valPSNR increase very slowly

[hogeman2]

Nice work thanks!

I have a question about convergence speed

I've run your training code and almost passed 50 epoch. The valPSNR is still under 15.5.

Is that normal?

also created clips seems to be not sequencial is that ok?

[avinashpaliwal]

• I have added tensorboard. How do the interpolated images look on the validation set?

• Are you using a custom dataset or adobe240fps dataset?

  • If you are using a custom dataset, then it should have slow motion videos. The dataloader takes frame1 and frame8 as reference and interpolates the frames in between.

• Did you use the create_dataset script?

[hogeman2]

Thanks for your reply
First interpolated images looks not good and blurred.

And I used adobe240dataset using create_dataset script

[avinashpaliwal]

Can you attach the interpolated results from tensorboard? Also specify your setup- CPU/GPU, PyTorch version, OS, etc. 15.5 valPSNR is really low. Maybe there is a problem with dataset or setup.

[hogeman2]

I'm using GTX 1060
Is that problem reducing batch size to 4?

I added validation image, and graph

[hogeman2]

#22
seems like this was the problem
I've changed code and graph looks good.

[avinashpaliwal]

#22
seems like this was the problem
I've changed code and graph looks good.

Right. The two reference (left and right corner) look very different. Frames in a clip look very similar since they are slow motion videos. Will fix the code.

[hogeman2]

@avinashpaliwal line 69 in create_dataset.py.
I changed line
images = os.listdir(os.path.join(root, file))
to
images = sorted(os.listdir(os.path.join(root, file)))

[shliang0603]

@avinashpaliwal I have the same problem. I had trained the model to 270 epoch, but the valPSNR keep bellow 15, can you help me ?

the follow is valPSNR changing curve

[avinashpaliwal]

@avinashpaliwal I have the same problem. I had trained the model to 270 epoch, but the valPSNR keep bellow 15, can you help me ?

the follow is valPSNR changing curve

Try the new create_dataset script to generate the dataset and try training again. The model starts converging right from the beginning.