TensorRT

NVIDIA Jetson シリーズはディープ・ニューラルネットワークの学習済みモデルを使って推論を高速に行うために利用するのが一般的です。ディープ・ニューラルネットワークの学習には通常ディープラーニング・フレームワーク ¹ と呼ばれるソフトウェアを使います。ディープラーニング・フレームワークはもちろん推論にも使えますが、推論用には もっと効率的な仕組み があります。NVIDIA TensorRT はその一つで NVIDIA GPU で推論処理を行う場合、最良な選択肢だと思います。Jetson の他にもあらゆる NVIDIA GPU に対応しています。Jetson Nano 開発者キット向けの OS イメージ・ファイルに TensorRT は含まれていますので TensorRT を新たにインストールする必要はありません。
この TensorRT を使うにはディープラーニング・フレームワークで生成した学習済みモデルを TensorRT 専用のデータ形式に変換する必要がありますが、TensorRT にその機能が用意されていますのでそれ程難しくありません。

ONNX

ディープ・ニューラルネットワークの学習済みモデル ² の形式はディープラーニング・フレームワーク毎に異なるのが普通でした。しかし、最近ではその共通形式が提唱されています。ONNX はその代表格です。この ONNX 形式に変換してしまえばディープラーニング・フレームワーク間でデータの交換が可能になります。TensorRT もこの ONNX 形式の学習済みモデルをインポートすることができます。

YOLO v3 サンプル・プログラム

TensorRT は C++ と Python の API を提供しています。サンプル・プログラムが /usr/src/tensorrt/samples ディレクトリにありますのでこれと TensorRT Developer Guide で TensorRT の使い方を知ることができます。
ここでは物体検出用ディープ・ニューラルネットワークとしてとても有名な YOLO (You Only Look Once) ³ のサンプル・プログラムを動作させてみましょう。

このプログラムは以下のとおり2ステップに分かれています。

YOLO 形式のモデルをダウンロードして ONNX 形式に変換する。
ONNX 形式のモデルを TensorRT にロードして入力として与えたファイル内の画像に対して物体検出を実行する。

第1ステップのコードは Python2 でのみ動作するので注意してください。第2ステップのコードは Python2 でも Python3 でも動作します。第1ステップは最初に一回だけ実行すれば良く、推論の度に再度実行する必要はありません。

CUDA ツールキットのパスを通す

念のため、CUDA のバージョン名を含まないパス名で参照できるように以下のとおりシンボリック・リンクを作成します。

$ sudo ln -s /usr/local/cuda-10.0 /usr/local/cuda

以下の行を ~/.bashrc ファイルに加えます。

export PATH=$PATH:/usr/local/cuda/bin

上記のパス設定を反映させます。

$ source ~/.bashrc

パスが通ったかテストしてみましょう。

$ nvcc --version
nvcc: NVIDIA (R) Cuda compiler driver
Copyright (c) 2005-2018 NVIDIA Corporation
Built on Sun_Sep_30_21:09:22_CDT_2018
Cuda compilation tools, release 10.0, V10.0.166

必要なライブラリ／モジュールのインストール

$ sudo apt-get update

$ sudo apt-get install python-pip protobuf-compiler libprotoc-dev libjpeg-dev cmake

$ cd /usr/src/tensorrt

$ sudo chmod -R a+rw samples

$ cd samples/python/yolov3_onnx/

今後、TensorRT のバージョンアップに従い状況は異なると思いますが、最新版の ONNX モジュールを使うとサンプル・プログラム実行中にエラーが発生しました。/usr/src/tensorrt/samples/python/yolov3_onnx/requirements.txt を以下のように変更して ONNX モジュールのバージョンを指定します。

numpy>=1.15.1
onnx==1.4.1
pycuda>=2017.1.1
Pillow>=5.2.0
wget>=3.2

以下のコマンドでモジュールをインストールします。

$ python2 -m pip install -r requirements.txt

スワップ領域の割り当て

YOLO 形式モデルを ONNX 形式モデルに変換する第1ステップの処理はメモリを多量に必要とするようです。以下のとおりスワップ領域を割り当ててこの問題を乗り切ります。⁴

$ sudo fallocate -l 4G /var/swapfile
$ sudo chmod 600 /var/swapfile
$ sudo mkswap /var/swapfile
$ sudo swapon /var/swapfile
$ sudo bash -c 'echo "/var/swapfile swap swap defaults 0 0" >> /etc/fstab'

