Training questions

[ScriptingShogun]

Took me a while to get everything installed and working. Right now I am doing noise training with about 300 images. It is currently on run #53. Just curious if there is a set number of runs it goes through?

[nagadomi]

training step runs #200 epochs by default.

[an-electric-sheep]

Another question about training: does the batch_size make any quality difference or is it just for performance tuning?

[nagadomi]

In waifu2x, computing performance mostly not influenced by batch_size (but slow performance when batch_size=1).
I recommend small batch_size (2~4).
small batch_size: low memory usage. fast convergence. reducing overtting.

NOTICE: waifu2x is using Adam for CNN training (http://arxiv.org/abs/1412.6980)

[drdaxxy]

I have another question: how long did the entire training process take for your training set?

[nagadomi]

2x scaling model: 16 hours
Noise reduction model: 28 hours * 2
Total: around 72 hours
(on GTX760, 3000 images)

[c0d3rman]

The facts brought up here (how long training takes, batch size recommendation, epochs to training end) should be added to the documentation - they would be EXTREMELY useful to someone like me trying to train their own model.

[djdjoko]

Could you please let me know how many iterations did you train it for? Also did you use all the current default settings when training your anime model?

[nagadomi]

See https://github.com/nagadomi/waifu2x/blob/master/train.sh
The default setting is defined at https://github.com/nagadomi/waifu2x/blob/master/lib/settings.lua
I was used 3000 images for training.

And I'm currently running the traning script for the next version using dev branch code and 6000 images.

[djdjoko]

Thanks! Could you share what the next version is and what the differences are?

[nagadomi]

I released new version yesterday.
Summary of difference in training script:

• Add data compression method for on-memory image (reducing memory usage for dataset)
• Change loss function (MSE to Customized Huber Loss)
• Improve some image processing functions
• Add some data augmentation methods
• Tuning default parameters
• Improve code readability
• Improve training time (4x~ faster)

Sorry, my english skill is very poor. Please refer the code for more details.

[djdjoko]

I would like to ask you a hypothetical question. If I have only one specific type of image I want to scale up, would it be possible to train the algorithm only on one perfect image with maximum resolution?