metric of nerf-synthetic/ficus

[DHPark98]

Hello, thank you for your work and code. For Blender x4(200x200, ss & w/o refine), I ran the "train_blender_downX.sh". For ficus dataset of NeRF-synthetic, psnr that I got was 22.69, which is quite lower that I expected. So could you give your psnr number of ficus?

Of course, I changed the hyperparameters : H=800, W=800, and downscale=4.

I read events file for psnr : logs/nerf-sr/blender-down4-ficus-800x800-ni64-dp-ds4/fine_psnr_ori_val_full -> 22.69

And after I render the test set("test_blender_downX.sh"), I also calculated the psnr between rendered views and test GT manually: 22.70

[cwchenwang]

Hi, the psnr for ficus is 26.635. I htink the fine_psnr_ori_val_full is only evaluated on one view, not the full scene.

[DHPark98]

I trained nerf-sr w/ ficus dataset once again, and rendered test views. Unfortunately, PSNR didn't increase : 22.73 Could you check my bash file below..?

dataset="ficus" epoch=20 W=800 H=800 accelerator="dp" downscale=4 N_importance=64 batch_size=1024 CUDA_VISIBLE_DEVICE=0 python train_blender.py --name blender-down${downscale}-$dataset-${H}x${W}-ni${N_importance}-${accelerator}-ds${downscale} --accelerator $accelerator
--dataset_mode blender_downX --dataset_root dataset/nerf_synthetic/${dataset} --val_epoch_split test
--checkpoints_dir ./checkpoints/nerf-sr --summary_dir ./logs/nerf-sr
--img_wh $W $H --batch_size $batch_size
--n_epochs ${epoch} --n_epochs_decay 10
--print_freq 100 --vis_freq 1000 --val_freq 1000 --val_epoch_freq 20 --vis_epoch_freq 20 --save_epoch_freq 5
--model nerf_downX --N_coarse 64 --N_importance $N_importance
--lr_policy exp --sigma_activation relu --lr 5e-4 --lr_final 5e-6
--downscale ${downscale}

[cwchenwang]

How did you get the PSNR? On all images?

[DHPark98]

yes, average PSNR of 200 test images.

[cwchenwang]

That's weird. The training script looks ok to me, I can try to find my previous checkpoints, but it may take some time.