概要

PyTorchのモデルをFlaskでデプロイします。画像をBase64エンコードで送信する方式を取ります。

imagenet_class_index.jsonはこちらからダウンロードしました。
sample.jpgは分類を試す画像で、なんでもいいです。

コード

フォルダ構造

.
├── Dockerfile
├── client.py
├── docker-compose.yml
├── imagenet_class_index.json
├── main.py
├── model.py
├── requirements.txt
└── sample.jpg

requirements.txt

numpy
pillow
flask
flask_cors
requests
torch
torchvision
gunicorn

サーバーのコード

main.py

import base64
import json
from io import BytesIO


from flask import Flask, jsonify, request
from flask_cors import CORS
from PIL import Image


from model import Predictor

app = Flask(__name__)
CORS(app)

predictor = Predictor('imagenet_class_index.json')


@app.route("/", methods=["POST"])
def predict():
    data = request.data.decode('utf-8')
    data = json.loads(data)

    img_stream = base64.b64decode(data['image'])
    img_pil = Image.open(BytesIO(img_stream))
    idx, label = predictor.predict(img_pil)

    response = {'result': label}
    return jsonify(response)


if __name__ == "__main__":
    app.run(host='0.0.0.0', port=5000, debug=True)

model.py

import json
import shutil

import numpy as np
import requests
import torch
import torch.nn.functional as F
from torchvision import models, transforms


def get_weight():
    url = 'https://s3xxxxxxxxxxxxxx'
    try:
        res = requests.get(url)
        with open('./weight.pth', mode='wb') as f:
            shutil.copyfileobj(res.raw, f)
    except Exception as e:
        raise RuntimeError(e)


def create_net(param_file=None, device="cpu"):
    net = models.resnet34(pretrained=True)
    if param_file:
        net.load_state_dict(torch.load(
            param_file,
            map_location=torch.device(device))
        )
    net.to(device)
    return net


def load_label(label_path):
    with open(label_path) as f:
        labels = json.load(f)
    return labels


class Predictor():
    def __init__(self, label_path, param_file=None, device="cpu"):
        self.net = create_net(param_file, device="cpu")
        self.net.eval()
        self.trans = transforms.Compose([
            transforms.Resize((224, 224)),
            transforms.ToTensor()
        ])
        self.classes = load_label(label_path)

    @torch.no_grad()
    def predict(self, img_pil):
        """
        Parameters
        ----------
        img_pil : PIL.Image.Image
        """
        img_tensor = self.trans(img_pil)
        img_tensor = torch.unsqueeze(img_tensor, 0)
        y_pred = self.net(img_tensor)
        y_pred = F.softmax(y_pred, dim=1).detach().numpy()

        y_idx = np.argmax(y_pred)

        return y_idx, self.classes[str(y_idx)][1]

Dockerfile

FROM python:3.9

ENV PYTHONUNBUFFERED 1

WORKDIR /usr/src/app

COPY requirements.txt /usr/src/app/

RUN pip install -r requirements.txt

COPY . /usr/src/app/

CMD ["gunicorn", "main:app", "-b", "0.0.0.0:5000", "--access-logfile", "-", "-w", "2", "--preload"]

docker-compose.yml

version: '2.3'
services:
  worker:
    build:
      context: ./
    container_name: flask_app
    ports:
    - "5000:5000"
    working_dir: /usr/src/app/

クライアントのコード

client.py

import base64
import os

import requests

os.environ['no_proxy'] = '*'
# Base64でエンコードする画像のパス
target_file = "sample.jpg"


with open(target_file, 'rb') as f:
    data = f.read()

# Base64で画像をエンコード
encoded = base64.b64encode(data).decode('utf-8')

url = 'http://localhost:5000'
payload = {'image': encoded}

res = requests.post(url, json=payload)
print(res.text)

実行

docker compose upでサーバーのコンテナ起動。
python client.pyを実行して画像を送信、分類結果を受信。