my version is:
python 3.7 cuda 11.3 torch==1.9.0+cu111 torchvision==0.10.0+cu111 pytorch-lightning: 1.6.0 mmcv-full 1.7.0 mmdet 2.25.3 mmdet3d 1.0.0rc5 mmsegmentation 0.29.1
i used bev_depth_fusion_lss_r50_256x704_128x128_24e.py
and set precision=16, but still got this #error.14
eg:
parser.set_defaults(profiler='simple', deterministic=False, max_epochs=24, accelerator='ddp', num_sanity_val_steps=0, gradient_clip_val=5, limit_val_batches=0, enable_checkpointing=True, precision=16, default_root_dir=os.path.join('./outputs/', exp_name))
i found that when building the model "precision" of the args still is 16,
but after building the model, like this:
then trainer = pl.Trainer.from_argparse_args(args) will return i used "bf16", like this:
Do you have any good suggestions besides modifying the version of torch?
my version is:
python 3.7 cuda 11.3 torch==1.9.0+cu111 torchvision==0.10.0+cu111 pytorch-lightning: 1.6.0 mmcv-full 1.7.0 mmdet 2.25.3 mmdet3d 1.0.0rc5 mmsegmentation 0.29.1i used
bev_depth_fusion_lss_r50_256x704_128x128_24e.pyand set precision=16, but still got this #error.14 eg:parser.set_defaults(profiler='simple', deterministic=False, max_epochs=24, accelerator='ddp', num_sanity_val_steps=0, gradient_clip_val=5, limit_val_batches=0, enable_checkpointing=True, precision=16, default_root_dir=os.path.join('./outputs/', exp_name))i found that when building the model "precision" of the args still is 16,
but after building the model, like this:
then
trainer = pl.Trainer.from_argparse_args(args)will return i used "bf16", like this:Do you have any good suggestions besides modifying the version of torch?
Obviously, I didn't change the precision mode.