TensorFlow Object Detection API中的Faster R-CNN /SSD模型参数调整
关于TensorFlow Object Detection API配置,可以参考之前的文章https://becominghuman.ai/tensorflow-object-detection-api-tutorial-training-and-evaluating-custom-object-detector-ed2594afcf73
在本文中,我将讨论如何更改预训练模型的配置。本文的目的是您可以根据您的应用程序配置TensorFlow/models,而API将不再是一个黑盒!
本文的概述:
了解协议缓冲区和proto文件。
利用proto文件知识,我们如何了解模型的配置文件
遵循3个步骤来更新模型的参数
其他示例:
更改重量初始值设定项
更改体重优化器
评估预训练模型
协议缓冲区
要修改模型,我们需要了解它的内部机制。TensorFlow对象检测API使用协议缓冲区(Protocol Buffers),这是与语言无关,与平台无关且可扩展的机制,用于序列化结构化数据。就像XML规模较小,但更快,更简单。API使用协议缓冲区语言的proto2版本。我将尝试解释更新预配置模型所需的语言。有关协议缓冲区语言的更多详细信息,请参阅此文档和Python教程。
协议缓冲区的工作可分为以下三个步骤:
在.proto文件中定义消息格式。该文件的行为就像所有消息的蓝图一样,它显示消息所接受的所有参数是什么,参数的数据类型应该是什么,参数是必需的还是可选的,参数的标记号是什么,什么是参数的默认值等。API的protos文件可在此处找到。为了理解,我使用grid_anchor_generator.proto文件。
复制代码
syntax = "proto2";
package object_detection.protos;
// Configuration proto for GridAnchorGenerator. See
// anchor_generators/grid_anchor_generator.py for details.
message GridAnchorGenerator {
// Anchor height in pixels.
optional int32 height = 1 [default = 256];
// Anchor width in pixels.
optional int32 width = 2 [default = 256];
// Anchor stride in height dimension in pixels.
optional int32 height_stride = 3 [default = 16];
// Anchor stride in width dimension in pixels.
optional int32 width_stride = 4 [default = 16];
// Anchor height offset in pixels.
optional int32 height_offset = 5 [default = 0];
// Anchor width offset in pixels.
optional int32 width_offset = 6 [default = 0];
// At any given location, len(scales) * len(aspect_ratios) anchors are
// generated with all possible combinations of scales and aspect ratios.
// List of scales for the anchors.
repeated float scales = 7;
// List of aspect ratios for the anchors.
repeated float aspect_ratios = 8;
}
复制代码
它是从线30-33的参数明确scales,并aspect_ratios是强制性的消息GridAnchorGenerator,而参数的其余部分都是可选的,如果不通过,将采取默认值。
定义消息格式后,我们需要编译协议缓冲区。该编译器将从文件生成类.proto文件。在安装API的过程中,我们运行了以下命令,该命令将编译协议缓冲区:
# From tensorflow/models/research/
protoc object_detection/protos/*.proto --python_out=.
在定义和编译协议缓冲区之后,我们需要使用Python协议缓冲区API来写入和读取消息。在我们的例子中,我们可以将配置文件视为协议缓冲区API,它可以在不考虑TensorFlow API的内部机制的情况下写入和读取消息。换句话说,我们可以通过适当地更改配置文件来更新预训练模型的参数。
了解配置文件
显然,配置文件可以帮助我们根据需要更改模型的参数。弹出的下一个问题是如何更改模型的参数?本节和下一部分将回答这个问题,在这里proto文件的知识将很方便。出于演示目的,我正在使用faster_rcnn_resnet50_pets.config文件。
