feat: add webp decode support by adding tfio

[fumiama]

The method

tfio.experimental.color.rgba_to_rgb

is experimental and I'm not sure whether this will effect training or predicting. I have tested under the Pretrained Model and found just a little different between webp and png.

webp format

(0.999) 1girl
(0.575) areolae
(0.715) arm_up
(0.641) armpits
(0.802) blue_eyes
(0.653) blush
(0.982) breasts
(0.674) brown_hair
(0.515) collarbone
(0.546) double_bun
(0.648) eyebrows_visible_through_hair
(0.956) long_hair
(0.892) looking_at_viewer
(0.820) lying
(0.735) medium_breasts
(0.845) navel
(0.824) nude
(0.778) on_back
(0.682) parted_lips
(0.589) smile
(0.986) solo
(0.956) rating:questionable

png format

(0.999) 1girl
(0.585) areolae
(0.647) arm_up
(0.626) armpits
(0.681) blue_eyes
(0.655) blush
(0.984) breasts
(0.519) brown_hair
(0.602) collarbone
(0.713) double_bun
(0.731) eyebrows_visible_through_hair
(0.637) hair_bun
(0.967) long_hair
(0.892) looking_at_viewer
(0.855) lying
(0.743) medium_breasts
(0.795) navel
(0.843) nude
(0.838) on_back
(0.746) parted_lips
(0.519) pasties
(0.987) solo
(0.934) rating:questionable

[KichangKim]

Current training script uses only png and jpeg files. So I think it does not effect to training.

[fumiama]

OK, then I hope my modification would make it possible to evaluate custom webp images.

[DHG-Dav]

Hello, i'm just passing by and noticed you finally added webp compatibility, however you may want to also add the ability to evaluate folders containing webp and train on webp : main.py line 82 : @click.option('--folder-filters', default='.[Pp][Nn][Gg],.[Jj][Pp][Gg],.[Jj][Pp][Ee][Gg],.[Gg][Ii][Ff],*.[Ww][Ee][Bb][Pp]'

and dataset.py line 21 : cursor.execute( "SELECT md5, file_ext, tag_string FROM posts WHERE (file_ext = 'png' OR file_ext = 'jpg' OR file_ext = 'JPG' OR file_ext = 'PNG' OR file_ext = 'jpeg' OR file_ext = 'webp') AND (tag_count_general >= ?) ORDER BY id", (minimum_tag_count,), )

thanks again for that great project !

[fumiama]

Exactly. I only fixed the predict part because I had never used it to train a model. Maybe you can open a new PR for that.