Update tips.md

yawudede · Sep 17, 2018 · 5db1837 · 5db1837
1 parent b0e64d4
commit 5db1837
Showing 1 changed file with 2 additions and 0 deletions.
diff --git a/docs/tips.md b/docs/tips.md
@@ -4,3 +4,5 @@
 - Visualization: during training, the current results can be viewed using two methods. First, if you set `--display_id` > 0, the results and loss plot will appear on a local graphics web server launched by [visdom](https://github.com/facebookresearch/visdom). To do this, you should have `visdom` installed and a server running by the command `python -m visdom.server`. The default server URL is `http://localhost:8097`. `display_id` corresponds to the window ID that is displayed on the `visdom` server. The `visdom` display functionality is turned on by default. To avoid the extra overhead of communicating with `visdom` set `--display_id -1`. Second, the intermediate results are saved to `[opt.checkpoints_dir]/[opt.name]/web/` as an HTML file. To avoid this, set `--no_html`.
 - Preprocessing: images can be resized and cropped in different ways using `--resize_or_crop` option. The default option `'resize_and_crop'` resizes the image to be of size `(opt.loadSize, opt.loadSize)` and does a random crop of size `(opt.fineSize, opt.fineSize)`. `'crop'` skips the resizing step and only performs random cropping. `'scale_width'` resizes the image to have width `opt.fineSize` while keeping the aspect ratio. `'scale_width_and_crop'` first resizes the image to have width `opt.loadSize` and then does random cropping of size `(opt.fineSize, opt.fineSize)`. `'none'` tries to skip all these preprocessing steps. However, if the image size is not a multiple of some number depending on the number of downsamplings of the generator, you will get an error because the size of the output image may be different from the size of the input image. Therefore, `'none'` option still tries to adjust the image size to be a multiple of 4. You might need a bigger adjustment if you change the generator architecture. Please see `data/base_datset.py` do see how all these were implemented.
 - Fine-tuning/Resume training: to fine-tune a pre-trained model, or resume the previous training, use the `--continue_train` flag. The program will then load the model based on `epoch`. By default, the program will initialize the epoch count as 1. Set `--epoch_count <int>` to specify a different starting epoch count.
+- About image size: Because the generator architecture in CycleGAN involves a series of downsampling / upsampling operations, the size of the input and output image may not match if the input image size is not a multiple of 4. As a result, you may get a runtime error because the L1 identity loss cannot be enforced with images of different size. Therefore, we slightly resize the image to become multiples of 4 even with `--resize_or_crop none` option. For the same reason, `--fineSize` needs to be a multiple of 4. 
+- Training/Testing with high res images: CycleGAN is quite memory intensive because 4 networks (2 generators and 2 discriminators) need to be loaded on one GPU, so a large image cannot be entirely loaded. In this case, we recommend training with cropped images. For example, to generate 1024px results, you can train with `--resize_or_crop scale_width_and_crop --loadSize 1024 --fineSize 360`, and test with `--resize_or_crop scale_width --fineSize 1024`. This way makes sure the training and test will be at the same scale. At test time, you can afford higher resolution because you don’t need to load all networks.