how to train on one gpu

[quxu91]

Hi, thanks for your wonderful works! I have only one gpu, can i train the network? If it is possible, how to train on one gpu?

[briannlongzhao]

In my case in the config file that I am using I changed IMS_PER_BATCH to 1 and used --num_gpus 1 in argument

[jaysonpryde]

@briannlongzhao, @xingyizhou @chenxwh for 1 GPU training, should IMS_PER_BATCH need to be set to 1 so that the asserts in:

• single_batch_collator

  def single_batch_collator(batch):
     """
     A batch collator that does nothing.
     """
     assert len(batch) == 1
     return batch[0]

• build_gtr_train_loader

  def build_gtr_train_loader(cfg, mapper):
     ...
     ...
     world_size = get_world_size()
     batch_size = cfg.SOLVER.IMS_PER_BATCH // world_size
     assert batch_size == 1

  in GTR/gtr/data/gtr_dataset_dataloader.py will not raise errors? In my config, my current IMS_PER_BATCH is 16 and it's always raising errors on those asserts. Will it be ok if the asserts will just be commented out?