About the pipeline about implicit-internal-video-inpainting

[Tengfei-Wang]

Hi, during the training, you can use Tensorboard to watch the training process (and logged results), and early-stop when the results are satisfactory. In our experiment, "BMX-trees" takes ~20,000 iterations (~2 hours) to converge with a single GPU. To accelerate the training, multi-GPUs may be used.
BTW, we put our results of 70 videos from DAVIS in https://hkustconnect-my.sharepoint.com/:f:/g/personal/tfwang_connect_ust_hk/EmtOPEAarqdFj4yDLECg2oABsylzC4lKjuFyyg7o2TTdyQ. If you want our results on DAVIS for reference, you can easily download them without training :)

[sydney0zq]

@Tengfei-Wang I will check the code and results. Thanks for your reply and suggestion.

[sydney0zq]

Sorry for reopening the issue.
Another question is that when the model for 4k videos will be released? I want to test them on some real videos.
Could you please send a copy of rough code to me，I want to read and try to run it. My email is [email protected].

Thanks in advance!

[Tengfei-Wang]

Hi, the code for 4K videos is expected to be released in about two weeks. You can refer to 'RefineModelUp' in model.py and 'build_dataset_up' in dataloader.py for the model and data loader used for the progressive training.

[sydney0zq]

Thanks for your reply.

[BarnabasTakacs]

Hello,

Great work. I wold also like to test the 4K pipeline. Could you provide a sample or some hints on how long it took to train?
Thank you inadvance.
Barnabas