search

klightz / PSF

Official Implementation for Fast Point Cloud Generation with Straight Flows
MIT License
32 stars 9 forks source link

Details of Experimental Settings #13

Open namhainguyen2803 opened 1 month ago

namhainguyen2803 commented 1 month ago

Hi, I have some questions concerning about the experimental settings:

  1. In train_flow.py, the default of number of epoch ("niter" I suppose) that the authors set is 20.000 epochs, but in paper, the authors claimed 200.000 steps, which is conflict. Moreover, the argument "bs" which I believe it is "batch-size" is set to 96, but in paper, the authors set to 256 (I dont know how many gpus did the authors have used to train the model but I have used 4 NVDIA-A100 and it does not fit 256 at all, I have to reduce "bs" to 128)
  2. In the paper, the authors set to 50.000 data pairs, but in sample_flow.py, I dont see any "50000" keyword. Perharp the authors set the number of generating sample in lines: for i in range(200): x_gen_eval = model.gen_samples(new_x_chain(x, 64).shape, x.device, clip_denoised=False). I understand that the model will generate 200 times and 64 data each time, but it just has 64x200 data, not enough for 50.000 data.

I hope the authors will address my concern soon.

dreamcubeblock commented 3 weeks ago

I successfully reproduced the code (but not as well as in the paper), my suggestion is to set the experimental parameters according to your computational resources, I don't think it will have much impact on the final result, 100 epochs basically gives better results, in his reply to me he said that he suggests to sample the same number of pairs as the dataset and train the same number of epochs as the flow in the reflow epoch

Mike001-wq commented 2 days ago

I successfully reproduced the code (but not as well as in the paper), my suggestion is to set the experimental parameters according to your computational resources, I don't think it will have much impact on the final result, 100 epochs basically gives better results, in his reply to me he said that he suggests to sample the same number of pairs as the dataset and train the same number of epochs as the flow in the reflow epoch

How much GPU memory is needed?

dreamcubeblock commented 2 days ago

i used one V100 about 24GB, but i have generated 4096 points so if you generate less than that it shouldn't take that much.