TensorFlow 安装及使用

安装

（1）安装包安装：pip install tensorflow==1.14 -i https://pypi.douban.com/simple

virtualenv -p /usr/bin/python2.7 venv-python2.7-tf1.14.0
source ./venv-python2.7-tf1.14.0/bin/activate
pip list
python
pip install numpy==1.16.5 opt-einsum==2.3.2 future -i https://pypi.douban.com/simple
pip install tensorflow==1.14.0 -i https://pypi.douban.com/simple

（2）源码编译安装：https://tensorflow.google.cn/install/source

Install bazel-0.25.2
# wget https://github.com/bazelbuild/bazel/releases/download/0.25.2/bazel-0.25.2-linux-x86_64
# chmod u+x bazel-0.25.2-linux-x86_64
# ln -s /path/bazel-0.25.2-linux-x86_64 /usr/bin/bazel
# bazel version
Build label: 0.25.2

Install tensorflow-1.14.0
# git clone https://github.com/tensorflow/tensorflow.git
# cd tensorflow
# git checkout v1.14.0
# ./configure   # /usr/bin/python3, others are default
# bazel build --config=opt //tensorflow/tools/pip_package:build_pip_package
# ./bazel-bin/tensorflow/tools/pip_package/build_pip_package /tmp/tensorflow_pkg
# mv /tmp/tensorflow_pkg/tensorflow-1.14.0-cp36-cp36m-linux_x86_64.whl ./
# python3 -m pip install ./tensorflow-1.14.0-cp36-cp36m-linux_x86_64.whl

install tensorflow-1.14.0 with MKL&Patch
# git clone https://github.com/tensorflow/tensorflow.git
# cd tensorflow/
# git checkout v1.14.0
# patch -p0 < /path/tf-mkl.patch
# ./configure   # /usr/bin/python3, others are default
# bazel build --config=mkl --config=opt //tensorflow/tools/pip_package:build_pip_package
# ./bazel-bin/tensorflow/tools/pip_package/build_pip_package /tmp/tensorflow_pkg
# mv /tmp/tensorflow_pkg/tensorflow-1.14.0-cp36-cp36m-linux_x86_64.whl ./
# python3 -m pip install ./tensorflow-1.14.0-cp36-cp36m-linux_x86_64.whl

# python3 -m pip list
tensorboard (1.14.0)
tensorflow (1.14.0)
tensorflow-estimator (1.14.0)

使用

模型优化

（1）查看 saved_model 模型的输入和输出

# bazel build tensorflow/python/tools:saved_model_cli
# saved_model_cli show --dir detection/ --all
或者
# python3 /usr/local/lib/python3.6/site-packages/tensorflow/python/tools/saved_model_cli.py show --dir detection/ --all
MetaGraphDef with tag-set: 'serve' contains the following SignatureDefs:
signature_def['serving_default']:
  The given SavedModel SignatureDef contains the following input(s):
    inputs['image'] tensor_info:
        dtype: DT_UINT8
        shape: (1, -1, -1, 3)
        name: image:0
    inputs['true_image_shape'] tensor_info:
        dtype: DT_INT32
        shape: (1, 3)
        name: true_image_shape:0
  The given SavedModel SignatureDef contains the following output(s):
    outputs['detection_boxes'] tensor_info:
        dtype: DT_FLOAT
        shape: (1, -1, 4)
        name: ChangeCoordToOriginalImage/stack:0
    outputs['detection_classes'] tensor_info:
        dtype: DT_INT32
        shape: (1, -1)
        name: add:0
    outputs['detection_keypoints'] tensor_info:
        dtype: DT_FLOAT
        shape: (1, -1, 4, 2)
        name: TextKeypointPostProcess/Reshape_2:0
    outputs['detection_scores'] tensor_info:
        dtype: DT_FLOAT
        shape: (1, -1)
        name: strided_slice_3:0
    outputs['num_detections'] tensor_info:
        dtype: DT_INT32
        shape: (1)
        name: BatchMultiClassNonMaxSuppression/stack_8:0
  Method name is: tensorflow/serving/predict

