+            neg_coords.unsqueeze(1),
+            align_corners=False,
+        ).squeeze(2).transpose(1, 2)
+
+        anchor_expanded = anchor.unsqueeze(2).expand(-1, -1, negative_candidates.size(1), -1)
+        negative_expanded = negative_candidates.unsqueeze(1).expand(-1, anchor.size(1), -1, -1)
+        manhattan_dist = torch.sum(torch.abs(anchor_expanded - negative_expanded), dim=3)
+
+        k = max(anchor.size(1) // 2, 1)
+        hard_indices = torch.topk(manhattan_dist, k=k, largest=False)[1]
+        idx_expand = hard_indices.unsqueeze(-1).expand(-1, -1, -1, negative_candidates.size(2))
+        negative = torch.gather(negative_candidates.unsqueeze(1).expand(-1, anchor.size(1), -1, -1), 2, idx_expand)
+        negative = negative.mean(dim=2)
+    else:
+        negative = torch.zeros_like(anchor)
+
+    triplet_loss = nn.TripletMarginLoss(margin=margin, p=1, reduction='mean')
+    geometric_triplet = triplet_loss(anchor, positive, negative)
+
+    manhattan_loss = 0.0
+    for i in range(anchor.size(1)):
+        anchor_norm = F.normalize(anchor[:, i], p=2, dim=1)
+        positive_norm = F.normalize(positive[:, i], p=2, dim=1)
+        cos_sim = torch.sum(anchor_norm * positive_norm, dim=1)
+        manhattan_loss += torch.mean(1 - cos_sim)
+
+    manhattan_loss = manhattan_loss / max(anchor.size(1), 1)
+    sparsity_loss = torch.mean(torch.abs(anchor)) + torch.mean(torch.abs(positive))
+    binary_loss = torch.mean(torch.abs(torch.sign(anchor) - torch.sign(positive)))
+
+    return geometric_triplet + 0.1 * manhattan_loss + 0.01 * sparsity_loss + 0.05 * binary_loss