add CI for formatting (#84)

* update.

* update version.

* update clang version.

* update.

* update.

* update.

* update.

* align with 21.12 for black.

* update.

.github/workflows/format-ubuntu.yml

@@ -0,0 +1,44 @@
+name: Check code formatting (Ubuntu)
+
+on:
+  push:
+    branches:
+      - main
+  pull_request:
+    types: [ assigned, opened, synchronize, reopened ]
+  release:
+    types: [ published, edited ]
+
+jobs:
+  build:
+    name: ${{ matrix.config.os }} formatting 
+    runs-on: ${{ matrix.config.os }}
+    strategy:
+      matrix:
+        config: [
+          {
+            os: ubuntu-22.04,
+            checkCodeFormat: true,
+          },
+        ]
+
+    env:
+      COMPILER_CACHE_VERSION: 1
+      COMPILER_CACHE_DIR: ${{ github.workspace }}/compiler-cache
+      CCACHE_DIR: ${{ github.workspace }}/compiler-cache/ccache
+      CCACHE_BASEDIR: ${{ github.workspace }}
+      CTCACHE_DIR: ${{ github.workspace }}/compiler-cache/ctcache
+
+    steps:
+      - uses: actions/checkout@v4
+      - name: Check code format
+        run: |
+          if [ "${{ matrix.config.checkCodeFormat }}" != "true" ]; then
+            exit 0
+          fi
+          set +x -euo pipefail
+          sudo apt-get update && sudo apt-get install -y clang-format-14 black
+          ./scripts/format/clang_format.sh
+          ./scripts/format/black.sh
+          git diff --name-only
+          git diff --exit-code || (echo "Code formatting failed" && exit 1)

limap/base/unit_test.py

@@ -34,10 +34,10 @@ def report_error(imagecols_pred, imagecols):
         R_error = (
             imagecols_pred.camimage(img_id).R() - imagecols.camimage(img_id).R()
         )
-        R_error = np.sqrt(np.sum(R_error**2))
+        R_error = np.sqrt(np.sum(R_error ** 2))
         T_error = (
             imagecols_pred.camimage(img_id).T() - imagecols.camimage(img_id).T()
         )
-        T_error = np.sqrt(np.sum(T_error**2))
+        T_error = np.sqrt(np.sum(T_error ** 2))
         pose_errors.append(np.array([R_error, T_error]))
     print("pose_error: (R, T)", np.array(pose_errors).mean(0))

limap/ceresbase/interpolation.h

@@ -87,9 +87,9 @@ template <typename dtype> struct InterpolationQuery {
                      const Eigen::Vector2d &scale_,
                      const std::array<int, 3> &shape_, const dtype *data_ptr_)
       : scale(scale_), shape(shape_), corner(corner_), data_ptr(data_ptr_) {}
-  Eigen::Vector2d scale; // Could be a const-reference? --> With little change
-                         // to pybind this can be done
-  Eigen::Vector2i corner;   // Could be a const-reference?
+  Eigen::Vector2d scale;  // Could be a const-reference? --> With little change
+                          // to pybind this can be done
+  Eigen::Vector2i corner; // Could be a const-reference?
   std::array<int, 3> shape; // Could be a const-reference?
   const dtype *data_ptr;
 

limap/estimators/absolute_pose/_pl_estimate_absolute_pose.py

@@ -110,12 +110,15 @@ def _pl_estimate_absolute_pose(
         ransac_options.data_type_weights_ = np.array(
             [ransac_cfg["weight_point"], ransac_cfg["weight_line"]]
         )
-        ransac_options.data_type_weights_ *= np.array(
-            [
-                ransac_options.squared_inlier_thresholds_[1],
-                ransac_options.squared_inlier_thresholds_[0],
-            ]
-        ) / np.sum(ransac_options.squared_inlier_thresholds_)
+        ransac_options.data_type_weights_ *= (
+            np.array(
+                [
+                    ransac_options.squared_inlier_thresholds_[1],
+                    ransac_options.squared_inlier_thresholds_[0],
+                ]
+            )
+            / np.sum(ransac_options.squared_inlier_thresholds_)
+        )
         ransac_options.min_num_iterations_ = ransac_cfg["min_num_iterations"]
         ransac_options.final_least_squares_ = ransac_cfg["final_least_squares"]
 

limap/features/models/s2dnet.py

@@ -66,7 +66,7 @@ def print_gpu_memory():
     a = torch.cuda.memory_allocated(0)
     f = r - a  # free inside reserved
 
