Is the batch size works?

[trocolit2]

Hi, thanks for sharing your great job. I tested the experiment with 3DMatch dataset from PREDATOR, however, I am trying to test it but I have faced a problem related to the GPU memory.

RuntimeError: CUDA out of memory. Tried to allocate 404.00 MiB (GPU 0; 5.81 GiB total capacity; 4.00 GiB already allocated; 236.38 MiB free; 4.45 GiB reserved in total by PyTorch)

I changed in config.py the batch size from 1 to 2 (or other any number bigger than 1), but now, another error is rising...

Traceback (most recent call last):
  File "trainval.py", line 62, in <module>
    main()
  File "trainval.py", line 58, in main
    trainer.run()
  File "/home/mcit/GeoTransformer/geotransformer/engine/epoch_based_trainer.py", line 180, in run
    self.train_epoch()
  File "/home/mcit/GeoTransformer/geotransformer/engine/epoch_based_trainer.py", line 95, in train_epoch
    output_dict, result_dict = self.train_step(self.epoch, self.inner_iteration, data_dict)
  File "trainval.py", line 41, in train_step
    output_dict = self.model(data_dict)
  File "/home/mcit/anaconda3/envs/geotransformer/lib/python3.8/site-packages/torch/nn/modules/module.py", line 727, in _call_impl
    result = self.forward(*input, **kwargs)
  File "/home/mcit/GeoTransformer/experiments/geotransformer.3dmatch.stage4.gse.k3.max.oacl.stage2.sinkhorn/model.py", line 75, in forward
    transform = data_dict['transform'].detach()
AttributeError: 'list' object has no attribute 'detach'

With a quick search in the issue history, you reported the solution only works for a batch equal 1. Did you evolve the code to accept a size bigger than 1? Or could you help me with the issue I found after changing the batch size?

Thanks in advance for your time and work.

[qinzheng93]

No, we only support the batch size of 1. Please use DDP to increase the batch size as described in READMD.