An image tagger base using the onnxruntime. Because of extensibility, you will also need to interact with onnxruntime_go.
A good and recommended model to use is SmilingWolf/wd-vit-large-tagger-v3, this is an example code:
package main
import (
"cmp"
"fmt"
"image"
_ "image/png"
"maps"
"os"
"slices"
"strings"
"github.com/milosworks/gotagger"
ort "github.com/yalue/onnxruntime_go"
)
const (
modelPath = "./model.onnx"
tagsPath = "./selected_tags.csv"
imagePath = "./image.png"
// Example for windows
// Runtimes are not provided since onnxruntime_go *sometimes* handles it
// it is still recommended to specify the runtime as onnxruntime_go could have outdated runtimes
runtimePath = "./onnxruntime.dll"
)
func main() {
// You need to specify the shared library path to ort specifically, not gotagger
ort.SetSharedLibraryPath(runtimePath)
// VERY IMPORTANT: Initialize ort environment
err := ort.InitializeEnvironment()
if err != nil {
panic(err)
}
// Don't forget to destroy it at the end
defer ort.DestroyEnvironment()
// Create a new gotagger session
session, err := gotagger.New(modelPath, tagsPath)
if err != nil {
panic(err)
}
// Open example file
file, err := os.Open(imagePath)
if err != nil {
panic(err)
}
defer file.Close()
// Decode it into the image.Image
img, _, err := image.Decode(file)
if err != nil {
panic(err)
}
// Run the session with the provided images
// The library is smart enough to do images in batches even when the model has a batchsize of 1.
// You can provide the GeneralThreshold and CharacterThreshold for their respective tags.
// Also enable mcut on general or character tags
predictions, err := session.Run(
[]image.Image{img},
gotagger.DefaultGeneralThreshold,
gotagger.DefaultCharacterThreshold,
false,
false,
)
if err != nil {
panic(err)
}
println("Predictions:")
for i, pred := range predictions {
fmt.Printf("Image [%d]: %s\n", i+1, strings.Join(pred.Names(), ", "))
if len(pred.Character) != 0 {
chars := slices.Collect(maps.Keys(pred.Character))
slices.SortFunc(chars, func(a, b string) int {
return cmp.Compare(pred.Character[b], pred.Character[a])
})
fmt.Printf("Character: %v\n", chars[0])
}
ratings := slices.Collect(maps.Keys(pred.Rating))
slices.SortFunc(ratings, func(a, b string) int {
return cmp.Compare(pred.Rating[b], pred.Rating[a])
})
fmt.Printf("Rating: %v\n", ratings[0])
}
}Output predictions:
Predictions:
Image [1]: 1boy, blonde hair, male focus, red eyes, earrings, jewelry, short hair, gauntlets, solo, single gauntlet, turban, open mouth, tattoo, looking at viewer, sitting, throne, horns, arabian clothes, hair between eyes, own hands together, upper body
Character: gilgamesh (caster) (fate)
Rating: sensitive