-    print(np.array([t, r, a, f]) / 2**30)
+    print(np.array([t, r, a, f]) / 2 ** 30)
 
 
 class AdapLayers(nn.Module):
@@ -130,7 +130,7 @@ def _init(self, conf):
             if isinstance(layer, torch.nn.MaxPool2d):
                 current_scale += 1
             if i in self.hypercolumn_indices:
-                self.scales.append(2**current_scale)
+                self.scales.append(2 ** current_scale)
 
         self.adaptation_layers = AdapLayers(
             conf.hypercolumn_layers, conf.output_dim

limap/features/models/vggnet.py

@@ -31,7 +31,7 @@ def _init(self, conf=default_conf):
             if isinstance(layer, torch.nn.MaxPool2d):
                 current_scale += 1
             if i in self.hypercolumn_indices:
-                self.scales.append(2**current_scale)
+                self.scales.append(2 ** current_scale)
 
     def _forward(self, data):
         image = data  # data['image']

limap/line2d/LineTR/line_attention.py

@@ -38,7 +38,7 @@ def __init__(self, n_heads: int, d_feature: int, dropout=0.1):
         self.w_vs = nn.Linear(d_feature, n_heads * dim, bias=True)
         self.fc = nn.Linear(n_heads * dim, d_feature, bias=True)
 
-        self.attention = ScaledDotProduct(scale=dim**0.5)
+        self.attention = ScaledDotProduct(scale=dim ** 0.5)
 
         self.dropout = nn.Dropout(dropout)
         self.layer_norm = nn.LayerNorm(d_feature, eps=1e-6)

limap/line2d/LineTR/line_process.py

@@ -54,7 +54,7 @@ def point_on_line(line, dist_px):
     vec = ep - sp
     if vec[0] != 0:
         m = vec[1] / vec[0]
-        x = np.sqrt(dist_px**2 / (1 + m**2))
+        x = np.sqrt(dist_px ** 2 / (1 + m ** 2))
         y = m * x
     else:
         x = 0
@@ -275,7 +275,7 @@ def change_cv2_T_np(klines_cv):
             kline_ep = [sp_x, sp_y]
 
         # linelength = math.sqrt((kline_ep[0]-kline_sp[0])**2 +(kline_ep[1]-kline_sp[1])**2)
-        linelength = line.lineLength * (2**line.octave)
+        linelength = line.lineLength * (2 ** line.octave)
 
         klines_sp.append(kline_sp)
         klines_ep.append(kline_ep)

limap/line2d/LineTR/line_transformer.py

@@ -186,7 +186,7 @@ def forward(
 def attention(query, key, value):
     dim = query.shape[1]
     scores = (
-        torch.einsum("bdhn,bdhm->bhnm", query, key) / dim**0.5
+        torch.einsum("bdhn,bdhm->bhnm", query, key) / dim ** 0.5
     )  # [3, 64, 4, 512] -> [3, 4, 512, 512]
     prob = torch.nn.functional.softmax(scores, dim=-1)
     return torch.einsum("bhnm,bdhm->bdhn", prob, value), prob

limap/line2d/LineTR/linetr_pipeline.py

@@ -90,7 +90,6 @@ def detect_lsd_lines(self, x, max_n_lines=None):
         return torch.stack(lines).to(x)
 
     def _forward(self, data):
-
         def process_siamese(data, i):
             data_i = {k[:-1]: v for k, v in data.items() if k[-1] == i}
             if self.conf.extractor.name:
@@ -181,30 +180,34 @@ def process_siamese(data, i):
                 ):
                     # Several points are sampled per line, and we supervise them independently
                     b_size = len(pred["lines0"])
-                    samples_assignment, samples_m0, samples_m1 = (
-                        gt_matches_from_homography(
-                            pred["lines0"].reshape(b_size, -1, 2),
-                            pred["lines1"].reshape(b_size, -1, 2),
-                            **data,
-                            pos_th=self.conf.ground_truth.th_positive
-                        )
+                    (
+                        samples_assignment,
+                        samples_m0,
+                        samples_m1,
+                    ) = gt_matches_from_homography(
+                        pred["lines0"].reshape(b_size, -1, 2),
+                        pred["lines1"].reshape(b_size, -1, 2),
+                        **data,
+                        pos_th=self.conf.ground_truth.th_positive
                     )
                     pred["samples_gt_assignment"] = samples_assignment
                     pred["samples_gt_matches0"] = samples_m0
                     pred["samples_gt_matches1"] = samples_m1
 
