How to use transposed conv3d correctly?

[GCChen97]

Hi, thanks for the cool torchsparse! Recently I am trying to use it but I don't know how to upsample the input.
I made the following model to test, but it raised an empty kmaps error:

model = tsnn.Conv3d(3, 32, 2, 2, transposed=True).to('cuda')

After debugging, I found that it seems to require some "downsample" convolutions before transposed convolutions. Is my guess right?

Besides, how to set_kmap_mode correctly?

F.set_kmap_mode("hashmap")
model = tsnn.Conv3d(3, 3, 2, 2, transposed=True, generative=True).to('cuda')

I've set the mode as the code above. But it still raised an error and told me to set "hashmap" instead of "hashmap_on_the_fly".

[GCChen97]

I made a hack, and it seems to work currently:

def conv3d(...) 
    ...
    if kernel_size == (1, 1, 1) and stride == (1, 1, 1) and dilation == (1, 1, 1):
        ...    
    else:
        tensor_stride = tuple(input.stride[k] // stride[k] for k in range(3))
        if not generative:
            kmap = input._caches.kmaps.get(
                (tensor_stride, kernel_size, stride, dilation)
            )
            
            # region modification
            hashmap_keys, hashmap_vals = None, None
            if kmap is None:
                kmap = F.build_kernel_map(
                    coords,
                    feats.shape[0],
                    kernel_size,
                    stride,
                    padding,
                    hashmap_keys,
                    hashmap_vals,
                    input.spatial_range,
                    kmap_mode,
                    dataflow,
                    downsample_mode=config.downsample_mode,
                    training=training,
                    ifsort=config.ifsort,
                    split_mask_num=config.split_mask_num,
                    split_mask_num_bwd=config.split_mask_num_bwd,
                )

                hashmap = [kmap["hashmap_keys"], kmap["hashmap_vals"]]

                input._caches.kmaps[(input.stride, kernel_size, stride, dilation)] = kmap
                input._caches.hashmaps[input.stride] = hashmap
            # endregion
    ...

Besides, two convolutions are required before a tranposed one:

model = nn.Sequential(
    tsnn.Conv3d(3, 3, 3, 2),
    tsnn.Conv3d(3, 3, 3, 2),
    tsnn.Conv3d(3, 3, 3, 2),
    tsnn.Conv3d(3, 3, 2, 2, transposed=True),
    tsnn.Conv3d(3, 3, 2, 2, transposed=True),
).to('cuda')