（2）将 tf 的 saved_model 保存成 frozen_model

# bazel build tensorflow/python/tools:freeze_graph
# freeze_graph --input_saved_model_dir detection/ --output_graph detection_frozen_model.pb --output_node_names ChangeCoordToOriginalImage/stack,add,TextKeypointPostProcess/Reshape_2,strided_slice_3,BatchMultiClassNonMaxSuppression/stack_8
或者
# python3 /usr/local/lib/python3.6/site-packages/tensorflow/python/tools/freeze_graph.py --input_saved_model_dir detection/ --output_graph detection_frozen_model.pb --output_node_names ChangeCoordToOriginalImage/stack,add,TextKeypointPostProcess/Reshape_2,strided_slice_3,BatchMultiClassNonMaxSuppression/stack_8

（3）将 frozen_model 通过优化得到 optimized_model

# bazel build tensorflow/python/tools:optimize_for_inference   // ouput: bazel-bin/tensorflow/python/tools/optimize_for_inference
# optimize_for_inference --input detection_frozen_model.pb --output detection_optimized_model.pb --input_names image,true_image_shape --output_names ChangeCoordToOriginalImage/stack,add,TextKeypointPostProcess/Reshape_2,strided_slice_3,BatchMultiClassNonMaxSuppression/stack_8 --frozen_graph true --placeholder_type_enum 4,3,1,3,1,1,3
或者
# python3 /usr/local/lib/python3.6/site-packages/tensorflow/python/tools/optimize_for_inference.py --input detection_frozen_model.pb --output detection_optimized_model.pb --input_names image,true_image_shape --output_names ChangeCoordToOriginalImage/stack,add,TextKeypointPostProcess/Reshape_2,strided_slice_3,BatchMultiClassNonMaxSuppression/stack_8 --frozen_graph true --placeholder_type_enum 4,3,1,3,1,1,3

其中 placeholder_type_enum 详情如下:

https://github.com/tensorflow/tensorflow/blob/master/tensorflow/core/framework/types.proto

（4）将 pb 模型输出成 TensorFlow 的可视化 graph

# bazel build tensorflow/python/tools:import_pb_to_tensorboard
# import_pb_to_tensorboard --model_dir ./recognition_frozen_model.pb --log_dir ./recognition_log
或者
# python3 /usr/local/lib/python3.6/site-packages/tensorflow/python/tools/import_pb_to_tensorboard.py --model_dir ./recognition_frozen_model.pb --log_dir ./recognition_frozen_model.graph
# nohup tensorboard --logdir=./recognition_frozen_model.graph --port=6006 2>&1 &

可视化工具 TensorBoard 用法: https://blog.csdn.net/gg_18826075157/article/details/78440766

（5）量化、固化、优化 pb 模型

官方手册：https://github.com/tensorflow/tensorflow/tree/master/tensorflow/tools/graph_transforms

intel 量化手册：https://github.com/IntelAI/tools/tree/master/tensorflow_quantization

# bazel build tensorflow/tools/graph_transforms:transform_graph
# bazel-bin/tensorflow/tools/graph_transforms/transform_graph --in_graph="./detection_frozen_model.pb" --out_graph="./detection_transformed_model.pb" --inputs="image,true_image_shape" --outputs="ChangeCoordToOriginalImage/stack,add,TextKeypointPostProcess/Reshape_2,strided_slice_3,BatchMultiClassNonMaxSuppression/stack_8" --transforms='
  add_default_attributes
  strip_unused_nodes()
  remove_nodes(op=Identity, op=CheckNumerics)
  fold_constants(ignore_errors=true)
  fold_batch_norms
  fold_old_batch_norms
  quantize_weights'

PS：

模型优化 refer：https://blog.csdn.net/qq_14845119/article/details/78846372

模型量化：https://www.jianshu.com/p/d2637646cda1

tf的log和vlog输出配置

There are two flags, similarly named, but with somewhat different semantics:

TF_CPP_MIN_LOG_LEVEL - which has 3 or 4 basic levels - low numbers = more messages.

0 outputs Information, Warning, Error, and Fatals (default)

1 outputs Warning, and above

2 outputs Errors and above.

etc... I didn't check edge cases

TF_CPP_MIN_VLOG_LEVEL - which causes very very many extra Information errors - really for debugging only - low numbers = less messages.

3 Outputs lots and lots of stuff

2 Outputs less

1 Outputs even less

0 Outputs nothing extra (default)