Splitting 2D and 3D predictions into single loops to prevent CPU priority issues.

scripts/test_skelda_dataset.py

@@ -344,32 +344,19 @@ def main():
     # Print a dataset sample for debugging
     print(labels[0])
 
-    print("\nRunning predictions ...")
+    print("\nCalculating 2D predictions ...")
-    all_poses = []
+    all_poses_2d = []
-    all_ids = []
     times = []
-    triangulator = rpt.Triangulator(
-        min_match_score=min_match_score, min_group_size=min_group_size
-    )
-    old_scene = ""
-    old_index = -1
     for label in tqdm.tqdm(labels):
         images_2d = []
 
-        if old_scene != label.get("scene", "") or (
-            old_index + datasets[dataset_use]["take_interval"] < label["index"]
-        ):
-            # Reset last poses if scene changes
-            old_scene = label.get("scene", "")
-            triangulator.reset()
-
         try:
             start = time.time()
             for i in range(len(label["imgpaths"])):
                 imgpath = label["imgpaths"][i]
                 img = test_triangulate.load_image(imgpath)
                 images_2d.append(img)
-            print("IMG time:", time.time() - start)
+            time_imgs = time.time() - start
         except cv2.error:
             print("One of the paths not found:", label["imgpaths"])
             continue
@@ -391,7 +378,28 @@ def main():
         poses_2d = utils_2d_pose.get_2d_pose(kpt_model, images_2d)
         poses_2d = test_triangulate.update_keypoints(poses_2d, joint_names_2d)
         time_2d = time.time() - start
-        print("2D time:", time_2d)
+
+        all_poses_2d.append(poses_2d)
+        times.append([time_imgs, time_2d, 0])
+
+    print("\nCalculating 3D predictions ...")
+    all_poses_3d = []
+    all_ids = []
+    triangulator = rpt.Triangulator(
+        min_match_score=min_match_score, min_group_size=min_group_size
+    )
+    old_scene = ""
+    old_index = -1
+    for i in tqdm.tqdm(range(len(labels))):
+        label = labels[i]
+        poses_2d = all_poses_2d[i]
+
+        if old_scene != label.get("scene", "") or (
+            old_index + datasets[dataset_use]["take_interval"] < label["index"]
+        ):
+            # Reset last poses if scene changes
+            old_scene = label.get("scene", "")
+            triangulator.reset()
 
         start = time.time()
         if sum(np.sum(p) for p in poses_2d) == 0:
@@ -402,14 +410,12 @@ def main():
             poses3D = triangulator.triangulate_poses(
                 poses_2d, rpt_cameras, roomparams, joint_names_2d
             )
-
         time_3d = time.time() - start
-        print("3D time:", time_3d)
 
         old_index = label["index"]
-        all_poses.append(np.array(poses3D).tolist())
+        all_poses_3d.append(np.array(poses3D).tolist())
         all_ids.append(label["id"])
-        times.append((time_2d, time_3d))
+        times[i][2] = time_3d
 
     # Print per-step triangulation timings
     print("")
@@ -418,9 +424,11 @@ def main():
     warmup_iters = 10
     if len(times) > warmup_iters:
         times = times[warmup_iters:]
-    avg_time_2d = np.mean([t[0] for t in times])
+    avg_time_im = np.mean([t[0] for t in times])
-    avg_time_3d = np.mean([t[1] for t in times])
+    avg_time_2d = np.mean([t[1] for t in times])
+    avg_time_3d = np.mean([t[2] for t in times])
     tstats = {
+        "img_loading": avg_time_im,
         "avg_time_2d": avg_time_2d,
         "avg_time_3d": avg_time_3d,
         "avg_fps": 1.0 / (avg_time_2d + avg_time_3d),
@@ -430,7 +438,7 @@ def main():
 
     _ = evals.mpjpe.run_eval(
         labels,
-        all_poses,
+        all_poses_3d,
         all_ids,
         joint_names_net=joint_names_3d,
         joint_names_use=eval_joints,
@@ -438,7 +446,7 @@ def main():
     )
     _ = evals.pcp.run_eval(
         labels,
-        all_poses,
+        all_poses_3d,
         all_ids,
         joint_names_net=joint_names_3d,
         joint_names_use=eval_joints,