复制代码
# Faster R-CNN with Resnet-50 (v1), configured for Oxford-IIIT Pets Dataset.
# Users should configure the fine_tune_checkpoint field in the train config as
# well as the label_map_path and input_path fields in the train_input_reader and
# eval_input_reader. Search for "PATH_TO_BE_CONFIGURED" to find the fields that
# should be configured.
model {
faster_rcnn {
num_classes: 37
image_resizer {
keep_aspect_ratio_resizer {
min_dimension: 600
max_dimension: 1024
}
}
feature_extractor {
type: 'faster_rcnn_resnet50'
first_stage_features_stride: 16
}
first_stage_anchor_generator {
grid_anchor_generator {
scales: [0.25, 0.5, 1.0, 2.0]
aspect_ratios: [0.5, 1.0, 2.0]
height_stride: 16
width_stride: 16
}
}
first_stage_box_predictor_conv_hyperparams {
op: CONV
regularizer {
l2_regularizer {
weight: 0.0
}
}
initializer {
truncated_normal_initializer {
stddev: 0.01
}
}
}
first_stage_nms_score_threshold: 0.0
first_stage_nms_iou_threshold: 0.7
first_stage_max_proposals: 300
first_stage_localization_loss_weight: 2.0
first_stage_objectness_loss_weight: 1.0
initial_crop_size: 14
maxpool_kernel_size: 2
maxpool_stride: 2
second_stage_box_predictor {
mask_rcnn_box_predictor {
use_dropout: false
dropout_keep_probability: 1.0
fc_hyperparams {
op: FC
regularizer {
l2_regularizer {
weight: 0.0
}
}
initializer {
variance_scaling_initializer {
factor: 1.0
uniform: true
mode: FAN_AVG
}
}
}
}
}
second_stage_post_processing {
batch_non_max_suppression {
score_threshold: 0.0
iou_threshold: 0.6
max_detections_per_class: 100
max_total_detections: 300
}
score_converter: SOFTMAX
}
second_stage_localization_loss_weight: 2.0
second_stage_classification_loss_weight: 1.0
}
}
train_config: {
batch_size: 1
optimizer {
momentum_optimizer: {
learning_rate: {
manual_step_learning_rate {
initial_learning_rate: 0.0003
schedule {
step: 900000
learning_rate: .00003
}
schedule {
step: 1200000
learning_rate: .000003
}
}
}
momentum_optimizer_value: 0.9
}
use_moving_average: false
}
gradient_clipping_by_norm: 10.0
fine_tune_checkpoint: "PATH_TO_BE_CONFIGURED/model.ckpt"
from_detection_checkpoint: true
# Note: The below line limits the training process to 200K steps, which we
# empirically found to be sufficient enough to train the pets dataset. This
# effectively bypasses the learning rate schedule (the learning rate will
# never decay). Remove the below line to train indefinitely.
num_steps: 200000
data_augmentation_options {
random_horizontal_flip {
}
}
max_number_of_boxes: 50
}
train_input_reader: {
tf_record_input_reader {
input_path: "PATH_TO_BE_CONFIGURED/pet_train.record"
}
label_map_path: "PATH_TO_BE_CONFIGURED/pet_label_map.pbtxt"
}
eval_config: {
num_examples: 2000
# Note: The below line limits the evaluation process to 10 evaluations.
# Remove the below line to evaluate indefinitely.
max_evals: 10
}
eval_input_reader: {
tf_record_input_reader {
input_path: "PATH_TO_BE_CONFIGURED/pet_val.record"
}
label_map_path: "PATH_TO_BE_CONFIGURED/pet_label_map.pbtxt"
shuffle: false
num_readers: 1
}
复制代码
第7至10行表示这num_classes是faster_rcnnmessage 的参数之一,而后者又是message的参数model。同样,optimizer是父train_config消息的子消息,而message的batch_size另一个参数train_config。我们可以通过签出相应的protos文件来验证这一点。
复制代码
syntax = "proto2";
package object_detection.protos;
import "object_detection/protos/anchor_generator.proto";
import "object_detection/protos/box_predictor.proto";
import "object_detection/protos/hyperparams.proto";
import "object_detection/protos/image_resizer.proto";
import "object_detection/protos/losses.proto";
import "object_detection/protos/post_processing.proto";
// Configuration for Faster R-CNN models.
