fix nan

configs/server/server_split_dataset_config.yaml

@@ -6,17 +6,17 @@ runner:
     cuda_visible_devices: "0,1,2,3,4,5,6,7"
     
   experiment:
-    name: debug
+    name: server_split_dataset
     root_dir: "experiments"
   
   split: # 
-    root_dir: "/data/hofee/data/packed_preprocessed_data"
+    root_dir: "/data/hofee/data/new_full_data"
     type: "unseen_instance" # "unseen_category"
     datasets:
       OmniObject3d_train: 
-        path: "/data/hofee/data/OmniObject3d_train.txt"
+        path: "/data/hofee/data/new_full_data_list/OmniObject3d_train.txt"
         ratio: 0.9
 
       OmniObject3d_test: 
-        path: "/data/hofee/data/OmniObject3d_test.txt"
+        path: "/data/hofee/data/new_full_data_list/OmniObject3d_test.txt"
         ratio: 0.1

configs/server/server_train_config.yaml

@@ -7,7 +7,7 @@ runner:
     parallel: False
     
   experiment:
-    name: debug
+    name: train_ab_global_and_partial_global
     root_dir: "experiments"
     use_checkpoint: False
     epoch: -1 # -1 stands for last epoch
@@ -28,50 +28,50 @@ runner:
       #- OmniObject3d_test
       - OmniObject3d_val
 
-  pipeline: nbv_reconstruction_global_pts_n_num_pipeline
+  pipeline: nbv_reconstruction_pipeline
   
 dataset:
   OmniObject3d_train:
-    root_dir: "/home/data/hofee/project/nbv_rec/data/sample_for_training_new"
+    root_dir: "/data/hofee/data/new_full_data"
     model_dir: "../data/scaled_object_meshes"
     source: nbv_reconstruction_dataset
-    split_file: "/home/data/hofee/project/nbv_rec/data/sample.txt"
+    split_file: "/data/hofee/data/new_full_data_list/OmniObject3d_train.txt"
     type: train
     cache: True
     ratio: 1
-    batch_size: 160
-    num_workers: 16
+    batch_size: 80
+    num_workers: 128
     pts_num: 8192
     load_from_preprocess: True
 
   OmniObject3d_test:
-    root_dir: "/home/data/hofee/project/nbv_rec/data/sample_for_training_new"
+    root_dir: "/data/hofee/data/new_full_data"
     model_dir: "../data/scaled_object_meshes"
     source: nbv_reconstruction_dataset
-    split_file: "/home/data/hofee/project/nbv_rec/data/sample.txt"
+    split_file: "/data/hofee/data/new_full_data_list/OmniObject3d_test.txt"
     type: test
     cache: True
     filter_degree: 75
     eval_list:
       - pose_diff
-    ratio: 0.05
-    batch_size: 160
+    ratio: 1
+    batch_size: 80
     num_workers: 12
     pts_num: 8192
     load_from_preprocess: True
 
   OmniObject3d_val:
-    root_dir: "/home/data/hofee/project/nbv_rec/data/sample_for_training_new"
+    root_dir: "/data/hofee/data/new_full_data"
     model_dir: "../data/scaled_object_meshes"
     source: nbv_reconstruction_dataset
-    split_file: "/home/data/hofee/project/nbv_rec/data/sample.txt"
+    split_file: "/data/hofee/data/new_full_data_list/OmniObject3d_train.txt"
     type: test
     cache: True
     filter_degree: 75
     eval_list:
       - pose_diff
-    ratio: 0.005
-    batch_size: 160
+    ratio: 0.1
+    batch_size: 80
     num_workers: 12
     pts_num: 8192
     load_from_preprocess: True
@@ -97,7 +97,7 @@ module:
     feature_transform: False
 
   transformer_seq_encoder:
-    embed_dim: 256
+    embed_dim: 320
     num_heads: 4
     ffn_dim: 256
     num_layers: 3

core/nbv_dataset.py

@@ -7,6 +7,7 @@ from PytorchBoot.utils.log_util import Log
 import torch
 import os
 import sys
+import time
 
 sys.path.append(r"/data/hofee/project/nbv_rec/nbv_reconstruction")
 