                 # Compute the GT line association
-                line_assignment, line_m0, line_m1 = (
-                    gt_line_matches_from_homography(
-                        pred["lines0"],
-                        pred["lines1"],
-                        pred["valid_lines0"],
-                        pred["valid_lines1"],
-                        data,
-                        self.conf.ground_truth.n_line_sampled_pts,
-                        self.conf.ground_truth.line_perp_dist_th,
-                        self.conf.ground_truth.overlap_th,
-                    )
+                (
+                    line_assignment,
+                    line_m0,
+                    line_m1,
+                ) = gt_line_matches_from_homography(
+                    pred["lines0"],
+                    pred["lines1"],
+                    pred["valid_lines0"],
+                    pred["valid_lines1"],
+                    data,
+                    self.conf.ground_truth.n_line_sampled_pts,
+                    self.conf.ground_truth.line_perp_dist_th,
+                    self.conf.ground_truth.overlap_th,
                 )
                 pred["line_gt_matches0"] = line_m0
                 pred["line_gt_matches1"] = line_m1
@@ -232,31 +235,37 @@ def process_siamese(data, i):
                 ):
                     # Several points are sampled per line, and we supervise them independently
                     b_size = len(pred["lines0"])
-                    samples_assignment, samples_m0, samples_m1 = (
-                        gt_matches_from_pose_depth(
-                            pred["lines0"].reshape(b_size, -1, 2),
-                            pred["lines1"].reshape(b_size, -1, 2),
-                            **data,
-                            pos_th=self.conf.ground_truth.th_positive,
-                            neg_th=self.conf.ground_truth.th_negative
-                        )[:3]
-                    )
+                    (
+                        samples_assignment,
+                        samples_m0,
+                        samples_m1,
+                    ) = gt_matches_from_pose_depth(
+                        pred["lines0"].reshape(b_size, -1, 2),
+                        pred["lines1"].reshape(b_size, -1, 2),
+                        **data,
+                        pos_th=self.conf.ground_truth.th_positive,
+                        neg_th=self.conf.ground_truth.th_negative
+                    )[
+                        :3
+                    ]
                     pred["samples_gt_assignment"] = samples_assignment
                     pred["samples_gt_matches0"] = samples_m0
                     pred["samples_gt_matches1"] = samples_m1
 
                 # Compute the GT line association
-                line_assignment, line_m0, line_m1 = (
-                    gt_line_matches_from_pose_depth(
-                        pred["lines0"],
-                        pred["lines1"],
-                        pred["valid_lines0"],
-                        pred["valid_lines1"],
-                        data,
-                        self.conf.ground_truth.n_line_sampled_pts,
-                        self.conf.ground_truth.line_perp_dist_th,
-                        self.conf.ground_truth.overlap_th,
-                    )
+                (
+                    line_assignment,
+                    line_m0,
+                    line_m1,
+                ) = gt_line_matches_from_pose_depth(
+                    pred["lines0"],
+                    pred["lines1"],
+                    pred["valid_lines0"],
+                    pred["valid_lines1"],
+                    data,
+                    self.conf.ground_truth.n_line_sampled_pts,
+                    self.conf.ground_truth.line_perp_dist_th,
+                    self.conf.ground_truth.overlap_th,
                 )
                 pred["line_gt_matches0"] = line_m0
                 pred["line_gt_matches1"] = line_m1
@@ -299,21 +308,21 @@ def process_siamese(data, i):
         assert match_mat.shape[0] == 1
         bool_match_mat = match_mat[0] > 0
         pred["line_matches0"] = np.argmax(bool_match_mat, axis=1)
-        pred["line_matches0"][~np.any(bool_match_mat, axis=1)] = (
-            UNMATCHED_FEATURE
-        )
+        pred["line_matches0"][
+            ~np.any(bool_match_mat, axis=1)
+        ] = UNMATCHED_FEATURE
         pred["line_matches1"] = np.argmax(bool_match_mat, axis=0)
-        pred["line_matches1"][~np.any(bool_match_mat, axis=0)] = (
-            UNMATCHED_FEATURE
-        )
+        pred["line_matches1"][
+            ~np.any(bool_match_mat, axis=0)
+        ] = UNMATCHED_FEATURE
         pred["line_matches0"] = torch.from_numpy(pred["line_matches0"])[None]
         pred["line_matches1"] = torch.from_numpy(pred["line_matches1"])[None]
         lmatch_scores = torch.from_numpy(
             distance_matrix[(0,) + np.where(match_mat[0] > 0)]
         )
-        pred["line_match_scores0"] = pred["line_match_scores1"] = (
-            -lmatch_scores[None]
-        )
+        pred["line_match_scores0"] = pred[
+            "line_match_scores1"
+        ] = -lmatch_scores[None]
         return pred
 
     def loss(self, pred, data):

limap/line2d/SOLD2/model/line_detection.py

@@ -178,7 +178,9 @@ def detect(self, junctions, heatmap, device=torch.device("cpu")):
                     dim=-1,
                 )
             )
-            normalized_seg_length = segments_length / (((H**2) + (W**2)) ** 0.5)
+            normalized_seg_length = segments_length / (
+                ((H ** 2) + (W ** 2)) ** 0.5
+            )
 
