Train the model

[Zonobia-A]

I have no idea to train a new model on my dataset, if the author can provide a tiny dataset sample.

[rleaver152]

I'd like that too - a simple 'how to' walkthrough assuming knowledge of Python :-)

[sciart17]

I hope so too. Thanks a lot!

[songyn95]

@Zonobia-A @rleaver152 @sciart17 @zhangmozhe
Can you train normally? I am currently troubled by this problem. If you solve it, can you provide a detailed description，Thank you very much

[ndhieunguyen]

Is there anyone who has the dataset? It's very nice if I am shared in order to train the model. Thank you very much

[AsserOssama]

@ndhieunguyen did u know the structure of the dataset to train the model? Thank you very much

[yyang181]

I wrote a dataloader purposed to manage the training process. Kindly examine an exemplar below which shows the dataloader at: https://github.com/yyang181/NTIRE23-VIDEO-COLORIZATION/blob/main/BiSTNet-NTIRE2023/lib/videoloader_woAugImg.py#L1332C1-L1454C31.

Note that the computation of the optical flow between two sequential frames—specifically, flow_forward and flow_backward—might pose some complexities. A strategy I highly recommend to navigate this issue involves the prior computation of the optical flow. It is vitally important to ensure that it is coherently loaded in conjunction with the input images and references.

Please be aware that the authors also offer a script dedicated to the calculation of the optical flow. This valuable resource can be accessed via the following link: https://github.com/zhangmozhe/Deep-Exemplar-based-Video-Colorization/blob/main/lib/videoloader_imagenet.py#L199C13-L215C14.

Hope this helps.