@@ -34,7 +35,7 @@ class NBVReconstructionDataset(BaseDataset):
             #self.model_dir = config["model_dir"]
             self.filter_degree = config["filter_degree"]
         if self.type == namespace.Mode.TRAIN:
-            scale_ratio = 100
+            scale_ratio = 1
             self.datalist = self.datalist*scale_ratio
         if self.cache:
             expr_root = ConfigManager.get("runner", "experiment", "root_dir")
@@ -114,8 +115,13 @@ class NBVReconstructionDataset(BaseDataset):
         except Exception as e:
             Log.error(f"Save cache failed: {e}")
 
-    def voxel_downsample_with_mask(self, pts, voxel_size):
-        pass
+    def voxel_downsample_with_mapping(self, point_cloud, voxel_size=0.003):
+        voxel_indices = np.floor(point_cloud / voxel_size).astype(np.int32)
+        unique_voxels, inverse, counts = np.unique(voxel_indices, axis=0, return_inverse=True, return_counts=True)
+        idx_sort = np.argsort(inverse)
+        idx_unique = idx_sort[np.cumsum(counts)-counts]
+        downsampled_points = point_cloud[idx_unique]
+        return downsampled_points, inverse
 
 
     def __getitem__(self, index):
@@ -129,6 +135,9 @@ class NBVReconstructionDataset(BaseDataset):
             scanned_coverages_rate,
             scanned_n_to_world_pose,
         ) = ([], [], [])
+        start_time = time.time()
+        start_indices = [0]
+        total_points = 0
         for view in scanned_views:
             frame_idx = view[0]
             coverage_rate = view[1]
@@ -150,8 +159,12 @@ class NBVReconstructionDataset(BaseDataset):
             n_to_world_trans = n_to_world_pose[:3, 3]
             n_to_world_9d = np.concatenate([n_to_world_6d, n_to_world_trans], axis=0)
             scanned_n_to_world_pose.append(n_to_world_9d)
+            total_points += len(downsampled_target_point_cloud)
+            start_indices.append(total_points)
 
 
+        end_time = time.time()
+        #Log.info(f"load data time: {end_time - start_time}")
         nbv_idx, nbv_coverage_rate = nbv[0], nbv[1]
         nbv_path = DataLoadUtil.get_path(self.root_dir, scene_name, nbv_idx)
         cam_info = DataLoadUtil.load_cam_info(nbv_path)
@@ -164,14 +177,27 @@ class NBVReconstructionDataset(BaseDataset):
         best_to_world_9d = np.concatenate(
             [best_to_world_6d, best_to_world_trans], axis=0
         )
-
-        combined_scanned_views_pts = np.concatenate(scanned_views_pts, axis=0)
-        voxel_downsampled_combined_scanned_pts_np = PtsUtil.voxel_downsample_point_cloud(combined_scanned_views_pts, 0.002)
-        random_downsampled_combined_scanned_pts_np = PtsUtil.random_downsample_point_cloud(voxel_downsampled_combined_scanned_pts_np, self.pts_num)
         
+        combined_scanned_views_pts = np.concatenate(scanned_views_pts, axis=0)
+        voxel_downsampled_combined_scanned_pts_np, inverse = self.voxel_downsample_with_mapping(combined_scanned_views_pts, 0.003)
+        random_downsampled_combined_scanned_pts_np, random_downsample_idx = PtsUtil.random_downsample_point_cloud(voxel_downsampled_combined_scanned_pts_np, self.pts_num, require_idx=True)
+
+        all_idx_unique = np.arange(len(voxel_downsampled_combined_scanned_pts_np))
+        all_random_downsample_idx = all_idx_unique[random_downsample_idx]
+        scanned_pts_mask = []
+        for idx, start_idx in enumerate(start_indices):
+            if idx == len(start_indices) - 1:
+                break
+            end_idx = start_indices[idx+1]
+            view_inverse = inverse[start_idx:end_idx]
+            view_unique_downsampled_idx = np.unique(view_inverse)
+            view_unique_downsampled_idx_set = set(view_unique_downsampled_idx)
+            mask = np.array([idx in view_unique_downsampled_idx_set for idx in all_random_downsample_idx])
+            scanned_pts_mask.append(mask)
         data_item = {
             "scanned_pts": np.asarray(scanned_views_pts, dtype=np.float32), # Ndarray(S x Nv x 3)
             "combined_scanned_pts": np.asarray(random_downsampled_combined_scanned_pts_np, dtype=np.float32), # Ndarray(N x 3)
+            "scanned_pts_mask": np.asarray(scanned_pts_mask, dtype=np.bool), # Ndarray(N)
             "scanned_coverage_rate": scanned_coverages_rate, # List(S): Float, range(0, 1)
             "scanned_n_to_world_pose_9d": np.asarray(scanned_n_to_world_pose, dtype=np.float32), # Ndarray(S x 9)
             "best_coverage_rate": nbv_coverage_rate, # Float, range(0, 1)
@@ -197,7 +223,9 @@ class NBVReconstructionDataset(BaseDataset):
             collate_data["scanned_n_to_world_pose_9d"] = [
                 torch.tensor(item["scanned_n_to_world_pose_9d"]) for item in batch
             ]
-            
+            collate_data["scanned_pts_mask"] = [
+                torch.tensor(item["scanned_pts_mask"]) for item in batch
+            ]
             ''' ------ Fixed Length ------ '''
             