// See meta_architectures/faster_rcnn_meta_arch.py and models/model_builder.py
//
// Naming conventions:
// Faster R-CNN models have two stages: a first stage region proposal network
// (or RPN) and a second stage box classifier. We thus use the prefixes
// `first_stage_` and `second_stage_` to indicate the stage to which each
// parameter pertains when relevant.
message FasterRcnn {
// Whether to construct only the Region Proposal Network (RPN).
optional int32 number_of_stages = 1 [default=2];
// Number of classes to predict.
optional int32 num_classes = 3;
// Image resizer for preprocessing the input image.
optional ImageResizer image_resizer = 4;
复制代码
从第20行和第26行可以明显看出,这num_classes是optional消息的参数之一faster_rcnn。我希望到目前为止的讨论有助于理解配置文件的组织。现在,是时候正确更新模型的参数之一了。
步骤1:确定要更新的参数
假设我们需要更新fast_rcnn_resnet50_pets.config文件的image_resizer第10行中提到的参数。
步骤2:在存储库中搜索给定参数
目标是找到proto参数文件。为此,我们需要在存储库中搜索。
我们需要搜索以下代码:
复制代码
parameter_name path:research/object_detection/protos
#in our case parameter_name="image_resizer" thus,
image_resizer path:research/object_detection/protos
复制代码
在此path:research/object_detection/protos限制搜索域。在此处可以找到有关如何在GitHub上搜索的更多信息。搜索的输出image_resizer path:research/object_detection/protos如下所示:
从输出中很明显,要更新image_resizer参数,我们需要分析image_resizer.proto文件。
步骤3:分析proto档案
复制代码
syntax = "proto2";
package object_detection.protos;
// Configuration proto for image resizing operations.
// See builders/image_resizer_builder.py for details.
message ImageResizer {
oneof image_resizer_oneof {
KeepAspectRatioResizer keep_aspect_ratio_resizer = 1;
FixedShapeResizer fixed_shape_resizer = 2;
}
}
// Enumeration type for image resizing methods provided in TensorFlow.
enum ResizeType {
BILINEAR = 0; // Corresponds to tf.image.ResizeMethod.BILINEAR
NEAREST_NEIGHBOR = 1; // Corresponds to tf.image.ResizeMethod.NEAREST_NEIGHBOR
BICUBIC = 2; // Corresponds to tf.image.ResizeMethod.BICUBIC
AREA = 3; // Corresponds to tf.image.ResizeMethod.AREA
}
// Configuration proto for image resizer that keeps aspect ratio.
message KeepAspectRatioResizer {
// Desired size of the smaller image dimension in pixels.
optional int32 min_dimension = 1 [default = 600];
// Desired size of the larger image dimension in pixels.
optional int32 max_dimension = 2 [default = 1024];
// Desired method when resizing image.
optional ResizeType resize_method = 3 [default = BILINEAR];
// Whether to pad the image with zeros so the output spatial size is
// [max_dimension, max_dimension]. Note that the zeros are padded to the
// bottom and the right of the resized image.
optional bool pad_to_max_dimension = 4 [default = false];
// Whether to also resize the image channels from 3 to 1 (RGB to grayscale).
optional bool convert_to_grayscale = 5 [default = false];
// Per-channel pad value. This is only used when pad_to_max_dimension is True.
// If unspecified, a default pad value of 0 is applied to all channels.
repeated float per_channel_pad_value = 6;
}
// Configuration proto for image resizer that resizes to a fixed shape.
message FixedShapeResizer {
// Desired height of image in pixels.
optional int32 height = 1 [default = 300];
// Desired width of image in pixels.
optional int32 width = 2 [default = 300];
// Desired method when resizing image.
optional ResizeType resize_method = 3 [default = BILINEAR];
// Whether to also resize the image channels from 3 to 1 (RGB to grayscale).
