memcpy is super slow in cuda kernels....

.gitignore

@@ -3,3 +3,4 @@ __pycache__
 .autoenv*
 runs
 build
+checkpoints

data/Indoor3DSemSegLoader.py

@@ -5,7 +5,8 @@ import os, sys, h5py, subprocess, shlex
 
 
 def _get_data_files(list_filename):
-    return [line.rstrip() for line in open(list_filename)]
+    with open(list_filename) as f:
+        return [line.rstrip() for line in f]
 
 
 def _load_data_file(name):
@@ -76,7 +77,7 @@ class Indoor3DSemSeg(data.Dataset):
         pt_idxs = np.arange(0, self.num_points)
         np.random.shuffle(pt_idxs)
 
-        current_points = torch.from_numpy(self.points[idx, pt_idxs, :]).type(
+        current_points = torch.from_numpy(self.points[idx, pt_idxs]).type(
             torch.FloatTensor
         )
         current_labels = torch.from_numpy(self.labels[idx, pt_idxs]).type(

data/ModelNet40Loader.py

@@ -8,7 +8,8 @@ sys.path.append(BASE_DIR)
 
 
 def _get_data_files(list_filename):
-    return [line.rstrip()[5:] for line in open(list_filename)]
+    with open(list_filename) as f:
+        return [line.rstrip()[5:] for line in f]
 
 
 def _load_data_file(name):
@@ -80,6 +81,7 @@ class ModelNet40Cls(data.Dataset):
 
     def set_num_points(self, pts):
         self.num_points = pts
+        self.actual_number_of_points = pts
 
     def randomize(self):
         self.actual_number_of_points = min(

models/Pointnet2Cls.py

@@ -128,13 +128,15 @@ if __name__ == "__main__":
     from torch.autograd import Variable
     import numpy as np
     import torch.optim as optim
+    import torch.autograd.profiler as profiler
     B = 2
-    N = 32
+    N = 2048
     inputs = torch.randn(B, N, 9).cuda()
     labels = torch.from_numpy(np.random.randint(0, 3, size=B)).cuda()
-    model = Pointnet2SSG(3)
+    model = Pointnet2MSG(3)
     model.cuda()
 
+
     optimizer = optim.Adam(model.parameters(), lr=1e-2)
 
     model_fn = model_fn_decorator(nn.CrossEntropyLoss())

train_cls.py

@@ -16,6 +16,9 @@ import utils.pytorch_utils as pt_utils
 import utils.data_utils as d_utils
 import argparse
 
+torch.backends.cudnn.enabled = True
+torch.backends.cudnn.benchmark = True
+
 
 def parse_args():
     parser = argparse.ArgumentParser(
@@ -141,8 +144,8 @@ if __name__ == "__main__":
         model,
         model_fn,
         optimizer,
-        checkpoint_name="cls_checkpoint",
-        best_name="cls_best",
+        checkpoint_name="checkpoints/cls_xyz",
+        best_name="checkpoints/cls_xyz_best",
         lr_scheduler=lr_scheduler,
         bnm_scheduler=bnm_scheduler
     )

utils/csrc/group_points_gpu.cu

@@ -21,8 +21,9 @@ __global__ void group_points_kernel(int b, int n, int c, int npoints,
     for (int j = index; j < npoints; j += stride) {
 	for (int k = 0; k < nsample; ++k) {
 	    int ii = idx[j * nsample + k];
-	    memcpy(out + j * nsample * c + k * c, points + ii * c,
-		   sizeof(float) * c);
+	    for (int l = 0; l < c; ++l) {
+		out[j * nsample * c + k * c + l] = points[ii * c + l];
+	    }
 	}
     }
 }

utils/pointnet2_modules.py

@@ -31,7 +31,8 @@ class PointnetSAModuleMSG(nn.Module):
             radii: List[float],
             nsamples: List[int],
             mlps: List[List[int]],
-            bn: bool = True
+            bn: bool = True,
+            use_xyz: bool = True
     ):
         super().__init__()
 
@@ -43,7 +44,9 @@ class PointnetSAModuleMSG(nn.Module):
         for i in range(len(radii)):
             radius = radii[i]
             nsample = nsamples[i]
-            self.groupers.append(pointnet2_utils.QueryAndGroup(radius, nsample))
+            self.groupers.append(
+                pointnet2_utils.QueryAndGroup(radius, nsample, use_xyz=use_xyz)
+            )
             mlp_spec = mlps[i]
             self.mlps.append(pt_utils.SharedMLP(mlp_spec, bn=bn))
 
@@ -111,7 +114,8 @@ class PointnetSAModule(nn.Module):
             npoint: int = None,
             radius: float = None,
             nsample: int = None,
-            bn: bool = True
+            bn: bool = True,
+            use_xyz: bool = True
     ):
         super().__init__()
         self.npoint = npoint
@@ -119,9 +123,11 @@ class PointnetSAModule(nn.Module):
         if self.npoint is not None:
             assert radius is not None
             assert nsample is not None
-            self.grouper = pointnet2_utils.QueryAndGroup(radius, nsample)
+            self.grouper = pointnet2_utils.QueryAndGroup(
+                radius, nsample, use_xyz=use_xyz
+            )
         else:
-            self.grouper = pointnet2_utils.GroupAll()
+            self.grouper = pointnet2_utils.GroupAll(use_xyz=use_xyz)
 
         self.mlp = pt_utils.SharedMLP(mlp, bn=bn)