zhjy2016 / splut Goto Github PK

The LUTs generated according to this code are all saved using float32, and the actual value stored in is also a floating point number (e.g.: 2.9525747) not a fixed-point number, which is 4 times different from the actual storage space calculated in the paper. For example, SPLUT-S (X4) need 22Mb to store the LUTs, not 5.5M listed in the paper.
I wonder if the quantized storage is done when the table is saved, will it have a bigger impact on the performance?

When we execute python Inference_SPLUT_L.py, but got these problem:

FileNotFoundError: [Errno 2] No such file or directory: './transfer/SPLUT_L_0.001_4/LUT1_K122_Model_S_A.npy'

We don't know how to obtain the transfer dirs for our inference this model and we doubt that the released code is incomplete. If you can check this released version, please provide complete version. Thank you!

Hello.
Thanks for your work.
May I know how to train the model with x2 upscale?
Is it just change the UPSCALE = 2 in Train_SPLUT_L.py and Transfering_SPLUT_L.py?
Thanks.

Hi. Thank you for your works and for sharing your codes.

I tried to train the model, but there is a problem with the provided codes.

In the training process, network shows around 7.6dB of AVG PSNR consistently as shown above.

Do I miss something for proper training?

Thank you.

Great work!

I'm curious about what kinds of phone you've tested on? Does the model's performance tested on CPUs and phones stay the same ranking comparing with other method?

File "/share4/home/zqiang/SPLUT-main/training_testing_code/utils.py", line 347, in next
self.lr_pngs, self.sharp_pngs, self.sr_pngs = self.shuffle_list(self.lr_pngs, self.sharp_pngs, self.sr_pngs)
ValueError: not enough values to unpack (expected 3, got 0)
===> Training start
0%| | 0/2000000 [00:00<?, ?it/s]
Traceback (most recent call last):
File "Train_SPLUT_S.py", line 230, in
batch_L, batch_H, batch_Stage1 = Iter_H.dequeue()
TypeError: cannot unpack non-iterable NoneType object

We don't know that the '/DIV2K_train_SR/*.png' includes the what type images in ./utils.py, so that cause this problem.
Thank you!

We don't know how to calculate the size of SPLUT-S, SPLUT-M, and SPLUT-L in Table 1. and the way of the calculation isn't mentioned in the paper. Please provide this answer, thank you very much!

Thanks for your work.

I noticed that the data cached in SPLUT is float32 type. Will it affect performance if the float32 type is changed to uint8 type?

Thank you very much!

In Fig.4 from the paper, we don't understand why the LUT(HC) or LUT(WC) can enlarge the receptive fields. The LUT also has the enlarge the receptive fields? why only 1 RF? Thank you very much!!!