Docker image with DeepDanbooru

ImoutoChan commented 1 year ago

Is there any docker image with some kind of HTTP API (like your demo)?

If there is none I would like to make one, but I have zero knowledge in Python development, maybe you have some tips? I also don't see the trained model, only instructions on how to make my own. Can I just download it from somewhere?

The way I see it:

Write some kind of HTTP wrapper
Download trained model from somewhere
Take base image with Python 3.7
Call pip install -r requirements.txt
Start the python app and forward some port from outside to the wrapper

Am I right?

Superfloh commented 1 year ago

The models are under releases. Technically that should work, but I remember that there were some issues. You will have to try and see if you can get it working I fear.

kamuridesu commented 1 year ago

I have created a Dockerfile with and without the model: https://github.com/kamuridesu/DeepDanbooru Tags are:

kamuri/deepdanbooru:nomodel for the image without a model
kamuri/deepdanbooru:latest for image with model

Although it does not have any API but isn't so hard, I may create one later.

Also, check https://github.com/kamuridesu/deepdanbooryu, a project using events and Docker to create a fully scalable backend if you're interested.

ImoutoChan commented 1 year ago

Thank you, I'm definitely waiting for a docker image with API.

If anyone else is interested in using DeepDanbooru through docker you can:

linux docker run -v /images:/app/images kamuri/deepdanbooru:latest evaluate /app/images/1.png --project-path /app/model

windows docker run -v //c/images:/app/images kamuri/deepdanbooru:latest evaluate /app/images/1.png --project-path /app/model

for the image in C:\images\1.png or in /images/1.png

kamuridesu commented 1 year ago

A simple Web API implementation for evaluate_image:

import os
import deepdanbooru
from io import BytesIO
from deepdanbooru.commands import evaluate_image
from flask import Flask, request, render_template

PROJECT_PATH = os.path.abspath("model")
print("Loading model")
MODEL = deepdanbooru.project.load_model_from_project(
    PROJECT_PATH, compile_model=False
)
print("Loading tags")
TAGS = deepdanbooru.project.load_tags_from_project(PROJECT_PATH)

app = Flask("deepdanbooru", template_folder=os.path.abspath("templates"))

@app.route("/")
def index():
    return render_template("index.html")

@app.route("/upload", methods=["POST"])
async def upload():
    if "file" not in request.files:
        return "No file uploaded!"

    file = request.files["file"]

    result = evaluate_image(BytesIO(file.stream.read()), MODEL, TAGS, 0.5)
    results = {}
    for tag, score in result:
        results[tag] = float(f"{score:05.3f}")
    return results

if __name__ == "__main__":
    app.run(debug=True, host="0.0.0.0", port="8080")

The model is stored in a subdir called model, placed in the same folder as the script, just extract the .zip in releases to the folder. It also has a templates/index.html file to test the upload:

<!DOCTYPE html>
<html>
<head>
    <title>File Upload</title>
</head>
<body>
    <h1>Upload a File</h1>
    <form action="/upload" method="POST" enctype="multipart/form-data">
        <input type="file" name="file" accept="*">
        <input type="submit" value="Upload">
    </form>
</body>
</html>

sky-cake commented 2 months ago

@kamuridesu I took your example and added it to a DDB scripts repo at https://github.com/sky-cake/ddb_scripts. The scripts also include an image crawler which saves tags and scores in a db.

KichangKim / DeepDanbooru

Docker image with DeepDanbooru #94