YOLO モデルのダウンロードと ONNX 形式への変換

$ python2 yolov3_to_onnx.py

結果

$ python2 yolov3_to_onnx.py
Downloading from https://raw.githubusercontent.com/pjreddie/darknet/f86901f6177dfc6116360a13cc06ab680e0c86b0/cfg/yolov3.cfg, this may take a while...
100% [............................................................] 8342 / 8342
Downloading from https://pjreddie.com/media/files/yolov3.weights, this may take a while...
100% [..................................................] 248007048 / 248007048
Layer of type yolo not supported, skipping ONNX node generation.
Layer of type yolo not supported, skipping ONNX node generation.
Layer of type yolo not supported, skipping ONNX node generation.
graph YOLOv3-608 (
  %000_net[FLOAT, 64x3x608x608]
) initializers (
  %001_convolutional_bn_scale[FLOAT, 32]
  %001_convolutional_bn_bias[FLOAT, 32]
  %001_convolutional_bn_mean[FLOAT, 32]
  %001_convolutional_bn_var[FLOAT, 32]
  %001_convolutional_conv_weights[FLOAT, 32x3x3x3]
  %002_convolutional_bn_scale[FLOAT, 64]
  %002_convolutional_bn_bias[FLOAT, 64]
  %002_convolutional_bn_mean[FLOAT, 64]
  %002_convolutional_bn_var[FLOAT, 64]
  %002_convolutional_conv_weights[FLOAT, 64x32x3x3]
  %003_convolutional_bn_scale[FLOAT, 32]
  %003_convolutional_bn_bias[FLOAT, 32]
  %003_convolutional_bn_mean[FLOAT, 32]
  %003_convolutional_bn_var[FLOAT, 32]
  %003_convolutional_conv_weights[FLOAT, 32x64x1x1]
  %004_convolutional_bn_scale[FLOAT, 64]
  %004_convolutional_bn_bias[FLOAT, 64]
  %004_convolutional_bn_mean[FLOAT, 64]
  %004_convolutional_bn_var[FLOAT, 64]
  %004_convolutional_conv_weights[FLOAT, 64x32x3x3]
  %006_convolutional_bn_scale[FLOAT, 128]
  %006_convolutional_bn_bias[FLOAT, 128]
  %006_convolutional_bn_mean[FLOAT, 128]
  %006_convolutional_bn_var[FLOAT, 128]
  %006_convolutional_conv_weights[FLOAT, 128x64x3x3]
  %007_convolutional_bn_scale[FLOAT, 64]
  %007_convolutional_bn_bias[FLOAT, 64]
  %007_convolutional_bn_mean[FLOAT, 64]
  %007_convolutional_bn_var[FLOAT, 64]
  %007_convolutional_conv_weights[FLOAT, 64x128x1x1]
  %008_convolutional_bn_scale[FLOAT, 128]
  %008_convolutional_bn_bias[FLOAT, 128]
  %008_convolutional_bn_mean[FLOAT, 128]
  %008_convolutional_bn_var[FLOAT, 128]
  %008_convolutional_conv_weights[FLOAT, 128x64x3x3]
  %010_convolutional_bn_scale[FLOAT, 64]
  %010_convolutional_bn_bias[FLOAT, 64]
  %010_convolutional_bn_mean[FLOAT, 64]
  %010_convolutional_bn_var[FLOAT, 64]
  %010_convolutional_conv_weights[FLOAT, 64x128x1x1]
  %011_convolutional_bn_scale[FLOAT, 128]
  %011_convolutional_bn_bias[FLOAT, 128]
  %011_convolutional_bn_mean[FLOAT, 128]
  %011_convolutional_bn_var[FLOAT, 128]
  %011_convolutional_conv_weights[FLOAT, 128x64x3x3]
  %013_convolutional_bn_scale[FLOAT, 256]
  %013_convolutional_bn_bias[FLOAT, 256]
  %013_convolutional_bn_mean[FLOAT, 256]
  %013_convolutional_bn_var[FLOAT, 256]
  %013_convolutional_conv_weights[FLOAT, 256x128x3x3]
  %014_convolutional_bn_scale[FLOAT, 128]
  %014_convolutional_bn_bias[FLOAT, 128]
  %014_convolutional_bn_mean[FLOAT, 128]
  %014_convolutional_bn_var[FLOAT, 128]
  %014_convolutional_conv_weights[FLOAT, 128x256x1x1]
  %015_convolutional_bn_scale[FLOAT, 256]
  %015_convolutional_bn_bias[FLOAT, 256]
  %015_convolutional_bn_mean[FLOAT, 256]
  %015_convolutional_bn_var[FLOAT, 256]
  %015_convolutional_conv_weights[FLOAT, 256x128x3x3]
  %017_convolutional_bn_scale[FLOAT, 128]
  %017_convolutional_bn_bias[FLOAT, 128]
  %017_convolutional_bn_mean[FLOAT, 128]
  %017_convolutional_bn_var[FLOAT, 128]
  %017_convolutional_conv_weights[FLOAT, 128x256x1x1]
  %018_convolutional_bn_scale[FLOAT, 256]
  %018_convolutional_bn_bias[FLOAT, 256]
  %018_convolutional_bn_mean[FLOAT, 256]
  %018_convolutional_bn_var[FLOAT, 256]
  %018_convolutional_conv_weights[FLOAT, 256x128x3x3]
  %020_convolutional_bn_scale[FLOAT, 128]
  %020_convolutional_bn_bias[FLOAT, 128]
  %020_convolutional_bn_mean[FLOAT, 128]
  %020_convolutional_bn_var[FLOAT, 128]
  %020_convolutional_conv_weights[FLOAT, 128x256x1x1]
  %021_convolutional_bn_scale[FLOAT, 256]
  %021_convolutional_bn_bias[FLOAT, 256]
  %021_convolutional_bn_mean[FLOAT, 256]
  %021_convolutional_bn_var[FLOAT, 256]
  %021_convolutional_conv_weights[FLOAT, 256x128x3x3]
  %023_convolutional_bn_scale[FLOAT, 128]
  %023_convolutional_bn_bias[FLOAT, 128]
  %023_convolutional_bn_mean[FLOAT, 128]
  %023_convolutional_bn_var[FLOAT, 128]
  %023_convolutional_conv_weights[FLOAT, 128x256x1x1]
  %024_convolutional_bn_scale[FLOAT, 256]
  %024_convolutional_bn_bias[FLOAT, 256]
  %024_convolutional_bn_mean[FLOAT, 256]
  %024_convolutional_bn_var[FLOAT, 256]
  %024_convolutional_conv_weights[FLOAT, 256x128x3x3]
  %026_convolutional_bn_scale[FLOAT, 128]
  %026_convolutional_bn_bias[FLOAT, 128]
  %026_convolutional_bn_mean[FLOAT, 128]
  %026_convolutional_bn_var[FLOAT, 128]
  %026_convolutional_conv_weights[FLOAT, 128x256x1x1]
  %027_convolutional_bn_scale[FLOAT, 256]
  %027_convolutional_bn_bias[FLOAT, 256]
  %027_convolutional_bn_mean[FLOAT, 256]
  %027_convolutional_bn_var[FLOAT, 256]
  %027_convolutional_conv_weights[FLOAT, 256x128x3x3]
  %029_convolutional_bn_scale[FLOAT, 128]
  %029_convolutional_bn_bias[FLOAT, 128]
  %029_convolutional_bn_mean[FLOAT, 128]
  %029_convolutional_bn_var[FLOAT, 128]
  %029_convolutional_conv_weights[FLOAT, 128x256x1x1]
  %030_convolutional_bn_scale[FLOAT, 256]
  %030_convolutional_bn_bias[FLOAT, 256]
  %030_convolutional_bn_mean[FLOAT, 256]
  %030_convolutional_bn_var[FLOAT, 256]
  %030_convolutional_conv_weights[FLOAT, 256x128x3x3]
  %032_convolutional_bn_scale[FLOAT, 128]
  %032_convolutional_bn_bias[FLOAT, 128]
  %032_convolutional_bn_mean[FLOAT, 128]
  %032_convolutional_bn_var[FLOAT, 128]
  %032_convolutional_conv_weights[FLOAT, 128x256x1x1]
  %033_convolutional_bn_scale[FLOAT, 256]
  %033_convolutional_bn_bias[FLOAT, 256]
  %033_convolutional_bn_mean[FLOAT, 256]
  %033_convolutional_bn_var[FLOAT, 256]
  %033_convolutional_conv_weights[FLOAT, 256x128x3x3]
  %035_convolutional_bn_scale[FLOAT, 128]
  %035_convolutional_bn_bias[FLOAT, 128]
  %035_convolutional_bn_mean[FLOAT, 128]
  %035_convolutional_bn_var[FLOAT, 128]
  %035_convolutional_conv_weights[FLOAT, 128x256x1x1]
  %036_convolutional_bn_scale[FLOAT, 256]
  %036_convolutional_bn_bias[FLOAT, 256]
  %036_convolutional_bn_mean[FLOAT, 256]
  %036_convolutional_bn_var[FLOAT, 256]
  %036_convolutional_conv_weights[FLOAT, 256x128x3x3]
  %038_convolutional_bn_scale[FLOAT, 512]
  %038_convolutional_bn_bias[FLOAT, 512]
  %038_convolutional_bn_mean[FLOAT, 512]
  %038_convolutional_bn_var[FLOAT, 512]
  %038_convolutional_conv_weights[FLOAT, 512x256x3x3]
  %039_convolutional_bn_scale[FLOAT, 256]
  %039_convolutional_bn_bias[FLOAT, 256]
  %039_convolutional_bn_mean[FLOAT, 256]
  %039_convolutional_bn_var[FLOAT, 256]
  %039_convolutional_conv_weights[FLOAT, 256x512x1x1]
  %040_convolutional_bn_scale[FLOAT, 512]
  %040_convolutional_bn_bias[FLOAT, 512]
  %040_convolutional_bn_mean[FLOAT, 512]
  %040_convolutional_bn_var[FLOAT, 512]
  %040_convolutional_conv_weights[FLOAT, 512x256x3x3]
  %042_convolutional_bn_scale[FLOAT, 256]
  %042_convolutional_bn_bias[FLOAT, 256]
  %042_convolutional_bn_mean[FLOAT, 256]
  %042_convolutional_bn_var[FLOAT, 256]
  %042_convolutional_conv_weights[FLOAT, 256x512x1x1]
  %043_convolutional_bn_scale[FLOAT, 512]
  %043_convolutional_bn_bias[FLOAT, 512]
  %043_convolutional_bn_mean[FLOAT, 512]
  %043_convolutional_bn_var[FLOAT, 512]
  %043_convolutional_conv_weights[FLOAT, 512x256x3x3]
  %045_convolutional_bn_scale[FLOAT, 256]
  %045_convolutional_bn_bias[FLOAT, 256]
  %045_convolutional_bn_mean[FLOAT, 256]
  %045_convolutional_bn_var[FLOAT, 256]
  %045_convolutional_conv_weights[FLOAT, 256x512x1x1]
  %046_convolutional_bn_scale[FLOAT, 512]
  %046_convolutional_bn_bias[FLOAT, 512]
  %046_convolutional_bn_mean[FLOAT, 512]
  %046_convolutional_bn_var[FLOAT, 512]
  %046_convolutional_conv_weights[FLOAT, 512x256x3x3]
  %048_convolutional_bn_scale[FLOAT, 256]
  %048_convolutional_bn_bias[FLOAT, 256]
  %048_convolutional_bn_mean[FLOAT, 256]
  %048_convolutional_bn_var[FLOAT, 256]
  %048_convolutional_conv_weights[FLOAT, 256x512x1x1]
  %049_convolutional_bn_scale[FLOAT, 512]
  %049_convolutional_bn_bias[FLOAT, 512]
  %049_convolutional_bn_mean[FLOAT, 512]
  %049_convolutional_bn_var[FLOAT, 512]
  %049_convolutional_conv_weights[FLOAT, 512x256x3x3]
  %051_convolutional_bn_scale[FLOAT, 256]
  %051_convolutional_bn_bias[FLOAT, 256]
  %051_convolutional_bn_mean[FLOAT, 256]
  %051_convolutional_bn_var[FLOAT, 256]
  %051_convolutional_conv_weights[FLOAT, 256x512x1x1]
  %052_convolutional_bn_scale[FLOAT, 512]
  %052_convolutional_bn_bias[FLOAT, 512]
  %052_convolutional_bn_mean[FLOAT, 512]
