code commit

.gitignore

@@ -0,0 +1,5 @@
+__pycache__
+build
+data
+results
+*.so

configs/config_abc.yaml

@@ -0,0 +1,11 @@
+dataset_name: ABC
+dataset_root: data/3d_shapes_abc_training
+
+manifold_points: 3000
+non_manifold_points: 2000
+iter_nbr: 600000
+training_random_rotation_x: 180
+training_random_rotation_y: 180
+training_random_rotation_z: 180
+
+val_interval: 5

configs/config_default.yaml

@@ -0,0 +1,45 @@
+experiment_name: null
+dataset_name: null
+dataset_root: null
+save_dir: 'results'
+train_split: 'training'
+val_split: 'validation'
+test_split: 'test'
+filter_name: null
+
+manifold_points: 2048
+non_manifold_points: 2048
+random_noise: null # 0.1
+normals: False
+
+#training
+training_random_scale: null # 0.1
+training_random_rotation_x: null # 180
+training_random_rotation_y: null # 180
+training_random_rotation_z: null # 180
+training_batch_size: 16
+training_iter_nbr: 100000
+training_lr_start: 0.001
+
+resume: false
+
+network_backbone: FKAConv
+network_latent_size: 32
+network_decoder: InterpAttentionKHeadsNet
+network_decoder_k: 64
+network_n_labels: 2
+
+device: "cuda" 
+threads: 8
+log_mode: "no_log"
+logging: INFO
+
+val_num_mesh: null
+val_interval: 1
+
+
+
+
+
+
+

configs/config_shapenet.yaml

@@ -0,0 +1,8 @@
+dataset_name: ShapeNet
+dataset_root: data/ShapeNet
+
+manifold_points: 3000
+non_manifold_points: 2048
+random_noise: 0.005
+
+iter_nbr: 600000

configs/config_synthetic.yaml

@@ -0,0 +1,12 @@
+dataset_name: SyntheticRooms
+dataset_root: data/synthetic_room_dataset
+
+manifold_points: 10000
+non_manifold_points: 2048
+random_noise: 0.005
+
+training_batch_size: 8
+iter_nbr: 600000
+
+val_interval: 5
+val_num_mesh: 20

datasets/__init__.py

@@ -0,0 +1,15 @@
+from .shapenet import ShapeNet
+
+from .synthetic_room import ShapeNetSyntheticRooms as SyntheticRooms
+
+from .scenenet import SceneNet
+from .scenenet import SceneNet as SceneNet20
+from .scenenet import SceneNet as SceneNet100
+from .scenenet import SceneNet as SceneNet500
+from .scenenet import SceneNet as SceneNet1000
+
+from .abc import ABCTrain as ABC
+from .abc_test import ABCTest, ABCTestNoiseFree, ABCTestExtraNoise
+from .real_world import RealWorld
+from .famous_test import FamousTest, FamousTestNoiseFree, FamousTestExtraNoisy, FamousTestSparse, FamousTestDense
+from .thingi10k_test import Thingi10kTest, Thingi10kTestNoiseFree, Thingi10kTestExtraNoisy, Thingi10kTestSparse, Thingi10kTestDense

datasets/abc.py

@@ -0,0 +1,103 @@
+from torch_geometric.data import Dataset
+from lightconvpoint.datasets.data import Data
+import os
+import numpy as np
+import torch
+import logging
+
+
+class ABCTrain(Dataset):
+
+    def __init__(self, root, split="training", transform=None, filter_name=None, num_non_manifold_points=2048, dataset_size=None, **kwargs):
+
+        super().__init__(root, transform, None)
+
+        logging.info(f"Dataset  - ABC Training - {split} - {dataset_size}")
+
+
+        self.root = os.path.join(self.root, "abc_train")
+
+        self.filenames = []
+        if split in ["train", "training"]:
+            split_file = os.path.join(self.root, "trainset.txt")
+        elif split in ["val", "validation"]:
+            split_file = os.path.join(self.root, "valset.txt")
+        else:
+            raise ValueError("Unknown split")
+
+        with open(split_file) as f:
+            content = f.readlines()
+            content = [line.split("\n")[0] for line in content]
+            content = [os.path.join(self.root, "04_pts", line) for line in content]
+            self.filenames += content
+        self.filenames.sort()
+
+        if dataset_size is not None:
+            self.filenames = self.filenames[:dataset_size]
+
+        logging.info(f"Dataset - len {len(self.filenames)}")
+
+    def get_category(self, f_id):
+        return self.filenames[f_id].split("/")[-2]
+
+    def get_object_name(self, f_id):
+        return self.filenames[f_id].split("/")[-1]
+
+    def get_class_name(self, f_id):
+        return self.metadata[self.get_category(f_id)]["name"]
+
+    @property
+    def raw_file_names(self):
+        return []
+
+    @property
+    def processed_file_names(self):
+        return []
+
+    def _download(self): # override _download to remove makedirs
+        pass
+
+    def download(self):
+        pass
+
+    def process(self):
+        pass
+
+    def _process(self):
+        pass
+
+    def len(self):
+        return len(self.filenames)
+
+
+    def get_data_for_evaluation(self, idx):
+        filename = self.filenames[idx]
+        raise NotImplementedError
+        data_shape = np.load(os.path.join(filename, "pointcloud.npz"))
+        data_space = np.load(os.path.join(filename, "points.npz"))
+        return data_shape, data_space
+
+    def get(self, idx):
+        """Get item."""
+        filename = self.filenames[idx]
+
+        pts_shp = np.load(filename+".xyz.npy")
+
+        filename = filename.replace("04_pts", "05_query_pts")
+        pts_space = np.load(filename+".ply.npy")
+
+        filename = filename.replace("05_query_pts", "05_query_dist")
+        occupancies = np.load(filename+".ply.npy")
+        occupancies = (occupancies>0).astype(np.int64)
+
+        pts_shp = torch.tensor(pts_shp, dtype=torch.float)
+        pts_space = torch.tensor(pts_space, dtype=torch.float)
+        occupancies = torch.tensor(occupancies, dtype=torch.long)
+
+        data = Data(x = torch.ones_like(pts_shp),
+                    shape_id=idx, 
+                    pos=pts_shp,
+                    pos_non_manifold=pts_space, occupancies=occupancies, #
+                    )
+
+        return data

