Q: train datasets / which kind of images?

[tanreb]

hi, I have it all setup an running, (thanks for sharing the code) and now I'm trying to train new styles, and I got a question:

Which kind of images should we use to train a style?

example: If the intended usage is: style A hand-sketch + content B city landscapes,

• Should the training images be similar to A? (a lot of sketches and illustration images)
• or similar to B? (a lot of photos of cities and buildings)
• or to none of them? (like COCO, random stuff from the real world).

also, what is important when training a style?:

• the number of images in a dataset? min-max? (I tried with 20k images).
• the number of images in the validation? min-max? (I tried with 2k images).
• the number of iterations during train? min-max? (I tried with 1k, 5k, 10k, 20k, 40k iterations, but I don't see much difference).

style img: sketch Model trained with: COCO, 40k training iterations (default settings)

style img: sketch Model trained with: 20k images of cities, 20k training iterations

(training in a google cloud instance)

--- [edit] --- My goal is to achieve something like the "CurlyHair" filter from PrismaApp, and apply it to a folder. (I could do it with the slow_neural_style.lua without training, but I don't know how can I apply it to directories)

CurlyHair + Chigago

CurlyHair style image from PrismaApp

[HandsomeDevilv112]

My goal is to achieve something like the "CurlyHair" filter from PrismaApp, and apply it to a folder. (I could do it with the slow_neural_style.lua without training, but I don't know how can I apply it to directories)

To give you a band-aid until you get a better answer, you can save this bit of code as curlyhair.sh, and run it with "bash curlyhair.sh". It will run slow_neural_style.lua on all jpg in the folder with CurlyHair.png as the style image. I'm currently training up my first fast-neural-style model, so I'm afraid I'm not much help there. Additionally, just to mention it, this won't get you straight to the prismaapp output. Like, I have no idea what the app did to get rid of the clouds.

for i in *.jpg; do th slow_neural_style.lua -content_image /path/to/fast-neural-style/"$i" -style_image /path/to/directory/CurlyHair.png -output_image /path/to/output/directory/"${i%.*}.png"; done