  %052_convolutional_bn_var[FLOAT, 512]
  %052_convolutional_conv_weights[FLOAT, 512x256x3x3]
  %054_convolutional_bn_scale[FLOAT, 256]
  %054_convolutional_bn_bias[FLOAT, 256]
  %054_convolutional_bn_mean[FLOAT, 256]
  %054_convolutional_bn_var[FLOAT, 256]
  %054_convolutional_conv_weights[FLOAT, 256x512x1x1]
  %055_convolutional_bn_scale[FLOAT, 512]
  %055_convolutional_bn_bias[FLOAT, 512]
  %055_convolutional_bn_mean[FLOAT, 512]
  %055_convolutional_bn_var[FLOAT, 512]
  %055_convolutional_conv_weights[FLOAT, 512x256x3x3]
  %057_convolutional_bn_scale[FLOAT, 256]
  %057_convolutional_bn_bias[FLOAT, 256]
  %057_convolutional_bn_mean[FLOAT, 256]
  %057_convolutional_bn_var[FLOAT, 256]
  %057_convolutional_conv_weights[FLOAT, 256x512x1x1]
  %058_convolutional_bn_scale[FLOAT, 512]
  %058_convolutional_bn_bias[FLOAT, 512]
  %058_convolutional_bn_mean[FLOAT, 512]
  %058_convolutional_bn_var[FLOAT, 512]
  %058_convolutional_conv_weights[FLOAT, 512x256x3x3]
  %060_convolutional_bn_scale[FLOAT, 256]
  %060_convolutional_bn_bias[FLOAT, 256]
  %060_convolutional_bn_mean[FLOAT, 256]
  %060_convolutional_bn_var[FLOAT, 256]
  %060_convolutional_conv_weights[FLOAT, 256x512x1x1]
  %061_convolutional_bn_scale[FLOAT, 512]
  %061_convolutional_bn_bias[FLOAT, 512]
  %061_convolutional_bn_mean[FLOAT, 512]
  %061_convolutional_bn_var[FLOAT, 512]
  %061_convolutional_conv_weights[FLOAT, 512x256x3x3]
  %063_convolutional_bn_scale[FLOAT, 1024]
  %063_convolutional_bn_bias[FLOAT, 1024]
  %063_convolutional_bn_mean[FLOAT, 1024]
  %063_convolutional_bn_var[FLOAT, 1024]
  %063_convolutional_conv_weights[FLOAT, 1024x512x3x3]
  %064_convolutional_bn_scale[FLOAT, 512]
  %064_convolutional_bn_bias[FLOAT, 512]
  %064_convolutional_bn_mean[FLOAT, 512]
  %064_convolutional_bn_var[FLOAT, 512]
  %064_convolutional_conv_weights[FLOAT, 512x1024x1x1]
  %065_convolutional_bn_scale[FLOAT, 1024]
  %065_convolutional_bn_bias[FLOAT, 1024]
  %065_convolutional_bn_mean[FLOAT, 1024]
  %065_convolutional_bn_var[FLOAT, 1024]
  %065_convolutional_conv_weights[FLOAT, 1024x512x3x3]
  %067_convolutional_bn_scale[FLOAT, 512]
  %067_convolutional_bn_bias[FLOAT, 512]
  %067_convolutional_bn_mean[FLOAT, 512]
  %067_convolutional_bn_var[FLOAT, 512]
  %067_convolutional_conv_weights[FLOAT, 512x1024x1x1]
  %068_convolutional_bn_scale[FLOAT, 1024]
  %068_convolutional_bn_bias[FLOAT, 1024]
  %068_convolutional_bn_mean[FLOAT, 1024]
  %068_convolutional_bn_var[FLOAT, 1024]
  %068_convolutional_conv_weights[FLOAT, 1024x512x3x3]
  %070_convolutional_bn_scale[FLOAT, 512]
  %070_convolutional_bn_bias[FLOAT, 512]
  %070_convolutional_bn_mean[FLOAT, 512]
  %070_convolutional_bn_var[FLOAT, 512]
  %070_convolutional_conv_weights[FLOAT, 512x1024x1x1]
  %071_convolutional_bn_scale[FLOAT, 1024]
  %071_convolutional_bn_bias[FLOAT, 1024]
  %071_convolutional_bn_mean[FLOAT, 1024]
  %071_convolutional_bn_var[FLOAT, 1024]
  %071_convolutional_conv_weights[FLOAT, 1024x512x3x3]
  %073_convolutional_bn_scale[FLOAT, 512]
  %073_convolutional_bn_bias[FLOAT, 512]
  %073_convolutional_bn_mean[FLOAT, 512]
  %073_convolutional_bn_var[FLOAT, 512]
  %073_convolutional_conv_weights[FLOAT, 512x1024x1x1]
  %074_convolutional_bn_scale[FLOAT, 1024]
  %074_convolutional_bn_bias[FLOAT, 1024]
  %074_convolutional_bn_mean[FLOAT, 1024]
  %074_convolutional_bn_var[FLOAT, 1024]
  %074_convolutional_conv_weights[FLOAT, 1024x512x3x3]
  %076_convolutional_bn_scale[FLOAT, 512]
  %076_convolutional_bn_bias[FLOAT, 512]
  %076_convolutional_bn_mean[FLOAT, 512]
  %076_convolutional_bn_var[FLOAT, 512]
  %076_convolutional_conv_weights[FLOAT, 512x1024x1x1]
  %077_convolutional_bn_scale[FLOAT, 1024]
  %077_convolutional_bn_bias[FLOAT, 1024]
  %077_convolutional_bn_mean[FLOAT, 1024]
  %077_convolutional_bn_var[FLOAT, 1024]
  %077_convolutional_conv_weights[FLOAT, 1024x512x3x3]
  %078_convolutional_bn_scale[FLOAT, 512]
  %078_convolutional_bn_bias[FLOAT, 512]
  %078_convolutional_bn_mean[FLOAT, 512]
  %078_convolutional_bn_var[FLOAT, 512]
  %078_convolutional_conv_weights[FLOAT, 512x1024x1x1]
  %079_convolutional_bn_scale[FLOAT, 1024]
  %079_convolutional_bn_bias[FLOAT, 1024]
  %079_convolutional_bn_mean[FLOAT, 1024]
  %079_convolutional_bn_var[FLOAT, 1024]
  %079_convolutional_conv_weights[FLOAT, 1024x512x3x3]
  %080_convolutional_bn_scale[FLOAT, 512]
  %080_convolutional_bn_bias[FLOAT, 512]
  %080_convolutional_bn_mean[FLOAT, 512]
  %080_convolutional_bn_var[FLOAT, 512]
  %080_convolutional_conv_weights[FLOAT, 512x1024x1x1]
  %081_convolutional_bn_scale[FLOAT, 1024]
  %081_convolutional_bn_bias[FLOAT, 1024]
  %081_convolutional_bn_mean[FLOAT, 1024]
  %081_convolutional_bn_var[FLOAT, 1024]
  %081_convolutional_conv_weights[FLOAT, 1024x512x3x3]
  %082_convolutional_conv_bias[FLOAT, 255]
  %082_convolutional_conv_weights[FLOAT, 255x1024x1x1]
  %085_convolutional_bn_scale[FLOAT, 256]
  %085_convolutional_bn_bias[FLOAT, 256]
  %085_convolutional_bn_mean[FLOAT, 256]
  %085_convolutional_bn_var[FLOAT, 256]
  %085_convolutional_conv_weights[FLOAT, 256x512x1x1]
  %086_upsample_scale[FLOAT, 4]
  %088_convolutional_bn_scale[FLOAT, 256]
  %088_convolutional_bn_bias[FLOAT, 256]
  %088_convolutional_bn_mean[FLOAT, 256]
  %088_convolutional_bn_var[FLOAT, 256]
  %088_convolutional_conv_weights[FLOAT, 256x768x1x1]
  %089_convolutional_bn_scale[FLOAT, 512]
  %089_convolutional_bn_bias[FLOAT, 512]
  %089_convolutional_bn_mean[FLOAT, 512]
  %089_convolutional_bn_var[FLOAT, 512]
  %089_convolutional_conv_weights[FLOAT, 512x256x3x3]
  %090_convolutional_bn_scale[FLOAT, 256]
  %090_convolutional_bn_bias[FLOAT, 256]
  %090_convolutional_bn_mean[FLOAT, 256]
  %090_convolutional_bn_var[FLOAT, 256]
  %090_convolutional_conv_weights[FLOAT, 256x512x1x1]
  %091_convolutional_bn_scale[FLOAT, 512]
  %091_convolutional_bn_bias[FLOAT, 512]
  %091_convolutional_bn_mean[FLOAT, 512]
  %091_convolutional_bn_var[FLOAT, 512]
  %091_convolutional_conv_weights[FLOAT, 512x256x3x3]
  %092_convolutional_bn_scale[FLOAT, 256]
  %092_convolutional_bn_bias[FLOAT, 256]
  %092_convolutional_bn_mean[FLOAT, 256]
  %092_convolutional_bn_var[FLOAT, 256]
  %092_convolutional_conv_weights[FLOAT, 256x512x1x1]
  %093_convolutional_bn_scale[FLOAT, 512]
  %093_convolutional_bn_bias[FLOAT, 512]
  %093_convolutional_bn_mean[FLOAT, 512]
  %093_convolutional_bn_var[FLOAT, 512]
  %093_convolutional_conv_weights[FLOAT, 512x256x3x3]
  %094_convolutional_conv_bias[FLOAT, 255]
  %094_convolutional_conv_weights[FLOAT, 255x512x1x1]
  %097_convolutional_bn_scale[FLOAT, 128]
  %097_convolutional_bn_bias[FLOAT, 128]
  %097_convolutional_bn_mean[FLOAT, 128]
  %097_convolutional_bn_var[FLOAT, 128]
  %097_convolutional_conv_weights[FLOAT, 128x256x1x1]
  %098_upsample_scale[FLOAT, 4]
  %100_convolutional_bn_scale[FLOAT, 128]
  %100_convolutional_bn_bias[FLOAT, 128]
  %100_convolutional_bn_mean[FLOAT, 128]
  %100_convolutional_bn_var[FLOAT, 128]
  %100_convolutional_conv_weights[FLOAT, 128x384x1x1]
  %101_convolutional_bn_scale[FLOAT, 256]
  %101_convolutional_bn_bias[FLOAT, 256]
  %101_convolutional_bn_mean[FLOAT, 256]
  %101_convolutional_bn_var[FLOAT, 256]
  %101_convolutional_conv_weights[FLOAT, 256x128x3x3]
  %102_convolutional_bn_scale[FLOAT, 128]
  %102_convolutional_bn_bias[FLOAT, 128]
  %102_convolutional_bn_mean[FLOAT, 128]
  %102_convolutional_bn_var[FLOAT, 128]
  %102_convolutional_conv_weights[FLOAT, 128x256x1x1]
  %103_convolutional_bn_scale[FLOAT, 256]
  %103_convolutional_bn_bias[FLOAT, 256]
  %103_convolutional_bn_mean[FLOAT, 256]
  %103_convolutional_bn_var[FLOAT, 256]
  %103_convolutional_conv_weights[FLOAT, 256x128x3x3]
  %104_convolutional_bn_scale[FLOAT, 128]
  %104_convolutional_bn_bias[FLOAT, 128]
  %104_convolutional_bn_mean[FLOAT, 128]
  %104_convolutional_bn_var[FLOAT, 128]
  %104_convolutional_conv_weights[FLOAT, 128x256x1x1]
  %105_convolutional_bn_scale[FLOAT, 256]
  %105_convolutional_bn_bias[FLOAT, 256]
  %105_convolutional_bn_mean[FLOAT, 256]
  %105_convolutional_bn_var[FLOAT, 256]
  %105_convolutional_conv_weights[FLOAT, 256x128x3x3]
  %106_convolutional_conv_bias[FLOAT, 255]
  %106_convolutional_conv_weights[FLOAT, 255x256x1x1]
) {
  %001_convolutional = Conv[auto_pad = u'SAME_LOWER', dilations = [1, 1], kernel_shape = [3, 3], strides = [1, 1]](%000_net, %001_convolutional_conv_weights)
  %001_convolutional_bn = BatchNormalization[epsilon = 1e-05, momentum = 0.99](%001_convolutional, %001_convolutional_bn_scale, %001_convolutional_bn_bias, %001_convolutional_bn_mean, %001_convolutional_bn_var)
  %001_convolutional_lrelu = LeakyRelu[alpha = 0.1](%001_convolutional_bn)
  %002_convolutional = Conv[auto_pad = u'SAME_LOWER', dilations = [1, 1], kernel_shape = [3, 3], strides = [2, 2]](%001_convolutional_lrelu, %002_convolutional_conv_weights)
