• home
  • about
  • 全ての投稿
  • ソフトウェア・ハードウェアの設定のまとめ
  • 分析関連のまとめ
  • ヘルスケア関連のまとめ
  • 生涯学習関連のまとめ

TensorFlow Inceptionまとめ

date: 2017-01-12 excerpt: TensorFlow inception v3転移学習まとめ

tag: inception

ボトルネックを利用した転移学習

p2 tensorflow/tensorflow/examples/image_retraining/retrain.py  --bottleneck_dir $HOME/sdb/food-classify-bottlenecks  --how_many_training_steps 10000  --model_dir $HOME/sdb/food-classify-train  --output_graph $HOME/sdb/food-classify-easytrain/retrained_graph.pb  --output_labels $HOME/sdb/food-classify-easytrain/retrained_labels.txt  --image_dir $HOME/sdb/food-classify-data/raw-data/train/

2017-01-12 21:07:33.861809: precision @ 1 = 0.7317 recall @ 5 = 0.9707 [50016 examples

evalは基本的に学習の進捗を監視するウォッチドッグのような使い方をするものらしい

TensorFlowのバグ？modelが消える

引数にホームディレクトリである、~を混入するとファイルが消えることがあるようなので、基本的に~は使わず、絶対パスで記述する
eval

bazel-bin/inception/flowers_eval --checkpoint_dir $HOME/sdb/food-classify-train --eval_dir=$HOME/sdb/flowers-data --data_dir $HOME/sdb/food-classify-data

例2

bazel-bin/inception/flowers_eval2 --checkpoint_dir $HOME/sdb/food-classify-train --eval_dir=$HOME/sdb/flowers-data --data_dir $HOME/sdb/food-classify-data

Optional : ワーニングを止める

tf.logging.set_verbosity(tf.logging.ERROR)

bazel-bin/inception/flowers_train   --train_dir=~/sda/flowers-train/   --data_dir ~/sda/flowers-data/   --pretrained_model_checkpoint_path ~/sda/inception-v3/model.ckpt-157585   --fine_tune=True   --initial_learning_rate=0.001 --input_queue_memory_factor=1
  1. tensorflow/modelからinceptionを含んだ学習モデルをダウンロードする 
    wget https://github.com/tensorflow/models/tree/master/inception
  2. もとデータのダウンロードと、学習スクリプトのビルド 
    export DATA_DIR=$HOME/sdb/imagenet-data
cd ./models/inception/

build and download

bazel build inception/download_and_preprocess_imagenet
bazel-bin/inception/download_and_preprocess_imagenet ${DATA_DIR}

build flowers

bazel build inception/download_and_preprocess_flowers
bazel-bin/inception/download_and_preprocess_flowers ${DATA_DIR}

build imagenet

bazel build inception/imagenet_train 
bazel-bin/inception/imagenet_train --num_gpus=1 --batch_size=32 --train_dir=$HOME/sdb/imagenet_data --data_dir=/tmp/imagenet_data

train flowers

$ bazel build inception/flowers_train
$ bazel-bin/inception/flowers_train --num_gpus=1 --batch_size=32 --train_dir=$HOME
$ /sdb/flowers_data --data_dir=$HOME/sdb/tmp/flowers_data
  1. ダウンロードした画像データのシリアライズ 
    $ bazel-bin/inception/download_and_preprocess_flowers.runfiles/ 
$ python ~/sdb/models/inception/inception/data/build_image_data.py --train_directory=$HOME/sdb/flowers-data/raw-data/train/ $ validation_directory=$HOME/sdb/flowers-data/raw-data/validation/ --output_directory= --labels_file=$HOME/sdb/flowers-data/raw-data//labels.txt
  2. 実行 
    bazel-bin/inception/flowers_train --num_gpus=1 --batch_size=32 --train_dir ~/sdb/food-classify-out --data_dir ~/sdb/food-classify-data/
  3. 評価 
    bazel-bin/inception/flowers_eval --eval_dir=$HOME/food-classify-data/raw-data/validation --data_dir=$HOME/food-classify-data --subset=validation --num_examples=500 --checkpoint_dir=$HOME/sdb/food-classify-out --input_queue_memory_factor=1 --run_once
  4. resume 
    bazel-bin/inception/flowers_train --num_gpus 1 --batch_size=32 --train_dir ~/sdb/food-classify-out --data_dir ~/sdb/food-classify-data/ --checkpoint_dir ~/sdb/food-classify-out/model.ckpt-25000

inception Share Tweet