Seems not converge on Lego about nerf-pytorch

[yenchenlin]

Hello @xrenaa , I've added the cropping operation suggested by the authors here in the newest commit. It helped me solve the "all white" issue on drums and mic.

Do you want to try it again? You just need to re-clone the code and run the same instructions.

[xrenaa]

Hi @yenchenlin , I am happy to see it converge finally. Thanks for your update!

I will close this issue.

[yenchenlin]

Hello, may I know your config?

[Chan0081]

I tried both the configs:

1. current ./configs/lego.txt:

expname = blender_paper_lego
basedir = ./logs
datadir = ./data/nerf_synthetic/lego
dataset_type = blender

no_batching = True

use_viewdirs = True
white_bkgd = True
lrate_decay = 500

N_samples = 64
N_importance = 128
N_rand = 1024

precrop_iters = 500
precrop_frac = 0.5

half_res = True

2. previous config:

expname = lego_test
basedir = ./logs
datadir = ./data/nerf_synthetic/lego
dataset_type = blender

half_res = True

N_samples = 64
N_importance = 64

use_viewdirs = True

white_bkgd = True

N_rand = 1024

and the first output is all white, while the second produces a reasonable result ...

[xrenaa]

Hi, thanks for your reply! I just follow the setting in the repo:

expname = lego_test
basedir = ./logs
datadir = ./data/nerf_synthetic/lego
dataset_type = blender

half_res = True

N_samples = 64
N_importance = 64

use_viewdirs = True

white_bkgd = True

N_rand = 1024

And I follow the installation of the environment.

However, the loss seems not to change and gets stuck about 0.13. This is weird. I do not change any of the original repo. Thank you!

[yenchenlin]

Hello, this is abnormal. Here is my output for the first 250 steps. The loss should drop down to ~0.03 at step 250.

I am attaching my setting for the experiments here:

• config.txt
• args.txt

You can find above files in the specified log path. Can you confirm they match exactly?

Additionally, here is output before loss that contains the dataset information:

Loaded blender (138, 400, 400, 4) torch.Size([40, 4, 4]) [400, 400, 555.5555155968841] ./data/nerf_synthetic/lego
Found ckpts []
Not ndc!
get rays
done, concats
shuffle rays
done
Begin
TRAIN views are [ 0  1  2  3  4  5  6  7  8  9 10 11 12 13 14 15 16 17 18 19 20 21 22 23
 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47
 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71
 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95
 96 97 98 99]
TEST views are [113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130
 131 132 133 134 135 136 137]
VAL views are [100 101 102 103 104 105 106 107 108 109 110 111 112]

[yenchenlin]

@xrenaa any update on this?

[xrenaa]

@yenchenlin Hi, I tried to re-clone the file and create new conda environments multiple times of both python 3.6 and 3.7 versions. However, the loss still stuck about 0.13. I am still trying to figure out the reason. And my output is the same as yours.

[yenchenlin]

I see, I will do my best to help. Can you report the PyTorch and CUDA version?

[xrenaa]

Hi, I run on PyTorch 1.4 and Cuda 10.1.

And I find an interesting thing. At first, the RGB video is plain white and after about 600k iters, the loss come to 0.01 and the video becomes to like this:

[yenchenlin]

Hello @xrenaa , I think the package's version looks right. Do you have other machines to test this? Today, I test it on two brand new machines and it works normally. Let me know if you can reproduce the same issue on multiple machines.

I am using mini-conda to create the environment:

conda create -n tmp python=3.6
conda activate tmp

[yenchenlin]

Glad to know it work! In my experiences, adding this whenever white_bkgd = True helps a lot :)

[kwea123]

I also have this problem with my implementation. The problem is indeed due to all white sampling, which is totally by chance... so I just rerun the experiment if I see the loss doesn't go down after ~100 steps, and finally it will work when I'm lucky in the first iters..

[yenchenlin]

@kwea123 I recommend the cropping solution mentioned above, it eliminates the need for luck.

[kwea123]

Increasing the batch size or changing the optimizer (I use RAdam) also solves the problem for me.

[hengfei-wang]

I have the same problem on my own llff data. Actually, loss went down quickly and converged in a short time in my experiment. But the final rendering images are all black. I see that all your experiments didn't converge at the first stage which is different with my problem. Any idea?

[IntentionsY]

Hello，I need help.Please!!!
I inputed"python run_nerf.py --config configs/lego.txt",then

it shows"no ndc"

please help me ,thank you.

你好，我在pycharm中导入项目之后，加载数据lego.txt在data/nerf_synthetic目录下，开始训练数据，输入python run_nerf.py --config configs/lego.txt，却显示没有ndc，而且val不进行，为0。显示结果如下

请帮帮我，谢谢你

[IntentionsY]

the question is above