The total training time

[Runsong123]

Hi:
thanks for making this great work available here!
I want to know the total training time for 3000 epochs for the model in this paper (or if you use a single GPU ? ). I find it's so long for me to train the model on a single 2080Ti.

[pengsongyou]

Hi @csuzhuzhuxia ,

If you consider 1- or 3-plane model on the ShapeNet dataset, you should be able to at least run 300K iterations in a day. It is strange that it takes so long for you to train. Maybe you can check when the overhead is. I doubt that it might be in the data loading.

Best,
Songyou

[Runsong123]

Thanks for the quick reply!!
I will check it to find the overhead.

[pengsongyou]

@csuzhuzhuxia you have closed the issue but I am still curious where the overhead was? Thanks!

[Runsong123]

@csuzhuzhuxia you have closed the issue but I am still curious where the overhead was? Thanks!

Sorry for the late reply. I am busy in other things in these days, so I just close the issue although I don't find the reason. And thank you for your concern.
I think you mean the 300K iterations in a day not 300K epochs（for ShapNet, 1 epoch = 959 iteration). But when I do the experiment with the shapnet-1-plane config, it is nearly 22k iteration in a day. I guess it due to that the dataset contains too many small files, and the file is saved in HDD disk, so the loader is so slow.

Thank you!

[pengsongyou]

Still strange because I was also keeping the dataset in HDD (or actually SSD?) and it worked well. But yup, thanks for letting me know!

[Runsong123]

Still strange because I was also keeping the dataset in HDD (or actually SSD?) and it worked well. But yup, thanks for letting me know!

Hello, I train the model in another machine which installed SSD disk to verify the reason. The speed in SSD disk can reach to nearly 300K iteration a day using shapnet-1-plane setting.

zhu

[pengsongyou]

Still strange because I was also keeping the dataset in HDD (or actually SSD?) and it worked well. But yup, thanks for letting me know!

Hello, I train the model in another machine which installed SSD disk to verify the reason. The speed in SSD disk can reach to nearly 300K iteration a day using shapnet-1-plane setting.

zhu

I just checked the machine I used to run the model and I also kept the ShapeNet dataset in HDD. Therefore, this should not be the problem. For the 1-plane model on ShapeNet, when I used GTX 1080ti with HDD disk, I just checked again my output files and it took really only 8 hours to get 300K iterations.

I suggest you checking the following:

• The GPU that you are using. I was using 2080ti and 1080ti (20-30% slower than 2080ti, but not too much).
• num_workers that you use in the config file.
• Check the volatile GPU-Util in nvidia-smi. If there are often cases 0%, it means that GPU is waiting for the data loading from CPU. This happened sometimes to me as well but disappeared by simply running your model again.
• use time.time() in the code, to check where exactly slows down the code. For example, check runtime in between train_step or the train_loader.

Best,
Songyou

[Runsong123]

Still strange because I was also keeping the dataset in HDD (or actually SSD?) and it worked well. But yup, thanks for letting me know!

Hello, I train the model in another machine which installed SSD disk to verify the reason. The speed in SSD disk can reach to nearly 300K iteration a day using shapnet-1-plane setting.
zhu

I just checked the machine I used to run the model and I also kept the ShapeNet dataset in HDD. Therefore, this should not be the problem. For the 1-plane model on ShapeNet, when I used GTX 1080ti with HDD disk, I just checked again my output files and it took really only 8 hours to get 300K iterations.

I suggest you checking the following:

• The GPU that you are using. I was using 2080ti and 1080ti (20-30% slower than 2080ti, but not too much).
• num_workers that you use in the config file.
• Check the volatile GPU-Util in nvidia-smi. If there are often cases 0%, it means that GPU is waiting for the data loading from CPU. This happened sometimes to me as well but disappeared by simply running your model again.
• use time.time() in the code, to check where exactly slows down the code. For example, check runtime in between train_step or the train_loader.

Best,
Songyou

Thanks for your quick reply.
I also compared the running time under different disks(SSD, HDD) keeping the same config provided by this repo, I find the difference is very large. Now, the training speed is fine for me when I change the disk. So I guess there exists some problems with my old HDD disk although the difference between a normal SSD and a normal HDD is not large from your test. The last, thank you again for your patient reply!!

-Zhu