Håkan Ardö

Results 16 comments of Håkan Ardö

Sure, they'll have to be extracted from a bigger experimental tangle, but I'll get back to you with a stand alone micro-benchmark... On Thu, Sep 8, 2022 at 8:39 AM...

Hi, here is the benchmark. Including results as comments at the bottom. On Thu, Sep 8, 2022 at 8:48 AM Hakan Ardo ***@***.***> wrote: > Sure, they'll have to be...

[hub_bench.zip](https://github.com/activeloopai/Hub/files/9547665/hub_bench.zip)

I use a "Intel(R) Core(TM) i7-9700K CPU @ 3.60GHz" system with an "GeForce RTX 2080 Ti" GPU. Thanx for the pointer, I will have a look! On Thu, Sep 15,...

I've benchmarked the deeplake loader using the attached benchmark. It performs slightly better than my branch with shuffle=True, but significantly worse than the python loader with shuffle=False. Am I doing...

Thanx for the tip, num_workers=0 improves things, but I still only get some 80% GPU utilization in my trainings as opposed to about 98% with the shuffle_thread branch. I also...

Se also https://github.com/activeloopai/Hub/issues/1888

That worked, but requires the dataloader to be picklable, which is a bit annoying. But it seems to be enough to wrap it in a separate thread: ```python class DONE:...

I think the main effect comes from doing the data loading on the CPU in parallel with the training on the GPU instead of in series with it. Thereby the...