Jun-Yan Zhu

You don't need the FCN network to evaluate photos to labels. The label is produced by the trained generator. See [here](https://github.com/phillipi/pix2pix/issues/115) and [here](https://github.com/phillipi/pix2pix/issues/112) for more details.

We didn't have a complete script for CycleGAN project. We just modified the [evaluate.py](https://github.com/junyanz/pytorch-CycleGAN-and-pix2pix/blob/master/scripts/eval_cityscapes/evaluate.py#L48). You can replace the predicted labels with the generator's output. We included some functions in some...

I assume that you are using the Pytorch version with colab. For Lua, you can use `continue_train` [flag](https://github.com/junyanz/CycleGAN/blob/master/options.lua#L34). For PyTorch, please see this [Q & A](https://github.com/junyanz/pytorch-CycleGAN-and-pix2pix/blob/master/docs/qa.md#can-i-continueresume-my-training-350-275-234-87).

I would wait for more epochs. In our experiment, we trained the model for 200 epochs. (Each epoch has 1, 000 iterations). You also want to double-check if there is...

In the code, we first calculate the loss on the 30x30 map and then average the loss. @phillipi

I think you may want to use a smaller fine_size.

Yes. fineSize is the size of cropped images. Also, see more options for [resize_or_crop](https://github.com/junyanz/CycleGAN/blob/83077c9d40d729ca0dbb7ff029b90a507d32accd/options.lua#L51).

I haven't followed the latest update/version of Lua. If you have troubles, I recommend that you use our latest [PyTorch ](https://github.com/junyanz/pytorch-CycleGAN-and-pix2pix)version in Python.

The training time performance is often better than the test time performance. One thing you can do is to add these new images as part of the training.

I will recommend 1000+ images per dataset. But you may want to try your dataset and see how it goes.