new version number

CHANGELOG.md

 # Changelog
 
+## [2.3.5] - 2023-03-24
+### Fixed 
+- pypi project reach size limit, so we need to assign a new version number.
+
 ## [2.3.4] - 2023-03-23
 ### Added 
 - Add SparseGlobalMaxPool and SparseGlobalAvgPool for training only. libspconv don't support it.

docs/USAGE.md

@@ -145,15 +145,16 @@ class ExampleNet(nn.Module):
         return self.net(x)
 ```
 
-### How To Use SparseConvTranspose
+### Generative Model Usage
 
-```SparseConvTranspose``` (standard upsampling) should only be used in generative model. You need to use a classifier to check if a output coordicates is empty, then set batch indices (or xyz) of that sparse tensor to a negative number:
+```SparseConvTranspose``` (standard upsampling) should only be used in generative model. You need to use a classifier to check if a output coordicates is empty, then set batch indices (or xyz) of that sparse tensor to a negative number.
 
-```Python
-spt.indices[empty_mask, 0] = -1
-```
+1. use ```SparseConvTranspose``` to upsample your sparse conv tensor, this will generate lots of points.
+
+2. use a classifier to get valid indices
+
+3. use ```select_by_index``` to generate new sparse conv tensor
 
-In next sparse convolution, invalid coordinates will be removed until you perform next ```spt.indices[empty_mask, 0] = -1```.
 
 #### Common Mistake 
 * issue [#467](https://github.com/traveller59/spconv/issues/467)

spconv/pytorch/core.py

@@ -212,6 +212,14 @@ class SparseConvTensor(metaclass=SpConvTensorMeta):
 
         return new_spt
     
+    def select_by_index(self, valid_indices: torch.Tensor):
+        new_spt = self.shadow_copy()
+        new_spt.indices = self.indices[valid_indices] 
+        new_spt.features = self.features[valid_indices] 
+        # reuse data must be cleared after modify indices
+        new_spt.indice_dict.clear()
+        return new_spt
+
     def minus(self):
         return self.replace_feature(-self.features)
 

version.txt

-2.3.4
+2.3.5