search

JonasSchult / Mask3D

Mask3D predicts accurate 3D semantic instances achieving state-of-the-art on ScanNet, ScanNet200, S3DIS and STPLS3D.
MIT License
525 stars 103 forks source link

Batch size selection #131

Open guilherme-coelho-isi opened 12 months ago

guilherme-coelho-isi commented 12 months ago

Hello, thank you for the excellent work! I'm trying to train on smaller batches, so I can maintain the same resolution but still have the same results on a weaker GPU. Is there any parameter to change the batch size? Thank you in advance, Jonas.

Salutations, Guilherme Coelho Fermino.

guilherme-coelho-isi commented 11 months ago

Found it! However, now I'm a bit worried about the learning rate, which its implementation on Mask3D is inverselly proportional to the batch size, and I couldn't find a parameter to change that. Shouldn't it be direcly proportional do the batch size? (as I've only used softgroup, maybe I could be mistaken with the implementation of Mask3D)

bh-cai commented 11 months ago

Dear sir,I have the same issue, can you tell me where to change the batch_size? Thank you very much!

guilherme-coelho-isi commented 11 months ago

The parameter inputed is data.batch_size, then you input the value you want or your machine can handle. Like:

python main_instance_segmentation.py \ data.batch_size=1

0nandon commented 7 months ago

@guilherme-coelho-isi Hi! did u address the learning rate issue above? And did you try the training with 1 batch size? How is the performance compared to the original Mask3D?