  %002_convolutional_bn = BatchNormalization[epsilon = 1e-05, momentum = 0.99](%002_convolutional, %002_convolutional_bn_scale, %002_convolutional_bn_bias, %002_convolutional_bn_mean, %002_convolutional_bn_var)
  %002_convolutional_lrelu = LeakyRelu[alpha = 0.1](%002_convolutional_bn)
  %003_convolutional = Conv[auto_pad = u'SAME_LOWER', dilations = [1, 1], kernel_shape = [1, 1], strides = [1, 1]](%002_convolutional_lrelu, %003_convolutional_conv_weights)
  %003_convolutional_bn = BatchNormalization[epsilon = 1e-05, momentum = 0.99](%003_convolutional, %003_convolutional_bn_scale, %003_convolutional_bn_bias, %003_convolutional_bn_mean, %003_convolutional_bn_var)
  %003_convolutional_lrelu = LeakyRelu[alpha = 0.1](%003_convolutional_bn)
  %004_convolutional = Conv[auto_pad = u'SAME_LOWER', dilations = [1, 1], kernel_shape = [3, 3], strides = [1, 1]](%003_convolutional_lrelu, %004_convolutional_conv_weights)
  %004_convolutional_bn = BatchNormalization[epsilon = 1e-05, momentum = 0.99](%004_convolutional, %004_convolutional_bn_scale, %004_convolutional_bn_bias, %004_convolutional_bn_mean, %004_convolutional_bn_var)
  %004_convolutional_lrelu = LeakyRelu[alpha = 0.1](%004_convolutional_bn)
  %005_shortcut = Add(%004_convolutional_lrelu, %002_convolutional_lrelu)
  %006_convolutional = Conv[auto_pad = u'SAME_LOWER', dilations = [1, 1], kernel_shape = [3, 3], strides = [2, 2]](%005_shortcut, %006_convolutional_conv_weights)
  %006_convolutional_bn = BatchNormalization[epsilon = 1e-05, momentum = 0.99](%006_convolutional, %006_convolutional_bn_scale, %006_convolutional_bn_bias, %006_convolutional_bn_mean, %006_convolutional_bn_var)
  %006_convolutional_lrelu = LeakyRelu[alpha = 0.1](%006_convolutional_bn)
  %007_convolutional = Conv[auto_pad = u'SAME_LOWER', dilations = [1, 1], kernel_shape = [1, 1], strides = [1, 1]](%006_convolutional_lrelu, %007_convolutional_conv_weights)
  %007_convolutional_bn = BatchNormalization[epsilon = 1e-05, momentum = 0.99](%007_convolutional, %007_convolutional_bn_scale, %007_convolutional_bn_bias, %007_convolutional_bn_mean, %007_convolutional_bn_var)
  %007_convolutional_lrelu = LeakyRelu[alpha = 0.1](%007_convolutional_bn)
  %008_convolutional = Conv[auto_pad = u'SAME_LOWER', dilations = [1, 1], kernel_shape = [3, 3], strides = [1, 1]](%007_convolutional_lrelu, %008_convolutional_conv_weights)
  %008_convolutional_bn = BatchNormalization[epsilon = 1e-05, momentum = 0.99](%008_convolutional, %008_convolutional_bn_scale, %008_convolutional_bn_bias, %008_convolutional_bn_mean, %008_convolutional_bn_var)
  %008_convolutional_lrelu = LeakyRelu[alpha = 0.1](%008_convolutional_bn)
  %009_shortcut = Add(%008_convolutional_lrelu, %006_convolutional_lrelu)
  %010_convolutional = Conv[auto_pad = u'SAME_LOWER', dilations = [1, 1], kernel_shape = [1, 1], strides = [1, 1]](%009_shortcut, %010_convolutional_conv_weights)
  %010_convolutional_bn = BatchNormalization[epsilon = 1e-05, momentum = 0.99](%010_convolutional, %010_convolutional_bn_scale, %010_convolutional_bn_bias, %010_convolutional_bn_mean, %010_convolutional_bn_var)
  %010_convolutional_lrelu = LeakyRelu[alpha = 0.1](%010_convolutional_bn)
  %011_convolutional = Conv[auto_pad = u'SAME_LOWER', dilations = [1, 1], kernel_shape = [3, 3], strides = [1, 1]](%010_convolutional_lrelu, %011_convolutional_conv_weights)
  %011_convolutional_bn = BatchNormalization[epsilon = 1e-05, momentum = 0.99](%011_convolutional, %011_convolutional_bn_scale, %011_convolutional_bn_bias, %011_convolutional_bn_mean, %011_convolutional_bn_var)
  %011_convolutional_lrelu = LeakyRelu[alpha = 0.1](%011_convolutional_bn)
  %012_shortcut = Add(%011_convolutional_lrelu, %009_shortcut)
  %013_convolutional = Conv[auto_pad = u'SAME_LOWER', dilations = [1, 1], kernel_shape = [3, 3], strides = [2, 2]](%012_shortcut, %013_convolutional_conv_weights)
  %013_convolutional_bn = BatchNormalization[epsilon = 1e-05, momentum = 0.99](%013_convolutional, %013_convolutional_bn_scale, %013_convolutional_bn_bias, %013_convolutional_bn_mean, %013_convolutional_bn_var)
  %013_convolutional_lrelu = LeakyRelu[alpha = 0.1](%013_convolutional_bn)
  %014_convolutional = Conv[auto_pad = u'SAME_LOWER', dilations = [1, 1], kernel_shape = [1, 1], strides = [1, 1]](%013_convolutional_lrelu, %014_convolutional_conv_weights)
  %014_convolutional_bn = BatchNormalization[epsilon = 1e-05, momentum = 0.99](%014_convolutional, %014_convolutional_bn_scale, %014_convolutional_bn_bias, %014_convolutional_bn_mean, %014_convolutional_bn_var)
  %014_convolutional_lrelu = LeakyRelu[alpha = 0.1](%014_convolutional_bn)
  %015_convolutional = Conv[auto_pad = u'SAME_LOWER', dilations = [1, 1], kernel_shape = [3, 3], strides = [1, 1]](%014_convolutional_lrelu, %015_convolutional_conv_weights)
  %015_convolutional_bn = BatchNormalization[epsilon = 1e-05, momentum = 0.99](%015_convolutional, %015_convolutional_bn_scale, %015_convolutional_bn_bias, %015_convolutional_bn_mean, %015_convolutional_bn_var)
  %015_convolutional_lrelu = LeakyRelu[alpha = 0.1](%015_convolutional_bn)
  %016_shortcut = Add(%015_convolutional_lrelu, %013_convolutional_lrelu)
  %017_convolutional = Conv[auto_pad = u'SAME_LOWER', dilations = [1, 1], kernel_shape = [1, 1], strides = [1, 1]](%016_shortcut, %017_convolutional_conv_weights)
  %017_convolutional_bn = BatchNormalization[epsilon = 1e-05, momentum = 0.99](%017_convolutional, %017_convolutional_bn_scale, %017_convolutional_bn_bias, %017_convolutional_bn_mean, %017_convolutional_bn_var)
  %017_convolutional_lrelu = LeakyRelu[alpha = 0.1](%017_convolutional_bn)
  %018_convolutional = Conv[auto_pad = u'SAME_LOWER', dilations = [1, 1], kernel_shape = [3, 3], strides = [1, 1]](%017_convolutional_lrelu, %018_convolutional_conv_weights)
  %018_convolutional_bn = BatchNormalization[epsilon = 1e-05, momentum = 0.99](%018_convolutional, %018_convolutional_bn_scale, %018_convolutional_bn_bias, %018_convolutional_bn_mean, %018_convolutional_bn_var)
  %018_convolutional_lrelu = LeakyRelu[alpha = 0.1](%018_convolutional_bn)
  %019_shortcut = Add(%018_convolutional_lrelu, %016_shortcut)
  %020_convolutional = Conv[auto_pad = u'SAME_LOWER', dilations = [1, 1], kernel_shape = [1, 1], strides = [1, 1]](%019_shortcut, %020_convolutional_conv_weights)
  %020_convolutional_bn = BatchNormalization[epsilon = 1e-05, momentum = 0.99](%020_convolutional, %020_convolutional_bn_scale, %020_convolutional_bn_bias, %020_convolutional_bn_mean, %020_convolutional_bn_var)
  %020_convolutional_lrelu = LeakyRelu[alpha = 0.1](%020_convolutional_bn)
  %021_convolutional = Conv[auto_pad = u'SAME_LOWER', dilations = [1, 1], kernel_shape = [3, 3], strides = [1, 1]](%020_convolutional_lrelu, %021_convolutional_conv_weights)
  %021_convolutional_bn = BatchNormalization[epsilon = 1e-05, momentum = 0.99](%021_convolutional, %021_convolutional_bn_scale, %021_convolutional_bn_bias, %021_convolutional_bn_mean, %021_convolutional_bn_var)
  %021_convolutional_lrelu = LeakyRelu[alpha = 0.1](%021_convolutional_bn)
  %022_shortcut = Add(%021_convolutional_lrelu, %019_shortcut)
  %023_convolutional = Conv[auto_pad = u'SAME_LOWER', dilations = [1, 1], kernel_shape = [1, 1], strides = [1, 1]](%022_shortcut, %023_convolutional_conv_weights)
  %023_convolutional_bn = BatchNormalization[epsilon = 1e-05, momentum = 0.99](%023_convolutional, %023_convolutional_bn_scale, %023_convolutional_bn_bias, %023_convolutional_bn_mean, %023_convolutional_bn_var)
  %023_convolutional_lrelu = LeakyRelu[alpha = 0.1](%023_convolutional_bn)
  %024_convolutional = Conv[auto_pad = u'SAME_LOWER', dilations = [1, 1], kernel_shape = [3, 3], strides = [1, 1]](%023_convolutional_lrelu, %024_convolutional_conv_weights)
  %024_convolutional_bn = BatchNormalization[epsilon = 1e-05, momentum = 0.99](%024_convolutional, %024_convolutional_bn_scale, %024_convolutional_bn_bias, %024_convolutional_bn_mean, %024_convolutional_bn_var)
  %024_convolutional_lrelu = LeakyRelu[alpha = 0.1](%024_convolutional_bn)
  %025_shortcut = Add(%024_convolutional_lrelu, %022_shortcut)
  %026_convolutional = Conv[auto_pad = u'SAME_LOWER', dilations = [1, 1], kernel_shape = [1, 1], strides = [1, 1]](%025_shortcut, %026_convolutional_conv_weights)
  %026_convolutional_bn = BatchNormalization[epsilon = 1e-05, momentum = 0.99](%026_convolutional, %026_convolutional_bn_scale, %026_convolutional_bn_bias, %026_convolutional_bn_mean, %026_convolutional_bn_var)
  %026_convolutional_lrelu = LeakyRelu[alpha = 0.1](%026_convolutional_bn)
  %027_convolutional = Conv[auto_pad = u'SAME_LOWER', dilations = [1, 1], kernel_shape = [3, 3], strides = [1, 1]](%026_convolutional_lrelu, %027_convolutional_conv_weights)
