tensorflow训练模型配置文件详解

2019年1月4日

参考官方：Configuring the Object Detection Training Pipeline

Tensorflow目标检测API使用protobuf文件来配置训练和测试相关的东西。

文件格式可以参照object_detection/protos/pipeline.proto文件：

import "object_detection/protos/eval.proto";
import "object_detection/protos/graph_rewriter.proto";
import "object_detection/protos/input_reader.proto";
import "object_detection/protos/model.proto";
import "object_detection/protos/train.proto";

import "object_detection/protos/eval.proto";

import "object_detection/protos/graph_rewriter.proto";

import "object_detection/protos/input_reader.proto";

import "object_detection/protos/model.proto";

import "object_detection/protos/train.proto";

model_config。这里定义训练模型类型和使用参数。
model_config.proto

syntax = <span class="hljs-string">"proto2"</span>;

<span class="hljs-keyword">package</span> object_detection.protos;

<span class="hljs-keyword">import</span> <span class="hljs-string">"object_detection/protos/faster_rcnn.proto"</span>;
<span class="hljs-keyword">import</span> <span class="hljs-string">"object_detection/protos/ssd.proto"</span>;

<span class="hljs-comment">// Top level configuration for DetectionModels.</span>
message DetectionModel {
  oneof model {
    FasterRcnn faster_rcnn = <span class="hljs-number">1</span>;
    Ssd ssd = <span class="hljs-number">2</span>;
  }
}

syntax = "proto2";

package object_detection.protos;

import "object_detection/protos/faster_rcnn.proto";

import "object_detection/protos/ssd.proto";

// Top level configuration for DetectionModels.

message DetectionModel {

oneof model {

FasterRcnn faster_rcnn = 1;

Ssd ssd = 2;

}

train_config，定义哪些参数应该被用来训练模型参数。
train_config.proto

syntax = <span class="hljs-string">"proto2"</span>;

<span class="hljs-keyword">package</span> object_detection.protos;

<span class="hljs-keyword">import</span> <span class="hljs-string">"object_detection/protos/optimizer.proto"</span>;
<span class="hljs-keyword">import</span> <span class="hljs-string">"object_detection/protos/preprocessor.proto"</span>;

<span class="hljs-comment">// Message for configuring DetectionModel training jobs (train.py).</span>
message TrainConfig {
  <span class="hljs-comment">// Input queue batch size.</span>
  optional uint32 batch_size = <span class="hljs-number">1</span> [<span class="hljs-keyword">default</span>=<span class="hljs-number">32</span>];

  <span class="hljs-comment">// Data augmentation options.</span>
  repeated PreprocessingStep data_augmentation_options = <span class="hljs-number">2</span>;

syntax = "proto2";

package object_detection.protos;

import "object_detection/protos/optimizer.proto";

import "object_detection/protos/preprocessor.proto";

// Message for configuring DetectionModel training jobs (train.py).

message TrainConfig {

// Input queue batch size.

optional uint32 batch_size = 1 [default=32];

// Data augmentation options.

repeated PreprocessingStep data_augmentation_options = 2;

eval_config，设定的使用哪些指标将被报告进行评估。

syntax = <span class="hljs-string">"proto2"</span>;

<span class="hljs-keyword">package</span> object_detection.protos;

<span class="hljs-comment">// Message for configuring DetectionModel evaluation jobs (eval.py).</span>
message EvalConfig {
  <span class="hljs-comment">// Number of visualization images to generate.</span>
  optional uint32 num_visualizations = <span class="hljs-number">1</span> [<span class="hljs-keyword">default</span>=<span class="hljs-number">10</span>];

  <span class="hljs-comment">// Number of examples to process of evaluation.</span>
  optional uint32 num_examples = <span class="hljs-number">2</span> [<span class="hljs-keyword">default</span>=<span class="hljs-number">5000</span>];

syntax = "proto2";

package object_detection.protos;

// Message for configuring DetectionModel evaluation jobs (eval.py).

message EvalConfig {

// Number of visualization images to generate.

optional uint32 num_visualizations = 1 [default=10];

// Number of examples to process of evaluation.

optional uint32 num_examples = 2 [default=5000];

train_input_config，定义在哪些数据集是训练集。
eval_input_config，定义在哪些数据集进行评估的。通常，这应该与训练输入数据集不同。

配置文件分为5个部分：

model { 
(… 在这里添加model的配置…) 
} 
train_config : { 
(… Add train_config here…) 
} 
train_input_reader: { 
(… Add train_input configuration here…) 
} 
eval_config: { 
} 
eval_input_reader: { 
(… Add eval_input configuration here…) 
}