Is there performance comparison with others, such as original pytorch, DALI... about rising HOT 2 CLOSED

Got it, nice work, thank you!

from rising.

Hi, what do you want benchmarked? Transforms or loading? For both we're as fast as a decent PyTorch implementation can be (since we also use PyTorch internally). We haven't yet benchmarked against DALI since they provide another featureset than we do :)

One thing to note: the transforms and the dataloader itself are comparable with plain PyTorch when it comes to speed.
But since so far, PyTorch does not allow batches transforms (except if you do them manually afterwards, which only works without multiprocessing), the main advantage is to have these in rising, as they are faster then a per-sample transform.

So in general: for the same operation we're as fast as a plain PyTorch datapipeline, but we offer more possibilities when it comes to heavy augmentation workloads.

from rising.