How to create PyTorch Dataset and Dataloader from multiple point cloud files

[davodogster]

Hi there, Awesome work on the repo!

I have individual point cloud files (e.g a "per object" .txt or .las file containing the X,Y,Z's and a separate file for the target) - It's a multitask problem where each point cloud has 2 associated labels.
Could you give me any pointers on how I can convert this data into a PyTorch dataset or use it in a Dataloader for 3D Deep Learning?
To make things more complicated the point clouds all contain a different number of points.

I have been searching issues and looking for information but haven't come across much for this. Every repo for images and point clouds I would have thought the most important thing would be "train on custom dataset" or "dataloader for custom dataset".
Cheers, Sam

[rusty1s]

For each point cloud, you would create an individual data object, e.g.:

import pandas as pd

pos = torch.from_numpy(pd.read_csv(pos.txt).values).to(torch.float)
y = ... # read label into PyTorch tensor
data = Data(pos=pos, y=y)

You can then simply hold that in a Python list or merge it into a PyG InMemoryDataset and pass it to the DataLoader:

from torch_geometric.loader import DataLoader

loader = DataLoader([data1, data2, data3, ...], batch_size=16, ...)

[rusty1s]

I think it should still be possible for PointNet to learn this global property. I think you might have more success when normalizing both inputs (center point clouds, rotation aligned) and labels (scale them to a fixed interval).

[davodogster]

Thanks! Am currently training with these adjustments. Training MSE can get very good even without scaling the labels. What capability does PyG have for preventing overfitting, e.g TopKPooling? Remembering my Data is sparse / each PC is a different size. I wasn't able to get TopKPooling to work - it gave a tensor dimension error. Thanks, Sam

[rusty1s]

You might have some success in avoiding overfitting by making use of dropout (both on node features and edges via torch.nn.functional.dropout and torch_geometric.utils.dropout_adj).

[davodogster]

Hi @rusty1s , thanks I've been able to implement both of those dropout layers and they do give some improvement. I will warn that torch.nn.functional.dropout gives different predictions on different runs even when using model.eval() so I will be switching to torch.nn.Dropout. Also, for F.dropout If I add training = self.training it makes the model train very poorly train and val loss are much higher) compared to when I omit the training flag.

[rusty1s]

training=self.training should disable dropout during evaluation (model.eval()). As such, F.dropout and torch.nn.Dropout should behave equivalent.