Training code coming soon?

[tachibanayoshino]

Training code coming soon?

[maderix]

animeganv2-hayao-models.zip
Here are the converted models from the checkpoint files if anyone's interested. :)

[TachibanaYoshino]

Currently, it is not recommended to change the generator network structure. If you still want to use it on the mobile terminal, you can refer to here: AnimeGANv2をCore MLに変換してiOSでつかう【変換済みモデルあり】 - Qiita https://qiita.com/john-rocky/items/e4c757306b4d8f0f2c36
In addition, AnimeGANv2 has added a new loss function, which cannot be told before all open source. You can use it for fun.
But without money, I may not be able to do more research. Hahaha

[TachibanaYoshino]

Maybe you can find a solution from TF Select. Regarding normal padding, what kind of padding do you mean? You might as well try it first. If it doesn't work, I will change the padding method and retrain the weights for you.

[TachibanaYoshino]

You can refer to the training code of AnimeGAN. Since AnimeGANv2 is still in the research stage of technology sales, I am not sure whether all the code will be open sourced. And its generator network can still be compressed even smaller. Therefore, for the time being, I cannot have an accurate answer.

[maderix]

Thanks for your reply. Let me see what I can do with your suggestion 😊

[TachibanaYoshino]

In fact, to run in real-time on a mobile terminal, a model size of several hundred Kb is a common mode. Simply compressing the network structure is not the most important choice. Model pruning is a necessary method.

[maderix]

Thanks for your reply.
Yes I've tried quantization for tflite and I managed to get the model size down to 2 MB but full integer quantization isn't happening due to some ops like mirror padding being used in the network which are not supported by tflite.
Do you have any suggestions?
For example, can I replace mirror padding with normal padding?
Thanks for your time 😊

[maderix]

Thanks I'll give it a try with the possible alternatives.

[maderix]

I meant instead of pattern mode 'reflect' being used in the generator, is it possible to use a different pattern mode in the generator layers.
I think I might figure out a solution soon . If not with modifying the current generator weights , I can probably try training a smaller model with the generator model used as a teacher for the smaller model.
I'll keep you posted regarding my efforts.
Thank you.

[wenfuli]

animeganv2-hayao-models.zip Here are the converted models from the checkpoint files if anyone's interested. :)

hi,@maderix what your model input? i try rgb input but all output is zero, hope your relpy