Insufficient GPU memory

[sunck1]

Dear authors, I encountered an insufficient GPU memory problem, when I trained the VQGAN model based on Kits19 dataset. I found you trained this medical diffusion model with a 24GB memory GPU. According to your default parameter setting, I failed to train the VQGAN model( n_codes = 16384) on a 50GB memory GPU (although the batch size = 1). It seems like this model can be adopted to process the whole CT scan without cropping. What should I do to cope with this problem, not affecting the performance of your model? I would appreciate if you can give me a hand!!

[mufanq]

I encountered the same problem as well. When training on the A800 with 80GB of VRAM, I also experience out-of-memory issues.

[mufanq]

Dear authors, I encountered an insufficient GPU memory problem, when I trained the VQGAN model based on Kits19 dataset. I found you trained this medical diffusion model with a 24GB memory GPU. According to your default parameter setting, I failed to train the VQGAN model( n_codes = 16384) on a 50GB memory GPU (although the batch size = 1). It seems like this model can be adopted to process the whole CT scan without cropping. What should I do to cope with this problem, not affecting the performance of your model? I would appreciate if you can give me a hand!!

I have fixed this issue, and the modifications I made are as follows:

1. I downgraded the PyTorch version from 2.0 to 1.20 to match the version used by the author.
2. I modified the "downsample" parameter to be consistent with the supplementary materials of the paper.

[TWWinde]

Dear authors, I encountered an insufficient GPU memory problem, when I trained the VQGAN model based on Kits19 dataset. I found you trained this medical diffusion model with a 24GB memory GPU. According to your default parameter setting, I failed to train the VQGAN model( n_codes = 16384) on a 50GB memory GPU (although the batch size = 1). It seems like this model can be adopted to process the whole CT scan without cropping. What should I do to cope with this problem, not affecting the performance of your model? I would appreciate if you can give me a hand!!

I have fixed this issue, and the modifications I made are as follows:

1. I downgraded the PyTorch version from 2.0 to 1.20 to match the version used by the author.
2. I modified the "downsample" parameter to be consistent with the supplementary materials of the paper.

I also encounter this Problem, could you explain it more in detail. Thank you so much for your help!!

[script-Yang]

Dear authors, I encountered an insufficient GPU memory problem, when I trained the VQGAN model based on Kits19 dataset. I found you trained this medical diffusion model with a 24GB memory GPU. According to your default parameter setting, I failed to train the VQGAN model( n_codes = 16384) on a 50GB memory GPU (although the batch size = 1). It seems like this model can be adopted to process the whole CT scan without cropping. What should I do to cope with this problem, not affecting the performance of your model? I would appreciate if you can give me a hand!!

I have fixed this issue, and the modifications I made are as follows:

1. I downgraded the PyTorch version from 2.0 to 1.20 to match the version used by the author.
2. I modified the "downsample" parameter to be consistent with the supplementary materials of the paper.

I also encounter this Problem, could you explain it more in detail. Thank you so much for your help!!

I recently encountered this problem when reproducing the experiment on the Breast MRI dataset. The way I solved it was to replace the default parameter model.downsample=[2,2,2] in the command with model.downsample=[4,4,4]. This is also the recommended parameter mentioned by the author in the supplementary material (specifically for the Brats dataset).