  %027_convolutional_bn = BatchNormalization[epsilon = 1e-05, momentum = 0.99](%027_convolutional, %027_convolutional_bn_scale, %027_convolutional_bn_bias, %027_convolutional_bn_mean, %027_convolutional_bn_var)
  %027_convolutional_lrelu = LeakyRelu[alpha = 0.1](%027_convolutional_bn)
  %028_shortcut = Add(%027_convolutional_lrelu, %025_shortcut)
  %029_convolutional = Conv[auto_pad = u'SAME_LOWER', dilations = [1, 1], kernel_shape = [1, 1], strides = [1, 1]](%028_shortcut, %029_convolutional_conv_weights)
  %029_convolutional_bn = BatchNormalization[epsilon = 1e-05, momentum = 0.99](%029_convolutional, %029_convolutional_bn_scale, %029_convolutional_bn_bias, %029_convolutional_bn_mean, %029_convolutional_bn_var)
  %029_convolutional_lrelu = LeakyRelu[alpha = 0.1](%029_convolutional_bn)
  %030_convolutional = Conv[auto_pad = u'SAME_LOWER', dilations = [1, 1], kernel_shape = [3, 3], strides = [1, 1]](%029_convolutional_lrelu, %030_convolutional_conv_weights)
  %030_convolutional_bn = BatchNormalization[epsilon = 1e-05, momentum = 0.99](%030_convolutional, %030_convolutional_bn_scale, %030_convolutional_bn_bias, %030_convolutional_bn_mean, %030_convolutional_bn_var)
  %030_convolutional_lrelu = LeakyRelu[alpha = 0.1](%030_convolutional_bn)
  %031_shortcut = Add(%030_convolutional_lrelu, %028_shortcut)
  %032_convolutional = Conv[auto_pad = u'SAME_LOWER', dilations = [1, 1], kernel_shape = [1, 1], strides = [1, 1]](%031_shortcut, %032_convolutional_conv_weights)
  %032_convolutional_bn = BatchNormalization[epsilon = 1e-05, momentum = 0.99](%032_convolutional, %032_convolutional_bn_scale, %032_convolutional_bn_bias, %032_convolutional_bn_mean, %032_convolutional_bn_var)
  %032_convolutional_lrelu = LeakyRelu[alpha = 0.1](%032_convolutional_bn)
  %033_convolutional = Conv[auto_pad = u'SAME_LOWER', dilations = [1, 1], kernel_shape = [3, 3], strides = [1, 1]](%032_convolutional_lrelu, %033_convolutional_conv_weights)
  %033_convolutional_bn = BatchNormalization[epsilon = 1e-05, momentum = 0.99](%033_convolutional, %033_convolutional_bn_scale, %033_convolutional_bn_bias, %033_convolutional_bn_mean, %033_convolutional_bn_var)
  %033_convolutional_lrelu = LeakyRelu[alpha = 0.1](%033_convolutional_bn)
  %034_shortcut = Add(%033_convolutional_lrelu, %031_shortcut)
  %035_convolutional = Conv[auto_pad = u'SAME_LOWER', dilations = [1, 1], kernel_shape = [1, 1], strides = [1, 1]](%034_shortcut, %035_convolutional_conv_weights)
  %035_convolutional_bn = BatchNormalization[epsilon = 1e-05, momentum = 0.99](%035_convolutional, %035_convolutional_bn_scale, %035_convolutional_bn_bias, %035_convolutional_bn_mean, %035_convolutional_bn_var)
  %035_convolutional_lrelu = LeakyRelu[alpha = 0.1](%035_convolutional_bn)
  %036_convolutional = Conv[auto_pad = u'SAME_LOWER', dilations = [1, 1], kernel_shape = [3, 3], strides = [1, 1]](%035_convolutional_lrelu, %036_convolutional_conv_weights)
  %036_convolutional_bn = BatchNormalization[epsilon = 1e-05, momentum = 0.99](%036_convolutional, %036_convolutional_bn_scale, %036_convolutional_bn_bias, %036_convolutional_bn_mean, %036_convolutional_bn_var)
  %036_convolutional_lrelu = LeakyRelu[alpha = 0.1](%036_convolutional_bn)
  %037_shortcut = Add(%036_convolutional_lrelu, %034_shortcut)
  %038_convolutional = Conv[auto_pad = u'SAME_LOWER', dilations = [1, 1], kernel_shape = [3, 3], strides = [2, 2]](%037_shortcut, %038_convolutional_conv_weights)
  %038_convolutional_bn = BatchNormalization[epsilon = 1e-05, momentum = 0.99](%038_convolutional, %038_convolutional_bn_scale, %038_convolutional_bn_bias, %038_convolutional_bn_mean, %038_convolutional_bn_var)
  %038_convolutional_lrelu = LeakyRelu[alpha = 0.1](%038_convolutional_bn)
  %039_convolutional = Conv[auto_pad = u'SAME_LOWER', dilations = [1, 1], kernel_shape = [1, 1], strides = [1, 1]](%038_convolutional_lrelu, %039_convolutional_conv_weights)
  %039_convolutional_bn = BatchNormalization[epsilon = 1e-05, momentum = 0.99](%039_convolutional, %039_convolutional_bn_scale, %039_convolutional_bn_bias, %039_convolutional_bn_mean, %039_convolutional_bn_var)
  %039_convolutional_lrelu = LeakyRelu[alpha = 0.1](%039_convolutional_bn)
  %040_convolutional = Conv[auto_pad = u'SAME_LOWER', dilations = [1, 1], kernel_shape = [3, 3], strides = [1, 1]](%039_convolutional_lrelu, %040_convolutional_conv_weights)
  %040_convolutional_bn = BatchNormalization[epsilon = 1e-05, momentum = 0.99](%040_convolutional, %040_convolutional_bn_scale, %040_convolutional_bn_bias, %040_convolutional_bn_mean, %040_convolutional_bn_var)
  %040_convolutional_lrelu = LeakyRelu[alpha = 0.1](%040_convolutional_bn)
  %041_shortcut = Add(%040_convolutional_lrelu, %038_convolutional_lrelu)
  %042_convolutional = Conv[auto_pad = u'SAME_LOWER', dilations = [1, 1], kernel_shape = [1, 1], strides = [1, 1]](%041_shortcut, %042_convolutional_conv_weights)
  %042_convolutional_bn = BatchNormalization[epsilon = 1e-05, momentum = 0.99](%042_convolutional, %042_convolutional_bn_scale, %042_convolutional_bn_bias, %042_convolutional_bn_mean, %042_convolutional_bn_var)
  %042_convolutional_lrelu = LeakyRelu[alpha = 0.1](%042_convolutional_bn)
  %043_convolutional = Conv[auto_pad = u'SAME_LOWER', dilations = [1, 1], kernel_shape = [3, 3], strides = [1, 1]](%042_convolutional_lrelu, %043_convolutional_conv_weights)
  %043_convolutional_bn = BatchNormalization[epsilon = 1e-05, momentum = 0.99](%043_convolutional, %043_convolutional_bn_scale, %043_convolutional_bn_bias, %043_convolutional_bn_mean, %043_convolutional_bn_var)
  %043_convolutional_lrelu = LeakyRelu[alpha = 0.1](%043_convolutional_bn)
  %044_shortcut = Add(%043_convolutional_lrelu, %041_shortcut)
  %045_convolutional = Conv[auto_pad = u'SAME_LOWER', dilations = [1, 1], kernel_shape = [1, 1], strides = [1, 1]](%044_shortcut, %045_convolutional_conv_weights)
  %045_convolutional_bn = BatchNormalization[epsilon = 1e-05, momentum = 0.99](%045_convolutional, %045_convolutional_bn_scale, %045_convolutional_bn_bias, %045_convolutional_bn_mean, %045_convolutional_bn_var)
  %045_convolutional_lrelu = LeakyRelu[alpha = 0.1](%045_convolutional_bn)
  %046_convolutional = Conv[auto_pad = u'SAME_LOWER', dilations = [1, 1], kernel_shape = [3, 3], strides = [1, 1]](%045_convolutional_lrelu, %046_convolutional_conv_weights)
  %046_convolutional_bn = BatchNormalization[epsilon = 1e-05, momentum = 0.99](%046_convolutional, %046_convolutional_bn_scale, %046_convolutional_bn_bias, %046_convolutional_bn_mean, %046_convolutional_bn_var)
  %046_convolutional_lrelu = LeakyRelu[alpha = 0.1](%046_convolutional_bn)
  %047_shortcut = Add(%046_convolutional_lrelu, %044_shortcut)
  %048_convolutional = Conv[auto_pad = u'SAME_LOWER', dilations = [1, 1], kernel_shape = [1, 1], strides = [1, 1]](%047_shortcut, %048_convolutional_conv_weights)
  %048_convolutional_bn = BatchNormalization[epsilon = 1e-05, momentum = 0.99](%048_convolutional, %048_convolutional_bn_scale, %048_convolutional_bn_bias, %048_convolutional_bn_mean, %048_convolutional_bn_var)
  %048_convolutional_lrelu = LeakyRelu[alpha = 0.1](%048_convolutional_bn)
  %049_convolutional = Conv[auto_pad = u'SAME_LOWER', dilations = [1, 1], kernel_shape = [3, 3], strides = [1, 1]](%048_convolutional_lrelu, %049_convolutional_conv_weights)
  %049_convolutional_bn = BatchNormalization[epsilon = 1e-05, momentum = 0.99](%049_convolutional, %049_convolutional_bn_scale, %049_convolutional_bn_bias, %049_convolutional_bn_mean, %049_convolutional_bn_var)
  %049_convolutional_lrelu = LeakyRelu[alpha = 0.1](%049_convolutional_bn)
  %050_shortcut = Add(%049_convolutional_lrelu, %047_shortcut)
  %051_convolutional = Conv[auto_pad = u'SAME_LOWER', dilations = [1, 1], kernel_shape = [1, 1], strides = [1, 1]](%050_shortcut, %051_convolutional_conv_weights)
  %051_convolutional_bn = BatchNormalization[epsilon = 1e-05, momentum = 0.99](%051_convolutional, %051_convolutional_bn_scale, %051_convolutional_bn_bias, %051_convolutional_bn_mean, %051_convolutional_bn_var)
  %051_convolutional_lrelu = LeakyRelu[alpha = 0.1](%051_convolutional_bn)
  %052_convolutional = Conv[auto_pad = u'SAME_LOWER', dilations = [1, 1], kernel_shape = [3, 3], strides = [1, 1]](%051_convolutional_lrelu, %052_convolutional_conv_weights)
  %052_convolutional_bn = BatchNormalization[epsilon = 1e-05, momentum = 0.99](%052_convolutional, %052_convolutional_bn_scale, %052_convolutional_bn_bias, %052_convolutional_bn_mean, %052_convolutional_bn_var)
  %052_convolutional_lrelu = LeakyRelu[alpha = 0.1](%052_convolutional_bn)
  %053_shortcut = Add(%052_convolutional_lrelu, %050_shortcut)
