How to understand batch_size and batch in Grid_pool
Stronger-Huang opened this issue · comments
Hello, thanks for your contribution of PTv2.
You set batch_size=12 in the cfg file, but train_loader brings in each time is two-dimensional, for example: coord: [num_points, 3]. Is this data one of batch?(12 point clouds a time ? but about 720000 points in one batch, so 1 point cloud about 60k points, but s3dis only has 24k points, how to explain that)
In grid_pool function : what the 'batch' means in this function? The batch converted by offset is [0,0,...,1,1,...,5,5]. And it changes with batch_size.
So What the difference between " batch_size " and "batch" in Grid_pool?
I would be grateful if you could answer this question!!
Readme here (https://github.com/Pointcept/Pointcept?tab=readme-ov-file#offset) will be helpful. For real-world data, as each point cloud contains a different number of points, we have to merge the batch (b) and num_point (n) into one dimension.
Thank you for your reply!!
The input of PTv2 contains 720,000 points, so is this the result of adding up 12 batches? Then why is the batch from 0 to 5 (6 batches?), and how many points are there in s3dis?
The input of PTv2 contains 720,000 points, so is this the result of adding up 12 batches?
Yes.
Then why is the batch from 0 to 5 (6 batches?)
Sorry, I can't understand the question.
How many points are there in s3dis?
You can check the process data by printing the number.
Thank you for your reply!!
Then why is the batch from 0 to 5 (6 batches?)
I mean I print variable batch in Grid Pool: [0,0,0,...,1,1,1,...5,5,5], it's separator in 12 batches of Point Cloud.
But why only 0 to 5 (seems 6 batches)
As mix_prob=0.8. we randomly mixup two point cloud.
Thank you for your reply!!
how much can this mixing method improve the accuracy?