A question about another program

[marvellousxf]

In the Deep Photo Enhancer , you have written this:

I've tried with @laol777's shared code. The model trained without replacing FLAGS['data_image_size'] to None also works. Hope for further communication. @liulizhou

I mean you only need to replace FLAGS['data_image_size'] to None in infer process

I wonder if you mean I don not need to retrain the model。 I can use the pair-training model（604） to make inference on arbitary resolution image?
What should I do with the model to do the inference?

[laol777]

Why do you ask in this project?
You cannot get from the pretrained model from this repo prediction for arbitrary image size. I shared snippet before for inference. I order to make a prediction for arbitrary resolution i rebuilt the graph, it can help you https://gist.github.com/m12sl/5fc45edede65ec3f007ebbe2baa7ced4

[marvellousxf]

Since the related questions in Deep photo enhancer have been closed and I cannot reply you there. I am doing my work on the basis of deep photo enhancer. I hope I can make use of the pair training model to deal with my 1920*1080 images while the model fixed the input to 512*512。 I just saw you said “it also works just modify FLAGS[‘data_image_size’] to None in inference stage”. So I’d like to comfirm how to do it and if I can make use the pair model provided by the autor since it is not easy for me as a new learner to retrain the model now, 发送自 Windows 10 版邮件应用 发件人: laol 发送时间: 2019年7月27日 16:15 收件人: xiaozhi2015/small-code 抄送: marvellousxf; Author 主题: Re: [xiaozhi2015/small-code] A question about another program (#2) Why do you ask in this project? You cannot get from the pretrained model from this repo prediction for arbitrary image size. I shared snippet before for inference. I order to make a prediction for arbitrary resolution i rebuilt the graph, it can help you https://gist.github.com/m12sl/5fc45edede65ec3f007ebbe2baa7ced4 — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub, or mute the thread.

[marvellousxf]

And really thank you for you help !! 发送自 Windows 10 版邮件应用 发件人: laol 发送时间: 2019年7月27日 16:15 收件人: xiaozhi2015/small-code 抄送: marvellousxf; Author 主题: Re: [xiaozhi2015/small-code] A question about another program (#2) Why do you ask in this project? You cannot get from the pretrained model from this repo prediction for arbitrary image size. I shared snippet before for inference. I order to make a prediction for arbitrary resolution i rebuilt the graph, it can help you https://gist.github.com/m12sl/5fc45edede65ec3f007ebbe2baa7ced4 — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub, or mute the thread.

[laol777]

Excuse me, I forgot details about the project. You don't need to make operations with graphs, which i gave above (it need only for tensorflow-serving), All others advice (nothinglo/Deep-Photo-Enhancer#51) are correct

[xiaozhi2015]

The author has already released his official implementation for inference with arbitrary resolution.

[marvellousxf]

He has released the unpair training model 999 for inference arbitary resolution. But I really need model 604 which is a U-net with global feature and trained by image pairs. I hope to know how to train the model for arbitaty resolution.