app.py

from os import getenv
from pathlib import Path
from typing import Optional

import gradio as gr
import numpy as np
import onnxruntime as rt
from PIL import Image

from tagger.common import LabelData, load_labels, preprocess_image
from tagger.model import create_session

HF_TOKEN = getenv("HF_TOKEN", None)
WORK_DIR = Path.cwd().resolve()

MODEL_VARIANTS: dict[str, str] = {
    "MOAT": "SmilingWolf/wd-v1-4-moat-tagger-v2",
    "SwinV2": "SmilingWolf/wd-v1-4-swinv2-tagger-v2",
    "ConvNeXT": "SmilingWolf/wd-v1-4-convnext-tagger-v2",
    "ConvNeXTv2": "SmilingWolf/wd-v1-4-convnextv2-tagger-v2",
    "ViT": "SmilingWolf/wd-v1-4-vit-tagger-v2",
}

# allowed extensions
IMAGE_EXTENSIONS = [".jpg", ".jpeg", ".png", ".gif", ".webp", ".bmp", ".tiff", ".tif"]

# model input shape
IMAGE_SIZE = 448
example_images = sorted(
    [
        str(x.relative_to(WORK_DIR))
        for x in WORK_DIR.joinpath("examples").iterdir()
        if x.is_file() and x.suffix.lower() in IMAGE_EXTENSIONS
    ]
)
loaded_models: dict[str, Optional[rt.InferenceSession]] = {k: None for k, _ in MODEL_VARIANTS.items()}


def load_model(variant: str) -> rt.InferenceSession:
    global loaded_models

    # resolve the repo name
    model_repo = MODEL_VARIANTS.get(variant, None)
    if model_repo is None:
        raise ValueError(f"Unknown model variant: {variant}")

    if loaded_models.get(variant, None) is None:
        # save model to cache
        loaded_models[variant] = create_session(model_repo, token=HF_TOKEN)

    return loaded_models[variant]


def predict(
    image: Image.Image,
    variant: str,
    general_threshold: float = 0.35,
    character_threshold: float = 0.85,
):
    # Load model
    model: rt.InferenceSession = load_model(variant)
    # load labels
    labels: LabelData = load_labels()

    # get input size and name
    _, h, w, _ = model.get_inputs()[0].shape
    input_name = model.get_inputs()[0].name
    output_name = model.get_outputs()[0].name

    # preprocess image
    image = preprocess_image(image, (h, w))

    # turn into BGR24 numpy array of N,H,W,C since thats what these want
    inputs = image.convert("RGB").convert("BGR;24")
    inputs = np.array(inputs).astype(np.float32)
    inputs = np.expand_dims(inputs, axis=0)

    # Run the ONNX model
    probs = model.run([output_name], {input_name: inputs})

    # Convert indices+probs to labels
    probs = list(zip(labels.names, probs[0][0].astype(float)))

    # First 4 labels are actually ratings
    rating_labels = dict([probs[i] for i in labels.rating])

    # General labels, pick any where prediction confidence > threshold
    gen_labels = [probs[i] for i in labels.general]
    gen_labels = dict([x for x in gen_labels if x[1] > general_threshold])
    gen_labels = dict(sorted(gen_labels.items(), key=lambda item: item[1], reverse=True))

    # Character labels, pick any where prediction confidence > threshold
    char_labels = [probs[i] for i in labels.character]
    char_labels = dict([x for x in char_labels if x[1] > character_threshold])
    char_labels = dict(sorted(char_labels.items(), key=lambda item: item[1], reverse=True))

    # Combine general and character labels, sort by confidence
    combined_names = [x for x in gen_labels]
    combined_names.extend([x for x in char_labels])

    # Convert to a string suitable for use as a training caption
    caption = ", ".join(combined_names)
    booru = caption.replace("_", " ").replace("(", "\(").replace(")", "\)")

    return image, caption, booru, rating_labels, char_labels, gen_labels


with gr.Blocks(title="pi-chan's tagger") as demo:
    with gr.Row(equal_height=False):
        with gr.Column():
            img_input = gr.Image(
                label="Input",
                type="pil",
                image_mode="RGB",
                sources=["upload", "clipboard"],
            )
            variant = gr.Radio(choices=list(MODEL_VARIANTS.keys()), label="Model Variant", value="MOAT")
            gen_thresh = gr.Slider(0.0, 1.0, value=0.35, label="General Tag Threshold")
            char_thresh = gr.Slider(0.0, 1.0, value=0.85, label="Character Tag Threshold")
            show_processed = gr.Checkbox(label="Show Preprocessed", value=False)
            with gr.Row():
                submit = gr.Button(value="Submit", variant="primary", size="lg")
                clear = gr.ClearButton(
                    components=[],
                    variant="secondary",
                    size="lg",
                )
            with gr.Row():
                examples = gr.Examples(
                    examples=[
                        [imgpath, var, 0.35, 0.85]
                        for imgpath in example_images
                        for var in ["MOAT", "ConvNeXTv2"]
                    ],
                    inputs=[img_input, variant, gen_thresh, char_thresh],
                )
        with gr.Column():
            img_output = gr.Image(label="Preprocessed", type="pil", image_mode="RGB", scale=1, visible=False)
            with gr.Group():
                tags_string = gr.Textbox(
                    label="Caption", placeholder="Caption will appear here", show_copy_button=True
                )
                tags_booru = gr.Textbox(
                    label="Tags", placeholder="Tag string will appear here", show_copy_button=True
                )
            rating = gr.Label(label="Rating")
            character = gr.Label(label="Character")
            general = gr.Label(label="General")

    # tell clear button which components to clear
    clear.add([img_input, img_output, tags_string, rating, character, general])

    # show/hide processed image
    def on_select_show_processed(evt: gr.SelectData):
        return gr.update(visible=evt.selected)

    show_processed.select(on_select_show_processed, inputs=[], outputs=[img_output])

    submit.click(
        predict,
        inputs=[img_input, variant, gen_thresh, char_thresh],
        outputs=[img_output, tags_string, tags_booru, rating, character, general],
        api_name="predict",
    )

if __name__ == "__main__":
    demo.queue(max_size=10)
    demo.launch(server_name="0.0.0.0", server_port=7871)