Open lucasjinreal opened 1 year ago
Hi. You may disable the option deform_hash and vid_hash for training if you do not have tiny-cuda. We implement a version with vanilla pytorch using positional encoding. But the training time will greatly increase and thus it is not recommended.
inference also not support
In that case, for inference, you may also disable these two options. It should work.
disable deform_hash and vid_hash will effect result performance?
Certainly, the type of motion present within a video sequence can impact the performance of the model to varying degrees. Specifically, if the video primarily contains rigid motion, you can expect the performance to remain relatively consistent.
How do I switch to gloo backend?
Done.
How did you switch? Can you post a how to?
How did you switch? Can you post a how to?
replace code strategy="ddp_find_unused_parameters_true" in train.py line 547 by from pytorch_lightning.strategies import DDPStrategy ddp_gloo = DDPStrategy(process_group_backend="gloo",find_unused_parameters=True) trainer = Trainer(max_steps=hparams.num_steps, precision=16 if hparams.vid_hash == True else 32, callbacks=[checkpoint_callback], logger=logger, accelerator='gpu', devices=hparams.gpus, num_sanity_val_steps=1, benchmark=True, profiler="simple" if len(hparams.gpus) == 1 else None, val_check_interval=hparams.valid_iters, limit_val_batches=hparams.valid_batches, strategy=ddp_gloo)
Hello, tinycudann are really hard to install on windows, at least I failed on cuda11.8 + windows. Would consider add vanilla pytorch in modeling so that users can use without tinycudann?