             collate_data["best_to_world_pose_9d"] = torch.stack(
@@ -206,17 +234,14 @@ class NBVReconstructionDataset(BaseDataset):
             collate_data["combined_scanned_pts"] = torch.stack(
                 [torch.tensor(item["combined_scanned_pts"]) for item in batch]
             )
-            collate_data["scanned_pts_mask"] = torch.stack(
-                [torch.tensor(item["scanned_pts_mask"]) for item in batch]
-            )
-                
+            
             for key in batch[0].keys():
                 if key not in [
                     "scanned_pts",
-                    "scanned_pts_mask",
                     "scanned_n_to_world_pose_9d",
                     "best_to_world_pose_9d",
                     "combined_scanned_pts",
+                    "scanned_pts_mask",
                 ]:
                     collate_data[key] = [item[key] for item in batch]
             return collate_data
@@ -232,9 +257,9 @@ if __name__ == "__main__":
     torch.manual_seed(seed)
     np.random.seed(seed)
     config = {
-        "root_dir": "/data/hofee/data/packed_preprocessed_data",
+        "root_dir": "/data/hofee/nbv_rec_part2_preprocessed",
         "source": "nbv_reconstruction_dataset",
-        "split_file": "/data/hofee/data/OmniObject3d_train.txt",
+        "split_file": "/data/hofee/data/sample.txt",
         "load_from_preprocess": True,
         "ratio": 0.5,
         "batch_size": 2,

core/pipeline.py

@@ -20,8 +20,8 @@ class NBVReconstructionPipeline(nn.Module):
         self.pose_encoder = ComponentFactory.create(
             namespace.Stereotype.MODULE, self.module_config["pose_encoder"]
         )
-        self.transformer_seq_encoder = ComponentFactory.create(
-            namespace.Stereotype.MODULE, self.module_config["transformer_seq_encoder"]
+        self.seq_encoder = ComponentFactory.create(
+            namespace.Stereotype.MODULE, self.module_config["seq_encoder"]
         )
         self.view_finder = ComponentFactory.create(
             namespace.Stereotype.MODULE, self.module_config["view_finder"]
@@ -54,10 +54,7 @@ class NBVReconstructionPipeline(nn.Module):
         return perturbed_x, random_t, target_score, std
 
     def forward_train(self, data):
-        start_time = time.time()
         main_feat = self.get_main_feat(data)
-        end_time = time.time()
-        print("get_main_feat time: ", end_time - start_time)
         """ get std """
         best_to_world_pose_9d_batch = data["best_to_world_pose_9d"]
         perturbed_x, random_t, target_score, std = self.pertube_data(
@@ -92,25 +89,49 @@ class NBVReconstructionPipeline(nn.Module):
             "scanned_n_to_world_pose_9d"
         ]  # List(B): Tensor(S x 9)
 
+        scanned_pts_mask_batch = data["scanned_pts_mask"] # List(B): Tensor(N)
+
         device = next(self.parameters()).device
 
         embedding_list_batch = []
 
