daily updates

Codes/PointNet2.py

@@ -107,8 +107,8 @@ def _sample(self,xyz,npoints):
             sampled_xyz : sampled points [B,C(3),npoints]
 
         '''
-        print("xyz.shape ",xyz.shape)
-        print("npoints: -  ",npoints)
+        print("sample xyz.shape ",xyz.shape)
+        print("sample npoints: -  ",npoints)
 
         if npoints is None:
             return None
@@ -149,14 +149,14 @@ def _group(self,xyz,points,sampled_xyz):
 
             if points is not None:
                 grouped_points = index_points(points.transpose(-2,-1),group_idx).movedim(-1,-3)
-                grouped_points = torch.cat([grouped_points, grouped_xyz],dim=-3)
+                grouped_points = torch.cat([grouped_points, grouped_xyz],dim=1)
             else:
                 grouped_points = grouped_xyz
         else:#group all
             grouped_xyz = xyz.unsqueeze(-2)
             if points is not None:
-                grouped_points =points.unsqueeze(-2)
-                grouped_points = torch.cat([grouped_points,grouped_xyz],dim=-3)
+                grouped_points =points.unsqueeze(2)
+                grouped_points = torch.cat([grouped_points,grouped_xyz],dim=1)
             else:
                 grouped_points = grouped_xyz
         return grouped_points
@@ -174,7 +174,7 @@ def forward(self, xyz, points=None, sampled_xyz=None, npoints=None):
                 new_xyz: Tensor, (B, 3, npoint)
                 new_points: Tensor, (B, mlp[-1], npoint)
             """
-            print(f'npoints: {npoints}, self.nsample: {self.nsample}')
+            print(f'forward - npoints: {npoints}, self.nsample: {self.nsample}')
 
             assert (npoints is None) or (self.npoints is None)
             if npoints is None:

Codes/model.py

@@ -4,7 +4,9 @@
 from torch import einsum
 import warnings
 import logging
-from pytorch3d.ops import  knn_points
+from pytorch3d.ops import  knn_points,sample_farthest_points
+
+from Codes.utils.seedformer_utils import fps_subsample
 from PointNet2 import MLP_Res,MLP_CONV, PointNet_SA_Layer,vTransformer,grouping_operation
 logger = logging.getLogger(__name__)
 
@@ -13,12 +15,13 @@ class FeatureExtractor(nn.Module):
     def __init__(self, out_dim=1024, n_knn=20):
         """Encoder that encodes information of partial point cloud
         """
+        #in_channel values are not correct needs to be investigated
         super(FeatureExtractor, self).__init__()
-        self.sa_module_1 = PointNet_SA_Layer(npoints=512,nsample=16,in_channel=3,mlp_channels=[64,128] )
+        self.sa_module_1 = PointNet_SA_Layer(npoints=512,nsample=16,in_channel=6,mlp_channels=[64,128] )
         self.transformer_1 = vTransformer(128, dim=64, n_knn=n_knn)
-        self.sa_module_2 = PointNet_SA_Layer(npoints=128,nsample=16,in_channel=128,mlp_channels=[128,256])
+        self.sa_module_2 = PointNet_SA_Layer(npoints=128,nsample=16,in_channel=131,mlp_channels=[128,256])
         self.transformer_2 = vTransformer(256, dim=64, n_knn=n_knn)
-        self.sa_module_3 = PointNet_SA_Layer(npoints=None,nsample=None,in_channel=256,mlp_channels=[512,out_dim])
+        self.sa_module_3 = PointNet_SA_Layer(npoints=None,nsample=None,in_channel=259,mlp_channels=[512,out_dim])
 
     def forward(self, partial_cloud):
         """
@@ -39,8 +42,6 @@ def forward(self, partial_cloud):
 
         return l3_points, l2_xyz, l2_points
 
-
-
 class SeedGenerator(nn.Module):
     def __init__(self, feat_dim=512, seed_dim=128, n_knn=20, factor=2, attn_channel=True):
         super(SeedGenerator, self).__init__()
@@ -307,7 +308,8 @@ def forward_decoder(self, feat, partial_cloud, patch_xyz, patch_feat):
         pred_pcds.append(seed)
 
         # Upsample layers
-        pcd = fps_subsample(torch.cat([seed, partial_cloud], 1), self.num_p0) # (B, num_p0, 3)
+        #pcd = fps_subsample(torch.cat([seed, partial_cloud], 1), self.num_p0) # (B, num_p0, 3)
+        pcd,_=sample_farthest_points(torch.cat([seed, partial_cloud], 1), K=self.num_p0,random_start_point=True)
         K_prev = None
         pcd = pcd.permute(0, 2, 1).contiguous() # (B, 3, num_p0)
         seed = seed.permute(0, 2, 1).contiguous() # (B, 3, 256)

Codes/utils/dataloader.py

@@ -0,0 +1,2 @@
+import torch
+from tqdm import tqdm

Codes/utils/seedformer_utils.py

@@ -675,7 +675,7 @@ def forward(self, xyz, points, idx=None):
             return new_xyz, new_points
 
 
-def fps_subsample(pcd, n_points=2048):#Not so correct !!!
+def fps_subsample(pcd, n_points=512):
     """
     Args
         pcd: (b, 16384, 3)
@@ -689,13 +689,14 @@ def fps_subsample(pcd, n_points=2048):#Not so correct !!!
         raise ValueError(
             'FPS subsampling receives a larger n_points: {:d} > {:d}'.format(
                 n_points, pcd.shape[1]))
+    #indices = sample_farthest_points(pcd, K=n_points)
+
     new_pcd = gather_operation(
         pcd.permute(0, 2, 1).contiguous(),
-        sample_farthest_points(pcd, n_points))
+        sample_farthest_points(pcd, K=n_points))
     new_pcd = new_pcd.permute(0, 2, 1).contiguous()
     return new_pcd
 
-
 def get_nearest_index(target, source, k=1, return_dis=False):
     """
     Args: