About raw images

[tzayuan]

Hi @ChenyangLEI ,

I downloaded your raw images from onedrive, and I found that the ambient image is more bright than flash image, Is my understanding wrong?

Also, I tried to manually simulate the shooting process of the original dataset, but it seems that the reflection of the flash on the mirror is very strong, how does the reflection photo of such a situation look like?

Thanks a lot!

Best regards, Zeyuan

[ChenyangLEI]

Hi, sorry for the mistake. The names are wrong for these two images.

Yes, the reflection of the flash should be strong.

[tzayuan]

Hi @ChenyangLEI Thanks for your suggestions.

I still have some questions that I have not been able to solve myself and would like to ask you for advice. :(

1. What is the actual effect of the algorithm for shooting with strong flash reflection?

2. I tried to retrain the model follow your instructure, but I got some pure black result, what could be the cause of this? And the training log looks strange. Epc:002-1421 | percep_r: nan total: nan percep_t: nan time:0.060 Epc:002-0463 | percep_r: nan total: nan percep_t: nan time:0.058 Epc:002-0315 | percep_r: nan total: nan percep_t: nan time:0.061 Epc:002-1593 | percep_r: nan total: nan percep_t: nan time:0.056 Epc:002-1699 | percep_r: nan total: nan percep_t: 9.981 time:0.056 Epc:002-3259 | percep_r: nan total: nan percep_t: nan time:0.185 Epc:002-0598 | percep_r: nan total: nan percep_t: nan time:0.055 Epc:002-3284 | percep_r: nan total: nan percep_t: nan time:0.184 Epc:002-1334 | percep_r: nan total: nan percep_t: nan time:0.057 Epc:002-3907 | percep_r: nan total: nan percep_t: nan time:0.148 Epc:002-2935 | percep_r: nan total: nan percep_t: nan time:0.242 Epc:002-3281 | percep_r: nan total: nan percep_t: nan time:0.184 Epc:002-1383 | percep_r: nan total: nan percep_t: 7.391 time:0.055

3. Also, I found that there are so many ckpt checkpoints file, What is the difference between them? Do the naming of these files have any meaning?

Thanks a lot!

Best regards, Zeyuan

[tzayuan]

Hi @ChenyangLEI

Sorry to interrupt again. We have tried to use the dataset and code you provided and retrained and inferred using the TensorFlow version, but we found that the inferred predicted images etc. are black, what could be the potential reasons for this? Thanks!

Best regards, Zeyuan

[xjiangan]

Hi. Could you provide more information about your hardware? I have successfully reproduced the experiments on NVIDIA 2080 Ti. In addition, I have also tried training on NVIDIA 3080Ti and encountered similar NAN problem. This NAN error could be specific to NVIDIA 30 series GPU.

In addition, for the inference of the retrained model, did you set the model path correctly and did you see in the log messages that the retrained model is actually loaded?

[tzayuan]

Hi @xjiangan ,

I apologize for replying to your message so late, and thank you very much for the very valuable guidance.

My hardware information is as follows: NVIDIA GeForce RTX3090, which agrees with your guidance.

I am interested in reproducing this algorithm, and I will continue to try to reproduce the training and inference of this model below, but it may take a bit longer due to other tasks I am doing.

I'll provide you with further feedback on the results when I'm done, thank you very much!

[Dongj909]

Hi @tzayuan , I wonder how your training and inference of this model went? I encountered the same NaN issue as you did. Could you please share how you resolved it?I am using an NVIDIA A30 device, and my system is CentOS 7. Looking forward to your response.