search

FENGGENYU / CAPRI-Net

Code for CAPRI-Net
MIT License
43 stars 6 forks source link

How to get the results? #11

Open xiaoxiaoqiang9527 opened 5 months ago

xiaoxiaoqiang9527 commented 5 months ago

Hello, thanks for your work! I fist used python train.py -e ./data/abc_all -g 0 -p 0. And then python test_pretrain.py -e ./data/abc_all -g 0 -p 0 -c initial --test --voxel --start 0 --end 16 --mc_threshold 0.5. I wonder if I missed some steps, because the results in file 'Reconstructions_test' can't be distinguished. Some of them are even empty. image Looking forward to your reply!

xiaoxiaoqiang9527 commented 5 months ago

This is my data folder image

xiaoxiaoqiang9527 commented 5 months ago

After the fine-tune stage, the results are good. But one issue said that stage 0 is enough image Did I misunderstood something?

FENGGENYU commented 5 months ago

Hi, Xiaoxiaoqiang9527.

Thank you for your interest in my work!

First, you just need to train the first stage. In order to get similar performance as my provided pre-trained weights, I additionally used progressive training strategy as IM-Net (https://github.com/czq142857/IM-NET). Specifically, you could increase the sampled points during training. You could try set ---grid_sample=32 for the first 500 epoches and then set --grid_sample=64 for another 1000 epoches. Please make sure you load your previous checkpoint by using --continue. This should improve the performance of your current model. Sorry I didn't include these details in the main page description as most of people just used my provided pre-trained weights.

Second, the pre-trained weights at stage 0 are not good and could produce empty results. This is because we put the difference operation as the last operation and bad operands could result in empty mesh. Our results in the paper are produced by fine-tuning. In our supplementary material, I showed how much the fine-tuning could improve over pre-trained weights at stage 0.

Best,

Fenggen

xiaoxiaoqiang9527 commented 5 months ago

OK I understand, thank you very much for your detailed answer

xiaoxiaoqiang9527 commented 5 months ago

In the stage of fine-tune, I wonder if I can remove all files named 'last_x.pth' or some of them to reduce memory usage(I don't have a lot of memory space left) ? Sorry for this question, because I'm busy doing other jobs, I haven’t read the entire code yet. Getting your answer will help me a lot, thank you very much!

xiaoxiaoqiang9527 commented 5 months ago

image

FENGGENYU commented 4 months ago

Sure, you could add os.remove("path_to_your_weights") to remove weights after loading the weights at here