  %054_convolutional = Conv[auto_pad = u'SAME_LOWER', dilations = [1, 1], kernel_shape = [1, 1], strides = [1, 1]](%053_shortcut, %054_convolutional_conv_weights)
  %054_convolutional_bn = BatchNormalization[epsilon = 1e-05, momentum = 0.99](%054_convolutional, %054_convolutional_bn_scale, %054_convolutional_bn_bias, %054_convolutional_bn_mean, %054_convolutional_bn_var)
  %054_convolutional_lrelu = LeakyRelu[alpha = 0.1](%054_convolutional_bn)
  %055_convolutional = Conv[auto_pad = u'SAME_LOWER', dilations = [1, 1], kernel_shape = [3, 3], strides = [1, 1]](%054_convolutional_lrelu, %055_convolutional_conv_weights)
  %055_convolutional_bn = BatchNormalization[epsilon = 1e-05, momentum = 0.99](%055_convolutional, %055_convolutional_bn_scale, %055_convolutional_bn_bias, %055_convolutional_bn_mean, %055_convolutional_bn_var)
  %055_convolutional_lrelu = LeakyRelu[alpha = 0.1](%055_convolutional_bn)
  %056_shortcut = Add(%055_convolutional_lrelu, %053_shortcut)
  %057_convolutional = Conv[auto_pad = u'SAME_LOWER', dilations = [1, 1], kernel_shape = [1, 1], strides = [1, 1]](%056_shortcut, %057_convolutional_conv_weights)
  %057_convolutional_bn = BatchNormalization[epsilon = 1e-05, momentum = 0.99](%057_convolutional, %057_convolutional_bn_scale, %057_convolutional_bn_bias, %057_convolutional_bn_mean, %057_convolutional_bn_var)
  %057_convolutional_lrelu = LeakyRelu[alpha = 0.1](%057_convolutional_bn)
  %058_convolutional = Conv[auto_pad = u'SAME_LOWER', dilations = [1, 1], kernel_shape = [3, 3], strides = [1, 1]](%057_convolutional_lrelu, %058_convolutional_conv_weights)
  %058_convolutional_bn = BatchNormalization[epsilon = 1e-05, momentum = 0.99](%058_convolutional, %058_convolutional_bn_scale, %058_convolutional_bn_bias, %058_convolutional_bn_mean, %058_convolutional_bn_var)
  %058_convolutional_lrelu = LeakyRelu[alpha = 0.1](%058_convolutional_bn)
  %059_shortcut = Add(%058_convolutional_lrelu, %056_shortcut)
  %060_convolutional = Conv[auto_pad = u'SAME_LOWER', dilations = [1, 1], kernel_shape = [1, 1], strides = [1, 1]](%059_shortcut, %060_convolutional_conv_weights)
  %060_convolutional_bn = BatchNormalization[epsilon = 1e-05, momentum = 0.99](%060_convolutional, %060_convolutional_bn_scale, %060_convolutional_bn_bias, %060_convolutional_bn_mean, %060_convolutional_bn_var)
  %060_convolutional_lrelu = LeakyRelu[alpha = 0.1](%060_convolutional_bn)
  %061_convolutional = Conv[auto_pad = u'SAME_LOWER', dilations = [1, 1], kernel_shape = [3, 3], strides = [1, 1]](%060_convolutional_lrelu, %061_convolutional_conv_weights)
  %061_convolutional_bn = BatchNormalization[epsilon = 1e-05, momentum = 0.99](%061_convolutional, %061_convolutional_bn_scale, %061_convolutional_bn_bias, %061_convolutional_bn_mean, %061_convolutional_bn_var)
  %061_convolutional_lrelu = LeakyRelu[alpha = 0.1](%061_convolutional_bn)
  %062_shortcut = Add(%061_convolutional_lrelu, %059_shortcut)
  %063_convolutional = Conv[auto_pad = u'SAME_LOWER', dilations = [1, 1], kernel_shape = [3, 3], strides = [2, 2]](%062_shortcut, %063_convolutional_conv_weights)
  %063_convolutional_bn = BatchNormalization[epsilon = 1e-05, momentum = 0.99](%063_convolutional, %063_convolutional_bn_scale, %063_convolutional_bn_bias, %063_convolutional_bn_mean, %063_convolutional_bn_var)
  %063_convolutional_lrelu = LeakyRelu[alpha = 0.1](%063_convolutional_bn)
  %064_convolutional = Conv[auto_pad = u'SAME_LOWER', dilations = [1, 1], kernel_shape = [1, 1], strides = [1, 1]](%063_convolutional_lrelu, %064_convolutional_conv_weights)
  %064_convolutional_bn = BatchNormalization[epsilon = 1e-05, momentum = 0.99](%064_convolutional, %064_convolutional_bn_scale, %064_convolutional_bn_bias, %064_convolutional_bn_mean, %064_convolutional_bn_var)
  %064_convolutional_lrelu = LeakyRelu[alpha = 0.1](%064_convolutional_bn)
  %065_convolutional = Conv[auto_pad = u'SAME_LOWER', dilations = [1, 1], kernel_shape = [3, 3], strides = [1, 1]](%064_convolutional_lrelu, %065_convolutional_conv_weights)
  %065_convolutional_bn = BatchNormalization[epsilon = 1e-05, momentum = 0.99](%065_convolutional, %065_convolutional_bn_scale, %065_convolutional_bn_bias, %065_convolutional_bn_mean, %065_convolutional_bn_var)
  %065_convolutional_lrelu = LeakyRelu[alpha = 0.1](%065_convolutional_bn)
  %066_shortcut = Add(%065_convolutional_lrelu, %063_convolutional_lrelu)
  %067_convolutional = Conv[auto_pad = u'SAME_LOWER', dilations = [1, 1], kernel_shape = [1, 1], strides = [1, 1]](%066_shortcut, %067_convolutional_conv_weights)
  %067_convolutional_bn = BatchNormalization[epsilon = 1e-05, momentum = 0.99](%067_convolutional, %067_convolutional_bn_scale, %067_convolutional_bn_bias, %067_convolutional_bn_mean, %067_convolutional_bn_var)
  %067_convolutional_lrelu = LeakyRelu[alpha = 0.1](%067_convolutional_bn)
  %068_convolutional = Conv[auto_pad = u'SAME_LOWER', dilations = [1, 1], kernel_shape = [3, 3], strides = [1, 1]](%067_convolutional_lrelu, %068_convolutional_conv_weights)
  %068_convolutional_bn = BatchNormalization[epsilon = 1e-05, momentum = 0.99](%068_convolutional, %068_convolutional_bn_scale, %068_convolutional_bn_bias, %068_convolutional_bn_mean, %068_convolutional_bn_var)
  %068_convolutional_lrelu = LeakyRelu[alpha = 0.1](%068_convolutional_bn)
  %069_shortcut = Add(%068_convolutional_lrelu, %066_shortcut)
  %070_convolutional = Conv[auto_pad = u'SAME_LOWER', dilations = [1, 1], kernel_shape = [1, 1], strides = [1, 1]](%069_shortcut, %070_convolutional_conv_weights)
  %070_convolutional_bn = BatchNormalization[epsilon = 1e-05, momentum = 0.99](%070_convolutional, %070_convolutional_bn_scale, %070_convolutional_bn_bias, %070_convolutional_bn_mean, %070_convolutional_bn_var)
  %070_convolutional_lrelu = LeakyRelu[alpha = 0.1](%070_convolutional_bn)
  %071_convolutional = Conv[auto_pad = u'SAME_LOWER', dilations = [1, 1], kernel_shape = [3, 3], strides = [1, 1]](%070_convolutional_lrelu, %071_convolutional_conv_weights)
  %071_convolutional_bn = BatchNormalization[epsilon = 1e-05, momentum = 0.99](%071_convolutional, %071_convolutional_bn_scale, %071_convolutional_bn_bias, %071_convolutional_bn_mean, %071_convolutional_bn_var)
  %071_convolutional_lrelu = LeakyRelu[alpha = 0.1](%071_convolutional_bn)
  %072_shortcut = Add(%071_convolutional_lrelu, %069_shortcut)
  %073_convolutional = Conv[auto_pad = u'SAME_LOWER', dilations = [1, 1], kernel_shape = [1, 1], strides = [1, 1]](%072_shortcut, %073_convolutional_conv_weights)
  %073_convolutional_bn = BatchNormalization[epsilon = 1e-05, momentum = 0.99](%073_convolutional, %073_convolutional_bn_scale, %073_convolutional_bn_bias, %073_convolutional_bn_mean, %073_convolutional_bn_var)
  %073_convolutional_lrelu = LeakyRelu[alpha = 0.1](%073_convolutional_bn)
  %074_convolutional = Conv[auto_pad = u'SAME_LOWER', dilations = [1, 1], kernel_shape = [3, 3], strides = [1, 1]](%073_convolutional_lrelu, %074_convolutional_conv_weights)
  %074_convolutional_bn = BatchNormalization[epsilon = 1e-05, momentum = 0.99](%074_convolutional, %074_convolutional_bn_scale, %074_convolutional_bn_bias, %074_convolutional_bn_mean, %074_convolutional_bn_var)
  %074_convolutional_lrelu = LeakyRelu[alpha = 0.1](%074_convolutional_bn)
  %075_shortcut = Add(%074_convolutional_lrelu, %072_shortcut)
  %076_convolutional = Conv[auto_pad = u'SAME_LOWER', dilations = [1, 1], kernel_shape = [1, 1], strides = [1, 1]](%075_shortcut, %076_convolutional_conv_weights)
  %076_convolutional_bn = BatchNormalization[epsilon = 1e-05, momentum = 0.99](%076_convolutional, %076_convolutional_bn_scale, %076_convolutional_bn_bias, %076_convolutional_bn_mean, %076_convolutional_bn_var)
  %076_convolutional_lrelu = LeakyRelu[alpha = 0.1](%076_convolutional_bn)
  %077_convolutional = Conv[auto_pad = u'SAME_LOWER', dilations = [1, 1], kernel_shape = [3, 3], strides = [1, 1]](%076_convolutional_lrelu, %077_convolutional_conv_weights)
  %077_convolutional_bn = BatchNormalization[epsilon = 1e-05, momentum = 0.99](%077_convolutional, %077_convolutional_bn_scale, %077_convolutional_bn_bias, %077_convolutional_bn_mean, %077_convolutional_bn_var)
  %077_convolutional_lrelu = LeakyRelu[alpha = 0.1](%077_convolutional_bn)
  %078_convolutional = Conv[auto_pad = u'SAME_LOWER', dilations = [1, 1], kernel_shape = [1, 1], strides = [1, 1]](%077_convolutional_lrelu, %078_convolutional_conv_weights)
  %078_convolutional_bn = BatchNormalization[epsilon = 1e-05, momentum = 0.99](%078_convolutional, %078_convolutional_bn_scale, %078_convolutional_bn_bias, %078_convolutional_bn_mean, %078_convolutional_bn_var)
  %078_convolutional_lrelu = LeakyRelu[alpha = 0.1](%078_convolutional_bn)
  %079_convolutional = Conv[auto_pad = u'SAME_LOWER', dilations = [1, 1], kernel_shape = [3, 3], strides = [1, 1]](%078_convolutional_lrelu, %079_convolutional_conv_weights)
