LambdaでDetectron2(物体検出)して、実行速度はかってみた。

概要

• モデル：resnet50 + frcnn
• lambdaのメモリサイズ：8192MB

結論先に書くと、
コールドスタートありで9秒ほど。
コールドスタートなしでなんと3秒！

簡単にapi作るにはいいかもですね。

(メモ書きです。)

ファイル郡

FROM nvidia/cuda:10.1-cudnn7-devel

ENV DEBIAN_FRONTEND noninteractive
RUN apt-get update && apt-get install -y \
	python3-opencv ca-certificates python3-dev git wget sudo  \
	cmake ninja-build && \
  rm -rf /var/lib/apt/lists/*
RUN ln -sv /usr/bin/python3 /usr/bin/python

WORKDIR /home

ENV PATH="/home/.local/bin:${PATH}"
RUN wget https://bootstrap.pypa.io/get-pip.py && \
	python3 get-pip.py && \
	rm get-pip.py

# install dependencies
# See https://pytorch.org/ for other options if you use a different version of CUDA
RUN pip install tensorboard
RUN pip install torch==1.7 torchvision==0.8.1 -f https://download.pytorch.org/whl/cu101/torch_stable.html

RUN pip install 'git+https://github.com/facebookresearch/fvcore'
# install detectron2
RUN git clone https://github.com/facebookresearch/detectron2 detectron2_repo
# set FORCE_CUDA because during `docker build` cuda is not accessible
ENV FORCE_CUDA="0"
# This will by default build detectron2 for all common cuda architectures and take a lot more time,
# because inside `docker build`, there is no way to tell which architecture will be used.
ARG TORCH_CUDA_ARCH_LIST="Kepler;Kepler+Tesla;Maxwell;Maxwell+Tegra;Pascal;Volta;Turing"
ENV TORCH_CUDA_ARCH_LIST="${TORCH_CUDA_ARCH_LIST}"

RUN pip install -e detectron2_repo

# aws lambda run interface clientを iinstallする
RUN pip install awslambdaric

# opetion aws-rieを入れる。
ADD https://github.com/aws/aws-lambda-runtime-interface-emulator/releases/latest/download/aws-lambda-rie /usr/bin/aws-lambda-rie
RUN chmod 755 /usr/bin/aws-lambda-rie

# Set a fixed model cache directory.
ENV FVCORE_CACHE="/tmp"
ARG APP_DIR="/home/app/"
WORKDIR ${APP_DIR}
COPY app ${APP_DIR}

ENTRYPOINT [ "/bin/bash", "entry.sh" ]
CMD [ "icon_predict.main" ]

• entry.sh

#!/bin/sh
if [ -z "${AWS_LAMBDA_RUNTIME_API}" ]; then
    exec /usr/bin/aws-lambda-rie /usr/bin/python -m awslambdaric $1
    # exec /usr/bin/aws-lambda-rie  /usr/bin/python3 -m awslambdaric 
else
    exec /usr/bin/python -m awslambdaric $1
fi

• icon_predict.py

import os
import sys
import numpy as np
import cv2
import torch
import json
import boto3
from torchvision import transforms, utils

from detectron2 import model_zoo
from detectron2.engine import DefaultPredictor
from detectron2.config import get_cfg
from detectron2.utils.visualizer import Visualizer, ColorMode
from detectron2.data import MetadataCatalog, DatasetCatalog
from detectron2.data.datasets import register_coco_instances

# modelのダウンロード
S3_BUCKET = [bucket名]
s3 = boto3.resource('s3')
model_path = [modelのpath名]
s3_key = [modelのs3のkey名]
if not os.path.exists(model_path):
    s3.Bucket(S3_BUCKET).download_file(s3_key, model_path)

def gen_predictor(model_path):
    """
        事前にモデルを設定する。
    """
    device = os.getenv("DEVICE", "cpu")
    cfg = get_cfg()
    cfg.MODEL.DEVICE = device
    cfg.merge_from_file(model_zoo.get_config_file("PascalVOC-Detection/faster_rcnn_R_50_FPN.yaml"))

    cfg.MODEL.WEIGHTS = model_path # downloadで落としてきたモデル。
    cfg.SOLVER.IMS_PER_BATCH = 1
    cfg.MODEL.ROI_HEADS.BATCH_SIZE_PER_IMAGE = 128  # faster, and good enough for this toy dataset (default: 512)
    cfg.MODEL.ROI_HEADS.NUM_CLASSES = 6  # only has one class (ballon)
    cfg.MODEL.ROI_HEADS.NMS = 0.2
    cfg.MODEL.ROI_HEADS.SCORE_THRESH_TEST = 0.8   # set the testing threshold for this model
    predictor = DefaultPredictor(cfg)
    return predictor

def main():
    img_path = "image.png"

    img = cv2.imread(img_path)
    predictor = gen_predictor(model_path)
    bigicon_predictions = predictor(img)["instances"].to("cpu")
    boxes = bigicon_predictions.pred_boxes.__dict__["tensor"].tolist()
    scores = bigicon_predictions.scores.tolist()
    print("boxes :", boxes)
    print("scores :", scores)
    print("-- Finished!! --")
    return boxes

def handler(event, context):
    return main()

if __name__ == "__main__":
    main()

実行結果

テストすると3秒くらい。

• モデルをひとまずAPIにしたい。
• サーバレスで運用したい
• SageMakermめんどくさい...
• コストそんな高くしたくない...

って人にはそこそこ使える性能じゃないだろうか。