search

ken2576 / vision-nerf

Official PyTorch Implementation of paper "Vision Transformer for NeRF-Based View Synthesis from a Single Input Image", WACV 2023.
MIT License
107 stars 12 forks source link

question about training step #20

Closed crazy-stycxj closed 2 months ago

crazy-stycxj commented 2 months ago

HI! Great Work！ I find the global_step=500000 ，when i trained last time ，the code died，then start_step=100000，not 0，if i train the network again，i need train if global_step > model.start_step + args.n_iters + 1: 600000steps，why not 400000steps，Will this overfit？Isn't 500000step a reasonable training interval? please answer me ，thanks

ken2576 commented 2 months ago

Hi

So I think that part of the code makes it additive for the number of steps to train. For example, if you have a checkpoint at 10k and set the n_iters to be 50k, then the total training steps will be 60k instead of 50k. As for overfitting, it should be fine because at that point the learning rate will be very small. I didn't test more than 500k steps so you might need to check the eval PSNR and see if it goes back down (overfitting).

crazy-stycxj commented 2 months ago

OK，thanks