Can't restore the dense feature map

[WuZhaoyue]

Hello, thanks a lot for this great work. I have a question for you. When I use it on an image where some pixels are 0, I use scn.DenseToSparse to convert it to sparse input format, but when I don't do any processing and directly use scn.SparseToDense to restore this sparse tensor to the original dense tensor, it doesn't work as shown. I would like to ask what is the reason? Thanks a lot!!!

[bottler]

Not sure anyone can help, but can you share your code? It might help see what is happening.

[WuZhaoyue]

Not sure anyone can help, but can you share your code? It might help see what is happening.
Hello, thank you for your kindly reply. Here is the code of a simple example

##############################
import torch
import sparseconvnet as scn
import matplotlib.pyplot as plt

####Creating a simple dense image data and zeroing out parts of the image to simulate sparse data
image = torch.randn(1, 3, 100, 100) # (batch_size, channels, height, width)
image[:, :, 20:40, 30:50] = 0

dimension = 2
spatial_size = torch.Size([100, 100]) # spatial size
dense_to_sparse = scn.DenseToSparse(dimension)
sparse_image = dense_to_sparse(image)

sparse_to_dense = scn.SparseToDense(dimension, 3) # 3 is the number of channels
restored_image = sparse_to_dense(sparse_image)

restored_image = restored_image.squeeze(0)

####Visualize original and recovered images
original_image_np = image.squeeze(0).permute(1, 2, 0).numpy()
restored_image_np = restored_image.permute(1, 2, 0).numpy()

plt.figure(figsize=(12, 6))
plt.subplot(1, 2, 1)
plt.title("Original Image")
plt.imshow(original_image_np)

plt.subplot(1, 2, 2)
plt.title("Restored Image")
plt.imshow(restored_image_np)

plt.show()
#########################################

[bottler]

Looking at the code in https://github.com/facebookresearch/SparseConvNet/blob/main/sparseconvnet/sparseToDense.py and https://github.com/facebookresearch/SparseConvNet/blob/main/sparseconvnet/denseToSparse.py, it seems that the shapes expected and produced are not quite the same. I think you need to take this into account manually.

[btgraham]

Sorry, I am unable to maintain SparseConvNet any more.
However, more recent implementations of submanifold SparseConvnets are more highly optimized , ie.
TorchSparse https://github.com/mit-han-lab/torchsparse,
MinkowskiEngine https://github.com/NVIDIA/MinkowskiEngine
SpConv https://github.com/traveller59/spconv

[WuZhaoyue]

Ok, thank you very much for your friendly answers, I will try another solution based on your suggestions.