New API for datasets/PyTorch dataloaders

[jacobaustin123]

Several issues (#172, #224, #225) mention a new API for interacting with datasets, but I can't find any mention of this online. After adapting several examples to the current API, I have

train_set = kal.datasets.ModelNet(root=args.modelnet_root, categories=args.categories)
dataloader_train = DataLoader(train_set, batch_size=args.batchsize, shuffle=True, num_workers=1)

but iterating over the dataloader raises a type error caused by the ModelNet get_item not returning a Tensor. You can fix this with a custom collate function, like

dataloader = DataLoader(voxels, batch_size=10, shuffle=True, num_workers=1, collate_fn=lambda x : x)

but this also doesn't work with the examples. A few recent examples use torch.randperm to shuffle the dataset, but this doesn't support batch sampling. What is the current standard method for feeding data to DataLoaders?

[jacobaustin123]

I ended up using

transform = tfs.Compose([
    tfs.TriangleMeshToVoxelGrid(resolution=30),
])

# Setup Dataloader
train_set = kal.datasets.ModelNet(root=args.modelnet_root, categories=args.categories, transform=transform)
dataloader_train = DataLoader(train_set, batch_size=args.batchsize, shuffle=True, num_workers=8)

which works fine after you change some attributes around. Is this the desired API at the moment? This also doesn't work because there are some corrupt entries in ModelNet. I was able to modify the ModelNet dataset to catch and randomly replace invalid entries, but this isn't available in the main library. Is this a feature we want to add, or is there a workaround?

    def _get_data(self, index):
        try:
            data = TriangleMesh.from_off(self.filepaths[index])
        except ValueError:
            print(f"[WARNING] sample {self.filepaths[index]} does not work. Picking a random entry instead.")
            idx = np.random.randint(0, len(self))
            data = self._get_data(idx)
    return data

[adityavadalkar]

Hi. Can you tell me the suitable import for your tfs as I'm not able to find TriangleMeshToVoxelGrid in any module.

[MBrandt-NASA]

It would be extremely helpful to see a basic demo / example of how to add or integrate one's own "custom" 3d point cloud data sets to use with Kaolin. Hopefully such a demo / example. could be made for the new Dataset API.