Unable to increase batch_size

[ktiays]

I noticed that your code was written with the assumption that batch_size = 1, but when I increased the batch_size, it resulted in dimension errors. I want to know why batch_size is limited to 1.
If it cannot be increased, it will not be possible to more efficiently utilize my device resources.

TPVFormer/dataloader/dataset_wrapper.py

Lines 116 to 127 in bbed188

	def custom_collate_fn(data):
	img2stack = np.stack([d[0] for d in data]).astype(np.float32)
	meta2stack = [d[1] for d in data]
	label2stack = np.stack([d[2] for d in data]).astype(np.int)
	# because we use a batch size of 1, so we can stack these tensor together.
	grid_ind_stack = np.stack([d[3] for d in data]).astype(np.float)
	point_label = np.stack([d[4] for d in data]).astype(np.int)
	return torch.from_numpy(img2stack), \
	meta2stack, \
	torch.from_numpy(label2stack), \
	torch.from_numpy(grid_ind_stack), \
	torch.from_numpy(point_label)

[shadow2469]

I'm having the same problem, and I think it's a very unreasonable thing to put a limit on batch_size.

[huang-yh]

Apart from GPU memory constraint, the batch size is set to one because of 1) the different lengths of point cloud data, and 2) the for-loop we use here to filter out unnecessary sampling locations (same as BEVFormer).
For point cloud lengths, you can simply sample a fixed number of points from each point cloud, and that would fix the error as posted above.
For the for-loop, you can insert even another for-loop to take into account the batch size.

[amundra15]

I am running into the same issue. Was anyone able to resolve it?

Besides,

1. I noticed that the paper says "All models are trained for 24 epochs with a batch size of 8 on 8 A100 GPUs".
2. The issue pointed out above by the author would come up only for LiDAR segmentation and not occupancy prediction, right?