datasets/abc_test.py

@@ -0,0 +1,119 @@
+from torch_geometric.data import Dataset
+from lightconvpoint.datasets.data import Data
+import os
+import numpy as np
+import torch
+import glob
+import logging
+
+
+class ABCTest(Dataset):
+
+    def __init__(self, root, split="training", transform=None, filter_name=None, num_non_manifold_points=2048, dataset_size=None, variant_directory="abc", **kwargs):            
+        super().__init__(root, transform, None)
+
+        logging.info(f"Dataset  - ABC Test - Test only - {dataset_size}")
+
+
+        self.root = os.path.join(self.root, variant_directory)
+
+        self.filenames = []
+        split_file = os.path.join(self.root, "testset.txt")
+
+        with open(split_file) as f:
+            content = f.readlines()
+            content = [line.split("\n")[0] for line in content]
+            content = [os.path.join(self.root, "04_pts", line) for line in content]
+            self.filenames += content
+        self.filenames.sort()
+
+        if dataset_size is not None:
+            self.filenames = self.filenames[:dataset_size]
+
+        logging.info(f"Dataset - len {len(self.filenames)}")
+
+    def get_category(self, f_id):
+        return self.filenames[f_id].split("/")[-2]
+
+    def get_object_name(self, f_id):
+        return self.filenames[f_id].split("/")[-1]
+
+    def get_class_name(self, f_id):
+        return self.metadata[self.get_category(f_id)]["name"]
+
+    @property
+    def raw_file_names(self):
+        return []
+
+    @property
+    def processed_file_names(self):
+        return []
+
+    def _download(self): # override _download to remove makedirs
+        pass
+
+    def download(self):
+        pass
+
+    def process(self):
+        pass
+
+    def _process(self):
+        pass
+
+    def len(self):
+        return len(self.filenames)
+
+
+    def get_data_for_evaluation(self, idx):
+        filename = self.filenames[idx]
+        raise NotImplementedError
+        data_shape = np.load(os.path.join(filename, "pointcloud.npz"))
+        data_space = np.load(os.path.join(filename, "points.npz"))
+        return data_shape, data_space
+
+    def get(self, idx):
+        """Get item."""
+        filename = self.filenames[idx]
+
+        pts_shp = np.load(filename+".xyz.npy")
+
+        # np.savetxt("/root/no_backup/test.xyz", np.concatenate([pts_space, occupancies[:,np.newaxis]], axis=1))
+        # exit()
+
+        pts_shp = torch.tensor(pts_shp, dtype=torch.float)
+        pts_space = torch.ones((1,3), dtype=torch.float)
+        occupancies = torch.ones((1,), dtype=torch.long)
+
+        data = Data(x = torch.ones_like(pts_shp),
+                    shape_id=idx, 
+                    pos=pts_shp,
+                    pos_non_manifold=pts_space, occupancies=occupancies, #
+                    )
+
+        return data
+
+class ABCTestNoiseFree(ABCTest):
+
+    def __init__(self, root, split="training", transform=None, filter_name=None, num_non_manifold_points=2048, variant_directory="abc_noisefree", dataset_size=None, **kwargs):
+
+        super().__init__(root, 
+            split=split, 
+            transform=transform,
+            filter_name=filter_name, 
+            num_non_manifold_points=num_non_manifold_points, 
+            variant_directory=variant_directory, 
+            dataset_size=dataset_size, **kwargs)
+
+
+class ABCTestExtraNoise(ABCTest):
+
+    def __init__(self, root, split="training", transform=None, filter_name=None, num_non_manifold_points=2048, variant_directory="abc_extra_noisy", dataset_size=None, **kwargs):
+
+        super().__init__(root, 
+            split=split, 
+            transform=transform,
+            filter_name=filter_name, 
+            num_non_manifold_points=num_non_manifold_points, 
+            variant_directory=variant_directory, 
+            dataset_size=dataset_size, **kwargs)