Loping151
commented
7 months ago - In Table 1, we present metrics as mean ± covariance on all 6 EndoNeRF datasets. We can plot them like this(nothing to do with Lerplane article):
- In EndoNeRF, one batch is 2048 rays. The actual iter is calculated by: {iteration in config: 1200}*{our batch size: 16384}/2048 = 9k and 3600 for 32k. The number is not very precise, only used to distinguish them.
-
Sorry for the 3min error. I noticed that the version uploaded was the version I used to make the video in README...which means, this version saves an image every 6 iter, which caused the training to pause. You should notice your GPU usage drops very often and there's a warning of accumulation...Anyway, you can directly remove line 98 and 99 in
video_trainer.py, or just git pull. It should be:def train_step(self, data: Dict[str, Union[int, torch.Tensor]], **kwargs): scale_ok = super().train_step(data, **kwargs) # if self.global_step in [900,1800,2700,3600]: # self.validate(self.global_step) # if self.global_step % 6 == 0: # self.validate(video_frame = self.global_step // 6) if self.global_step == self.isg_step: self.train_dataset.enable_isg() raise StopIteration # Whenever we change the dataset if self.global_step == self.ist_step: self.train_dataset.switch_isg2ist() raise StopIteration # Whenever we change the dataset return scale_okI've tested it on an RTX4090 GPU and it takes 2:47 for 9k. Should take around 3min on an RTX3090.
Epsilon404
Hi, thanks for your work and I am interested in it!
But during the reproduction of your work, I'm suffering from some issues.
There are two datasets in EndoNeRF dataset,
cutting_tissues_twiceandpulling_soft_tissues, I wonder which dataset was used in the results of Table 1 in the paper, and how the error like ±1.387 was computed.The "9k" and "32k" in the paper represent the number of iterations, but in the config files
example-9k.pyandexample-32k.pythe parameternum_stepsare set to 1200 and 3600, and they are also the actual iteration number during training. So I'm wondering why and what is the relation between 9k and 1200.I set up the environment on Ubuntu 20.04 and trained with a single RTX3090, and here are my environment dependencies:
environment dependencies
But after the training with the unchanged parameters and 1200 or 3600
So could you please tell me how to reproduce the result of 3 min or are there any additional environment or config settings before training?
num_steps, on the two datasets of EndoNeRF, I got the results with much longer time than in Table 1 of paper:Thank you very much!