         combined_scanned_pts_batch = data["combined_scanned_pts"]  # Tensor(B x N x 3)
-        global_scanned_feat = self.pts_encoder.encode_points(
-            combined_scanned_pts_batch, require_per_point_feat=False
+        global_scanned_feat, per_point_feat_batch = self.pts_encoder.encode_points(
+            combined_scanned_pts_batch, require_per_point_feat=True
         )  # global_scanned_feat: Tensor(B x Dg)
-
-        for scanned_n_to_world_pose_9d in scanned_n_to_world_pose_9d_batch:
-            scanned_n_to_world_pose_9d = scanned_n_to_world_pose_9d.to(device)  # Tensor(S x 9)
+        batch_size = len(scanned_n_to_world_pose_9d_batch)
+        for i in range(batch_size):
+            seq_len = len(scanned_n_to_world_pose_9d_batch[i])
+            scanned_n_to_world_pose_9d = scanned_n_to_world_pose_9d_batch[i].to(device)  # Tensor(S x 9)
+            scanned_pts_mask = scanned_pts_mask_batch[i] # Tensor(S x N)
+            per_point_feat = per_point_feat_batch[i] # Tensor(N x Dp)
+            partial_point_feat_seq = []
+            for j in range(seq_len):
+                partial_per_point_feat = per_point_feat[scanned_pts_mask[j]]
+                if partial_per_point_feat.shape[0] == 0:
+                    partial_point_feat = torch.zeros(per_point_feat.shape[1], device=device)
+                else:
+                    partial_point_feat = torch.mean(partial_per_point_feat, dim=0) # Tensor(Dp)
+                partial_point_feat_seq.append(partial_point_feat)
+            partial_point_feat_seq = torch.stack(partial_point_feat_seq, dim=0) # Tensor(S x Dp)
+            
             pose_feat_seq = self.pose_encoder.encode_pose(scanned_n_to_world_pose_9d)  # Tensor(S x Dp) 
-            seq_embedding = pose_feat_seq
+
+            seq_embedding = torch.cat([partial_point_feat_seq, pose_feat_seq], dim=-1)
+            
             embedding_list_batch.append(seq_embedding) # List(B): Tensor(S x (Dp))
         
-        seq_feat = self.transformer_seq_encoder.encode_sequence(embedding_list_batch) # Tensor(B x Ds)
+        seq_feat = self.seq_encoder.encode_sequence(embedding_list_batch) # Tensor(B x Ds)
         main_feat = torch.cat([seq_feat, global_scanned_feat], dim=-1) # Tensor(B x (Ds+Dg))
 
         if torch.isnan(main_feat).any():
+            for i in range(len(main_feat)):
+                if torch.isnan(main_feat[i]).any():
+                    scanned_pts_mask = scanned_pts_mask_batch[i]
+                    Log.info(f"scanned_pts_mask shape: {scanned_pts_mask.shape}")
+                    Log.info(f"scanned_pts_mask sum: {scanned_pts_mask.sum()}")
+                    import ipdb
+                    ipdb.set_trace()
             Log.error("nan in main_feat", True)
 
         return main_feat

utils/pts.py

@@ -14,16 +14,38 @@ class PtsUtil:
             downsampled_points = point_cloud[idx_unique]
             return downsampled_points, idx_unique
         else:
-            unique_voxels = np.unique(voxel_indices, axis=0, return_inverse=True)
-            return unique_voxels[0]*voxel_size
+            import ipdb; ipdb.set_trace()
+            unique_voxels = np.unique(voxel_indices, axis=0, return_inverse=False)
+            return unique_voxels*voxel_size
+
+    @staticmethod
+    def voxel_downsample_point_cloud_o3d(point_cloud, voxel_size=0.005):
+        pcd = o3d.geometry.PointCloud()
+        pcd.points = o3d.utility.Vector3dVector(point_cloud)
+        pcd = pcd.voxel_down_sample(voxel_size)
+        return np.asarray(pcd.points)
     
     @staticmethod
-    def random_downsample_point_cloud(point_cloud, num_points, require_idx=False):
+    def voxel_downsample_point_cloud_and_trace_o3d(point_cloud, voxel_size=0.005):
+        pcd = o3d.geometry.PointCloud()
+        pcd.points = o3d.utility.Vector3dVector(point_cloud)
+        max_bound = pcd.get_max_bound()
+        min_bound = pcd.get_min_bound()
+        pcd = pcd.voxel_down_sample_and_trace(voxel_size, max_bound, min_bound, True)
+        
+        return np.asarray(pcd.points)
+       
+    @staticmethod
+    def random_downsample_point_cloud(point_cloud, num_points, require_idx=False, replace=True):
         if point_cloud.shape[0] == 0:
             if require_idx:
                 return point_cloud, np.array([])
             return point_cloud
-        idx = np.random.choice(len(point_cloud), num_points, replace=True)
+        if not replace and num_points > len(point_cloud):
+            if require_idx:
+                return point_cloud, np.arange(len(point_cloud))
+            return point_cloud
+        idx = np.random.choice(len(point_cloud), num_points, replace=replace)
         if require_idx:
             return point_cloud[idx], idx
         return point_cloud[idx]