Hi, interesting work indeed! one basic question on config setup: requirement.txt mentions: numpy==1.18.5 torch==1.3.1+cu100 opencv_contrib_python==4.2.0.34

However, above numpy version (1.18.5) is not compatible with python3.9 installed in my environment. For all above 3 requirements, do you recommend any specific python version (e.g. 3.6, 3.7 or 3.8 etc.)? would be good to know the version you had used..? Thank you!

What is [align ratio] used for? I'm confused about the align-ratio... How to set the align ratio in train_HDRUNet.yml if train on custom dataset?

#### datasets
datasets:
  train:
    name: Single_LDR2HDR
    mode: LQGT_condition
    dataroot_LQ: /mnt/diskc/even/Ldr2HdrData/medium_sub
    dataroot_GT: /mnt/diskc/even/Ldr2HdrData/gt_sub
    dataroot_ratio: /data0/NTIRE2021_HDR/000_Train_SingleFrame_FirstStage/alignratio
    use_shuffle: true
    n_workers: 4 #  8
    batch_size: 16
    GT_size: 160
    use_flip: true
    use_rot: true
    condition: image

Is the code

img = io.imread_uint16_png(img_path, alignratio_path)
img = np.clip(np.tanh(img), 0, 1) * 255
img = img.round().astype(np.uint8)
cv2.imwrite(save_path, cv2.cvtColor(img, cv2.COLOR_RGB2BGR))

enough for converting output to HDR image?

In your paper, do you directly use cropped Val dateset's PSNR as the PSNR in the final paper?
In other areas, the best test results should be compared to real restored images.

Thank you for your great work! When i upload 1050 images to ./data,why there are only 11 processed images in ./results/HDRUNet_tanh_l1_bs_16_ps_160/000_Valid_SingleFrame_FirstStage?

Another question： How can i modify the output path?

Hello, thank you for your work, looks like you had a great competition with excellent results! I'm a student who is working on HDRI and want to ask if dataset of this model is still available anywhere because the competiton ended and I cannot reach dataset anymore. Thanks in advance!

What does the value of alignratio in the training dataset mean? The range of gt_image/alignratio is not at [0,1]. What does the ratio do ? How are the alignratio in the train dataset derived?

Thanks for your impressive work!

I have read your code and generate some results using the test.py file. It seems that the reconstructed hdr image is saved as uint16 png format, and in the tonemapped_visualization.py, the hdr image is tone mapped to LDR image for visualization. Am I right? And I also try to directly save the sr_img to .hdr or .exr format using the following code: cv2.imwrite('hdr_img_path.hdr', sr_img) The result seems good. Can I do so to get the HDR image in .hdr or .exr format instead of uint16 png format?

Thank you~

First of all, thanks for your comments in this thread regarding align ratio. Wanted to understand few more details. For e.g.: How to find out what is the maximum nits represented in GT data? Is there any information or metadata indicated where GT is graded for say 1000nits, etc? Essentially, trying to understand whats the relationship of uint16 number in GT file and the nits it might represent. Current understanding is:: GT data is uint16 & gamma encoded and lets say if original range was [0 M], it is always scaled in a way that max of image content will be represented by 65535 where alignratio= 65535/M.

Current thought process to make sense of GT data:

1. GT_aligned = output of function imread_uint16_png() i.e. read GT uint16 from file and divide by alignratio to get original range, say [0 M].
2. At this point, GT_aligned is not normalized to [0 1]. Ideally, would like to normalize it to [0 1] before applying gamma correction. Q1. How to normalize GT_aligned to [0 1]? e.g. GT_norm = function(GT_aligned)..what function to use for normalization?
3. GT_linear = GT_norm ^ gamma
4. Now, GT_linear will be in normalized to [0 1], but trying to understand what 1 represents? For e.g. for linear SDR signal, 1 will represent 100nits. Q2. But, for this GT_linear of HDR signal, 1 will translate to what value of nits? Any idea/ suggestions to understand this? or any assumptions made here?

would be great if you can share your insight in this regard.

Thank you so much for releasing this code!

I had a question regarding the amount of time it takes to complete training. From the paper. I found the following information:

All models are built on the PyTorch framework and trained with NVIDIA 2080Ti GPU. When the patch size of input is set to 256 x 256, the total training time is about 5 days.

I have started training on my system with NVIDIA RTX 3090 GPU. However, after training for 6 days, I notice that only 410k iterations are completed and ~600k iterations are pending. I would expect training to be faster using the 3090 GPU card compared to the 2080Ti GPU card, but this seems to be extremely slow. I am using the code from this repository as is with no modifications apart from the training data paths.

Could you please let me know if I am missing something? Is some additional step required in order to accelerate the training?

Thank you.