  %079_convolutional_bn = BatchNormalization[epsilon = 1e-05, momentum = 0.99](%079_convolutional, %079_convolutional_bn_scale, %079_convolutional_bn_bias, %079_convolutional_bn_mean, %079_convolutional_bn_var)
  %079_convolutional_lrelu = LeakyRelu[alpha = 0.1](%079_convolutional_bn)
  %080_convolutional = Conv[auto_pad = u'SAME_LOWER', dilations = [1, 1], kernel_shape = [1, 1], strides = [1, 1]](%079_convolutional_lrelu, %080_convolutional_conv_weights)
  %080_convolutional_bn = BatchNormalization[epsilon = 1e-05, momentum = 0.99](%080_convolutional, %080_convolutional_bn_scale, %080_convolutional_bn_bias, %080_convolutional_bn_mean, %080_convolutional_bn_var)
  %080_convolutional_lrelu = LeakyRelu[alpha = 0.1](%080_convolutional_bn)
  %081_convolutional = Conv[auto_pad = u'SAME_LOWER', dilations = [1, 1], kernel_shape = [3, 3], strides = [1, 1]](%080_convolutional_lrelu, %081_convolutional_conv_weights)
  %081_convolutional_bn = BatchNormalization[epsilon = 1e-05, momentum = 0.99](%081_convolutional, %081_convolutional_bn_scale, %081_convolutional_bn_bias, %081_convolutional_bn_mean, %081_convolutional_bn_var)
  %081_convolutional_lrelu = LeakyRelu[alpha = 0.1](%081_convolutional_bn)
  %082_convolutional = Conv[auto_pad = u'SAME_LOWER', dilations = [1, 1], kernel_shape = [1, 1], strides = [1, 1]](%081_convolutional_lrelu, %082_convolutional_conv_weights, %082_convolutional_conv_bias)
  %085_convolutional = Conv[auto_pad = u'SAME_LOWER', dilations = [1, 1], kernel_shape = [1, 1], strides = [1, 1]](%080_convolutional_lrelu, %085_convolutional_conv_weights)
  %085_convolutional_bn = BatchNormalization[epsilon = 1e-05, momentum = 0.99](%085_convolutional, %085_convolutional_bn_scale, %085_convolutional_bn_bias, %085_convolutional_bn_mean, %085_convolutional_bn_var)
  %085_convolutional_lrelu = LeakyRelu[alpha = 0.1](%085_convolutional_bn)
  %086_upsample = Upsample[mode = u'nearest'](%085_convolutional_lrelu, %086_upsample_scale)
  %087_route = Concat[axis = 1](%086_upsample, %062_shortcut)
  %088_convolutional = Conv[auto_pad = u'SAME_LOWER', dilations = [1, 1], kernel_shape = [1, 1], strides = [1, 1]](%087_route, %088_convolutional_conv_weights)
  %088_convolutional_bn = BatchNormalization[epsilon = 1e-05, momentum = 0.99](%088_convolutional, %088_convolutional_bn_scale, %088_convolutional_bn_bias, %088_convolutional_bn_mean, %088_convolutional_bn_var)
  %088_convolutional_lrelu = LeakyRelu[alpha = 0.1](%088_convolutional_bn)
  %089_convolutional = Conv[auto_pad = u'SAME_LOWER', dilations = [1, 1], kernel_shape = [3, 3], strides = [1, 1]](%088_convolutional_lrelu, %089_convolutional_conv_weights)
  %089_convolutional_bn = BatchNormalization[epsilon = 1e-05, momentum = 0.99](%089_convolutional, %089_convolutional_bn_scale, %089_convolutional_bn_bias, %089_convolutional_bn_mean, %089_convolutional_bn_var)
  %089_convolutional_lrelu = LeakyRelu[alpha = 0.1](%089_convolutional_bn)
  %090_convolutional = Conv[auto_pad = u'SAME_LOWER', dilations = [1, 1], kernel_shape = [1, 1], strides = [1, 1]](%089_convolutional_lrelu, %090_convolutional_conv_weights)
  %090_convolutional_bn = BatchNormalization[epsilon = 1e-05, momentum = 0.99](%090_convolutional, %090_convolutional_bn_scale, %090_convolutional_bn_bias, %090_convolutional_bn_mean, %090_convolutional_bn_var)
  %090_convolutional_lrelu = LeakyRelu[alpha = 0.1](%090_convolutional_bn)
  %091_convolutional = Conv[auto_pad = u'SAME_LOWER', dilations = [1, 1], kernel_shape = [3, 3], strides = [1, 1]](%090_convolutional_lrelu, %091_convolutional_conv_weights)
  %091_convolutional_bn = BatchNormalization[epsilon = 1e-05, momentum = 0.99](%091_convolutional, %091_convolutional_bn_scale, %091_convolutional_bn_bias, %091_convolutional_bn_mean, %091_convolutional_bn_var)
  %091_convolutional_lrelu = LeakyRelu[alpha = 0.1](%091_convolutional_bn)
  %092_convolutional = Conv[auto_pad = u'SAME_LOWER', dilations = [1, 1], kernel_shape = [1, 1], strides = [1, 1]](%091_convolutional_lrelu, %092_convolutional_conv_weights)
  %092_convolutional_bn = BatchNormalization[epsilon = 1e-05, momentum = 0.99](%092_convolutional, %092_convolutional_bn_scale, %092_convolutional_bn_bias, %092_convolutional_bn_mean, %092_convolutional_bn_var)
  %092_convolutional_lrelu = LeakyRelu[alpha = 0.1](%092_convolutional_bn)
  %093_convolutional = Conv[auto_pad = u'SAME_LOWER', dilations = [1, 1], kernel_shape = [3, 3], strides = [1, 1]](%092_convolutional_lrelu, %093_convolutional_conv_weights)
  %093_convolutional_bn = BatchNormalization[epsilon = 1e-05, momentum = 0.99](%093_convolutional, %093_convolutional_bn_scale, %093_convolutional_bn_bias, %093_convolutional_bn_mean, %093_convolutional_bn_var)
  %093_convolutional_lrelu = LeakyRelu[alpha = 0.1](%093_convolutional_bn)
  %094_convolutional = Conv[auto_pad = u'SAME_LOWER', dilations = [1, 1], kernel_shape = [1, 1], strides = [1, 1]](%093_convolutional_lrelu, %094_convolutional_conv_weights, %094_convolutional_conv_bias)
  %097_convolutional = Conv[auto_pad = u'SAME_LOWER', dilations = [1, 1], kernel_shape = [1, 1], strides = [1, 1]](%092_convolutional_lrelu, %097_convolutional_conv_weights)
  %097_convolutional_bn = BatchNormalization[epsilon = 1e-05, momentum = 0.99](%097_convolutional, %097_convolutional_bn_scale, %097_convolutional_bn_bias, %097_convolutional_bn_mean, %097_convolutional_bn_var)
  %097_convolutional_lrelu = LeakyRelu[alpha = 0.1](%097_convolutional_bn)
  %098_upsample = Upsample[mode = u'nearest'](%097_convolutional_lrelu, %098_upsample_scale)
  %099_route = Concat[axis = 1](%098_upsample, %037_shortcut)
  %100_convolutional = Conv[auto_pad = u'SAME_LOWER', dilations = [1, 1], kernel_shape = [1, 1], strides = [1, 1]](%099_route, %100_convolutional_conv_weights)
  %100_convolutional_bn = BatchNormalization[epsilon = 1e-05, momentum = 0.99](%100_convolutional, %100_convolutional_bn_scale, %100_convolutional_bn_bias, %100_convolutional_bn_mean, %100_convolutional_bn_var)
  %100_convolutional_lrelu = LeakyRelu[alpha = 0.1](%100_convolutional_bn)
  %101_convolutional = Conv[auto_pad = u'SAME_LOWER', dilations = [1, 1], kernel_shape = [3, 3], strides = [1, 1]](%100_convolutional_lrelu, %101_convolutional_conv_weights)
  %101_convolutional_bn = BatchNormalization[epsilon = 1e-05, momentum = 0.99](%101_convolutional, %101_convolutional_bn_scale, %101_convolutional_bn_bias, %101_convolutional_bn_mean, %101_convolutional_bn_var)
  %101_convolutional_lrelu = LeakyRelu[alpha = 0.1](%101_convolutional_bn)
  %102_convolutional = Conv[auto_pad = u'SAME_LOWER', dilations = [1, 1], kernel_shape = [1, 1], strides = [1, 1]](%101_convolutional_lrelu, %102_convolutional_conv_weights)
  %102_convolutional_bn = BatchNormalization[epsilon = 1e-05, momentum = 0.99](%102_convolutional, %102_convolutional_bn_scale, %102_convolutional_bn_bias, %102_convolutional_bn_mean, %102_convolutional_bn_var)
  %102_convolutional_lrelu = LeakyRelu[alpha = 0.1](%102_convolutional_bn)
  %103_convolutional = Conv[auto_pad = u'SAME_LOWER', dilations = [1, 1], kernel_shape = [3, 3], strides = [1, 1]](%102_convolutional_lrelu, %103_convolutional_conv_weights)
  %103_convolutional_bn = BatchNormalization[epsilon = 1e-05, momentum = 0.99](%103_convolutional, %103_convolutional_bn_scale, %103_convolutional_bn_bias, %103_convolutional_bn_mean, %103_convolutional_bn_var)
  %103_convolutional_lrelu = LeakyRelu[alpha = 0.1](%103_convolutional_bn)
  %104_convolutional = Conv[auto_pad = u'SAME_LOWER', dilations = [1, 1], kernel_shape = [1, 1], strides = [1, 1]](%103_convolutional_lrelu, %104_convolutional_conv_weights)
  %104_convolutional_bn = BatchNormalization[epsilon = 1e-05, momentum = 0.99](%104_convolutional, %104_convolutional_bn_scale, %104_convolutional_bn_bias, %104_convolutional_bn_mean, %104_convolutional_bn_var)
  %104_convolutional_lrelu = LeakyRelu[alpha = 0.1](%104_convolutional_bn)
  %105_convolutional = Conv[auto_pad = u'SAME_LOWER', dilations = [1, 1], kernel_shape = [3, 3], strides = [1, 1]](%104_convolutional_lrelu, %105_convolutional_conv_weights)
  %105_convolutional_bn = BatchNormalization[epsilon = 1e-05, momentum = 0.99](%105_convolutional, %105_convolutional_bn_scale, %105_convolutional_bn_bias, %105_convolutional_bn_mean, %105_convolutional_bn_var)
  %105_convolutional_lrelu = LeakyRelu[alpha = 0.1](%105_convolutional_bn)
  %106_convolutional = Conv[auto_pad = u'SAME_LOWER', dilations = [1, 1], kernel_shape = [1, 1], strides = [1, 1]](%105_convolutional_lrelu, %106_convolutional_conv_weights, %106_convolutional_conv_bias)
  return %082_convolutional, %094_convolutional, %106_convolutional
}

