Train without using robust labels map from Mseg

[luda1013]

Hallo Pros,

i want to implement this and train the model. But i am still stuck at step collect robust label map using Mseg. So i want to try train without robust label map. In the description that we must first modify the dataset because it automatically loads the robust map. Can anyone show which python scripts and which line i need to modify to evade the calling of robust label map? so i can train withtout it. Thank you very much

[Zhaohy21]

Hey Luda, I am also curious about this question. Have you found a solution yet?

[luda1013]

hi @Zhaohy21 , no sadly. but i dont try it yet, because now i alr got some images and i want to trainm but still cannot train the EPE yet, do u already can train with the algorithm?

[Zhaohy21]

Hi Luda, unfortunately neither did I. I also had some problems with my training. Specifically with gt_labels: File "/content/gdrive/MyDrive/Training/RealCarla/PhotorealismEnhancement/code/epe/dataset/pfd.py", line 126, in __getitem__ gt_labels = material_from_gt_label(imageio.imread(gt_label_path)) File "/content/gdrive/MyDrive/Training/RealCarla/PhotorealismEnhancement/code/epe/dataset/pfd.py", line 24, in material_from_gt_label h,w = gt_labelmap.shape
ValueError: too many values to unpack (expected 2) Have you met the same Error?

[lm17918]

i have got the same error, it looks like the gt_label map should be a 1 dimensional images. Did you solve this issue? @Zhaohy21

[Zhaohy21]

@lm17918 yes, it's. I've forgotten how I got it, but I can give you an example, like this image from Pfd dataset:

Also, the problem with this network is that if you want to train your own network and achieve similar results to the paper, you have to use your own gbuffer, not the fake_gbuffer provided by the paper. i got stuck at this step, so i ended up replacing it with another network. I got stuck at this step, so I ended up replacing it with another network

[lm17918]

Thanks for the update!