search

mahyarnajibi / SSH

SSH: Single Stage Headless Face Detector
Other
835 stars 280 forks source link

How to config training on one GPU #23

Closed liuwuliuyun closed 6 years ago

liuwuliuyun commented 6 years ago

Hi, I have encounter several questions when training on one gpu.

  1. Where to change learning rate? I cant find learning rate para in config files.
  2. I keep encounter this error: Traceback (most recent call last): File "main_train.py", line 63, in max_iter=args.iters, gpus=gpus) File "/SSH/SSH/train.py", line 167, in train_net uid = caffe.NCCL.new_uid() AttributeError: type object 'NCCL' has no attribute 'new_uid' Thanks
liuwuliuyun commented 6 years ago

Solve 2 on my own, this error is due to I haven't install nccl. The code in train.py says we have to use nccl even if we only use one gpu card.

liuwuliuyun commented 6 years ago

About 1, you cant change learning rate from config files.

liuwuliuyun commented 6 years ago

You need to change from solver_ssh.prototxt

Brystone commented 5 years ago

Solve 2 on my own, this error is due to I haven't install nccl. The code in train.py says we have to use nccl even if we only use one gpu card.

Can you share your solution? For the second question, I installed NCLL, but still reported this error AttributeError: type object 'NCCL' has no attribute 'new_uid'.

siriusdemon commented 5 years ago

Solve 2 on my own, this error is due to I haven't install nccl. The code in train.py says we have to use nccl even if we only use one gpu card.

Can you share your solution? For the second question, I installed NCLL, but still reported this error AttributeError: type object 'NCCL' has no attribute 'new_uid'.

Recompile caffe and pycaffe

make clean
make all 
make pycaffe