             # Perform local max search
             num_cand = cand_h.shape[0]
@@ -550,7 +552,7 @@ def detect_local_max(
         """Detection by local maximum search."""
         # Compute the distance threshold
         dist_thresh = (
-            0.5 * (2**0.5) + self.lambda_radius * normalized_seg_length
+            0.5 * (2 ** 0.5) + self.lambda_radius * normalized_seg_length
         )
         # Make it N x 64
         dist_thresh = torch.repeat_interleave(

limap/line2d/SOLD2/model/loss.py

@@ -154,7 +154,7 @@ def space_to_depth(input_tensor, grid_size):
     # (N, bs, bs, C, H//bs, W//bs)
     x = x.permute(0, 3, 5, 1, 2, 4).contiguous()
     # (N, C*bs^2, H//bs, W//bs)
-    x = x.view(N, C * (grid_size**2), H // grid_size, W // grid_size)
+    x = x.view(N, C * (grid_size ** 2), H // grid_size, W // grid_size)
     return x
 
 

limap/line2d/SOLD2/train.py

@@ -373,7 +373,7 @@ def train_single_epoch(
             results = metric_func.metric_results
             average = average_meter.average()
             # Get gpu memory usage in GB
-            gpu_mem_usage = torch.cuda.max_memory_allocated() / (1024**3)
+            gpu_mem_usage = torch.cuda.max_memory_allocated() / (1024 ** 3)
             if compute_descriptors:
                 print(
                     "Epoch [%d / %d] Iter [%d / %d] loss=%.4f (%.4f), junc_loss=%.4f (%.4f), heatmap_loss=%.4f (%.4f), descriptor_loss=%.4f (%.4f), gpu_mem=%.4fGB"
@@ -734,7 +734,7 @@ def record_train_summaries(writer, global_step, scalars, images):
 
     # GPU memory part
     # Get gpu memory usage in GB
-    gpu_mem_usage = torch.cuda.max_memory_allocated() / (1024**3)
+    gpu_mem_usage = torch.cuda.max_memory_allocated() / (1024 ** 3)
     writer.add_scalar("GPU/GPU_memory_usage", gpu_mem_usage, global_step)
 
     # Loss part

limap/line2d/line_utils/merge_lines.py

@@ -104,8 +104,8 @@ def merge_line_cluster(lines):
     if b == 0:
         u = np.array([1, 0]) if a >= c else np.array([0, 1])
     else:
-        m = (c - a + np.sqrt((a - c) ** 2 + 4 * b**2)) / (2 * b)
-        u = np.array([1, m]) / np.sqrt(1 + m**2)
+        m = (c - a + np.sqrt((a - c) ** 2 + 4 * b ** 2)) / (2 * b)
+        u = np.array([1, m]) / np.sqrt(1 + m ** 2)
 
     # Get the center of gravity of all endpoints
     cross = np.mean(points, axis=0)

limap/point2d/superglue/superglue.py

@@ -91,7 +91,7 @@ def attention(
     query: torch.Tensor, key: torch.Tensor, value: torch.Tensor
 ) -> Tuple[torch.Tensor, torch.Tensor]:
     dim = query.shape[1]
-    scores = torch.einsum("bdhn,bdhm->bhnm", query, key) / dim**0.5
+    scores = torch.einsum("bdhn,bdhm->bhnm", query, key) / dim ** 0.5
     prob = torch.nn.functional.softmax(scores, dim=-1)
     return torch.einsum("bhnm,bdhm->bdhn", prob, value), prob
 

limap/pointsfm/database.py

@@ -38,7 +38,7 @@
 
 IS_PYTHON3 = sys.version_info[0] >= 3
 
-MAX_IMAGE_ID = 2**31 - 1
+MAX_IMAGE_ID = 2 ** 31 - 1
 
 CREATE_CAMERAS_TABLE = """CREATE TABLE IF NOT EXISTS cameras (
     camera_id INTEGER PRIMARY KEY AUTOINCREMENT NOT NULL,
@@ -144,7 +144,6 @@ def blob_to_array(blob, dtype, shape=(-1,)):
 
 
 class COLMAPDatabase(sqlite3.Connection):
-
     @staticmethod
     def connect(database_path):
         return sqlite3.connect(database_path, factory=COLMAPDatabase)