optional bool convert_to_grayscale = 4 [default = false];
}
复制代码
从第8-10行可以看出,我们可以使用keep_aspect_ratio_resizer或调整图像的大小fixed_shape_resizer。在分析行23-44,我们可以观察到的消息keep_aspect_ratio_resizer有参数:min_dimension,max_dimension,resize_method,pad_to_max_dimension,convert_to_grayscale,和per_channel_pad_value。此外,fixed_shape_resizer有参数:height,width,resize_method,和convert_to_grayscale。proto文件中提到了所有参数的数据类型。因此,要更改image_resizer类型,我们可以在配置文件中更改以下几行。
复制代码
#before
image_resizer {
keep_aspect_ratio_resizer {
min_dimension: 600
max_dimension: 1024
}
}
#after
image_resizer {
fixed_shape_resizer {
height: 600
width: 500
resize_method: AREA
}
}
复制代码
上面的代码将使用AREA调整大小方法将图像调整为500 * 600。TensorFlow中可用的各种调整大小的方法可以在这里找到。
其他例子
我们可以使用上一节中讨论的步骤更新/添加任何参数。我将在此处演示一些经常使用的示例,但是上面讨论的步骤可能有助于更新/添加模型的任何参数。
更改重量初始化器
决定更改fast_rcnn_resnet50_pets.config文件的initializer第35行的参数。
initializer path:research/object_detection/protos在存储库中搜索。根据搜索结果,很明显我们需要分析hyperparams.proto文件。
hyperparams.proto文件中的第68–74行说明了initializer配置。
复制代码
message Initializer {
oneof initializer_oneof {
TruncatedNormalInitializer truncated_normal_initializer = 1;
VarianceScalingInitializer variance_scaling_initializer = 2;
RandomNormalInitializer random_normal_initializer = 3;
}
}
复制代码
我们可以使用random_normal_intializer代替truncated_normal_initializer,因为我们需要分析hyperparams.proto文件中的第99–102行。
message RandomNormalInitializer {
optional float mean = 1 [default = 0.0];
optional float stddev = 2 [default = 1.0];
}
显然random_normal_intializer有两个参数mean和stddev。我们可以将配置文件中的以下几行更改为use random_normal_intializer。
复制代码
#before
initializer {
truncated_normal_initializer {
stddev: 0.01
}
}
#after
initializer {
random_normal_intializer{
mean: 1
stddev: 0.5
}
}
复制代码
更改体重优化器
决定更改faster_rcnn_resnet50_pets.config文件的第87行momentum_optimizer的父消息的参数。optimizer
optimizer path:research/object_detection/protos在存储库中搜索。根据搜索结果,很明显我们需要分析optimizer.proto文件。
optimizer.proto文件中的9-14行,解释optimizer配置。
复制代码
message Optimizer {
oneof optimizer {
RMSPropOptimizer rms_prop_optimizer = 1;
MomentumOptimizer momentum_optimizer = 2;
AdamOptimizer adam_optimizer = 3;
}
复制代码
显然,代替momentum_optimizer我们可以使用adam_optimizer已被证明是良好的优化程序。为此,我们需要在f aster_rcnn_resnet50_pets.config文件中进行以下更改。
复制代码
#before
optimizer {
momentum_optimizer: {
learning_rate: {
manual_step_learning_rate {
initial_learning_rate: 0.0003
schedule {
step: 900000
learning_rate: .00003
}
schedule {
step: 1200000
learning_rate: .000003
}
}
}
momentum_optimizer_value: 0.9
}
#after
optimizer {
adam_optimizer: {
learning_rate: {
manual_step_learning_rate {
initial_learning_rate: 0.0003
schedule {
step: 900000
learning_rate: .00003
}
schedule {
step: 1200000
learning_rate: .000003
}
}
}
}
复制代码
评估预训练模型
Eval等待300秒,以检查训练模型是否已更新!如果您的GPU不错,那么您可以同时进行训练和评估!通常,资源将被耗尽。为了克服这个问题,我们可以先训练模型,将其保存在目录中,然后再评估模型。为了稍后进行评估,我们需要在配置文件中进行以下更改:
复制代码
#Before
eval_config: {
num_examples: 2000
# Note: The below line limits the evaluation process to 10 evaluations.
# Remove the below line to evaluate indefinitely.
max_evals: 10
}
#after
eval_config: {