How to set the parameters in *.yml when training a new model

[neal668]

`## Basic configuration
style_image: img/snow.jpg # targeted style image
naming: "snow" # the name of this model. Determine the path to save checkpoint and events file.
model_path: models # root path to save checkpoint and events file. The final path would be <model_path>/

Weight of the loss

content_weight: 1.0 # weight for content features loss
style_weight: 200.0 # weight for style features loss
tv_weight: 0.0 # weight for total variation loss

The size, the iter number to run

image_size: 256
batch_size: 4
epoch: 2`

Hi, this is a yml file I want to set for training, I noticed the parameter of style_weight is different in those .yml file you proposed while others are the same. Any suggestions for parameter setting? Thanks a lot

[hzy46]

style_weight is a hyperparameter for the model. So you can adjust it by observing the generated image. If the generated image is full of patterns in the "style_image", and you can hardly find meaningful contents in it, decrease style_weight. If the content is very clear and without any style, increase style_weight.

[neal668]

Thank you for your reply. :)
By the way, would you mind introducing a suggested the number of epoch when training a new model? I tried from 6000 to 11000, the effect is different while quite small.
Another question is do I need to select those "simple" figure as the style one?
Thank you for your time.

[hzy46]

Epoch is usually 1~2. Maybe you mean 6000~11000 steps?
If you use larger learning rate, you can use less steps. Use the tensorboard to decide whether to stop training.