yolov3.onnx ファイルが生成されたか確認しましょう。⁵

推論の実行

$ python2 onnx_to_tensorrt.py
Downloading from https://github.com/pjreddie/darknet/raw/f86901f6177dfc6116360a13cc06ab680e0c86b0/data/dog.jpg, this may take a while...
100% [........................................................] 163759 / 163759
Loading ONNX file from path yolov3.onnx...
Beginning ONNX file parsing
Completed parsing of ONNX file
Building an engine from file yolov3.onnx; this may take a while...
Completed creating Engine
Running inference on image dog.jpg...
[[135.04631186 219.14289907 184.31729756 324.86079515]
 [ 98.95619619 135.56527022 499.10088664 299.16208427]
 [477.88941676  81.22835286 210.86738172  86.96319933]] [0.99852329 0.99881124 0.93929232] [16  1  7]
Saved image with bounding boxes of detected objects to dog_bboxes.png.

dog_bboxes.png ファイルに検出結果が表示されます。

推論コードを Python3 で実行

$ sudo apt-get install python3-pip

$ pip3 install -U numpy

$ python3 -m pip install -r requirements.txt

$ python3 onnx_to_tensorrt.py
Reading engine from file yolov3.trt
Running inference on image dog.jpg...
[[135.04631186 219.14289907 184.31729756 324.86079515]
 [ 98.95619619 135.56527022 499.10088664 299.16208427]
 [477.88941676  81.22835286 210.86738172  86.96319933]] [0.99852329 0.99881124 0.93929232] [16  1  7]
Saved image with bounding boxes of detected objects to dog_bboxes.png.

以上です。

TensorFlow や PyTorch が有名です。 ↩
ここでは重み、バイアス値に加えてニューラルネットワークの定義も含めるとします。 ↩
YOLO のアルゴリズムに関しては YOLO の解説記事を参照してください。 ↩
このページを参考にしました。 ↩
ONNX 形式のファイルは Netron などのモデル可視化ツールで表示することができます。 ↩

NVIDIA Jetson Nano 開発者キットで TensorRT サンプル・プログラムを試す