Jun-Yan Zhu

Could you share with us more details? For example, are you trying to improve pix2pix or CycleGAN? Which aspects of the model are you trying to improve? Regarding the idea...

@improgress any details?

See [training & test tips](https://github.com/junyanz/pytorch-CycleGAN-and-pix2pix/blob/master/docs/tips.md#trainingtesting-with-high-res-images) for more details.

They are training set images. You need to run `test.py` for test image results.

You can use `--preprocess none`. See [Training & Test Tips](https://github.com/junyanz/pytorch-CycleGAN-and-pix2pix/blob/master/docs/tips.md#trainingtesting-with-rectangular-images) for more details.

You probably need to proprocess your data before feeding them to CycleGAN. I recommend that you (1) apply portrait segmentation (e.g., this [one](https://github.com/dongwu92/AutoPortraitMatting)) to remove the background and (2) use...

You can also change this [line](https://github.com/junyanz/pytorch-CycleGAN-and-pix2pix/blob/master/test.py#L79) in the `test.py`.

#1018 might be relevant to your issues. Would it be possible to share with us your python command?

It could be caused by several reasons. (1) The monet2photo models are trained on landscape images. It may not work well for other types of inputs. (2) the scale of...

You should be able to call ``scheduler.get_last_lr()`` within the update_learning_rate() [function](https://github.com/junyanz/pytorch-CycleGAN-and-pix2pix/blob/f13aab8148bd5f15b9eb47b690496df8dadbab0c/models/base_model.py#L119).