Tensorflow object detection API源码分析之如何构建模型
模型的具体参数被定义在config文件中，如samples/configs/ssd_mobilenet_v2_coco.config
model {
ssd {
num_classes: 90
box_coder {
faster_rcnn_box_coder {
y_scale: 10.0
x_scale: 10.0
height_scale: 5.0
width_scale: 5.0
}
}
matcher {
argmax_matcher {
matched_threshold: 0.5
unmatched_threshold: 0.5
ignore_thresholds: false
negatives_lower_than_unmatched: true
force_match_for_each_row: true
}
}
similarity_calculator {
iou_similarity {
}
}
anchor_generator {
ssd_anchor_generator {
num_layers: 6
min_scale: 0.2
max_scale: 0.95
aspect_ratios: 1.0
aspect_ratios: 2.0
aspect_ratios: 0.5
aspect_ratios: 3.0
aspect_ratios: 0.3333
}
}
image_resizer {
fixed_shape_resizer {
height: 300
width: 300
}
}
box_predictor {
convolutional_box_predictor {
min_depth: 0
max_depth: 0
num_layers_before_predictor: 0
use_dropout: false
dropout_keep_probability: 0.8
kernel_size: 1
box_code_size: 4
apply_sigmoid_to_scores: false
conv_hyperparams {
activation: RELU_6,
regularizer {
l2_regularizer {
weight: 0.00004
}
}
initializer {
truncated_normal_initializer {
stddev: 0.03
mean: 0.0
}
}
batch_norm {
train: true,
scale: true,
center: true,
decay: 0.9997,
epsilon: 0.001,
}
}
}
}
feature_extractor {
type: 'ssd_mobilenet_v2'
min_depth: 16
depth_multiplier: 1.0
conv_hyperparams {
activation: RELU_6,
regularizer {
l2_regularizer {
weight: 0.00004
}
}
initializer {
truncated_normal_initializer {
stddev: 0.03
mean: 0.0
}
}
batch_norm {
train: true,
scale: true,
center: true,
decay: 0.9997,
epsilon: 0.001,
}
}
}train.py中读入config中关于模型的配置，通过以下方式 创建 检测模型
model_config = configs['model']
train_config = configs['train_config']
input_config = configs['train_input_config']
model_fn = functools.partial(
model_builder.build,
model_config=model_config,
is_training=True)在builders/model_builder.py（A function to build a DetectionModel from configuration.）
import functools
#导入构建模型的各个组件，有各种feature_extractor以及其他
from object_detection.builders import anchor_generator_builder
from object_detection.builders import box_coder_builder
from object_detection.builders import box_predictor_builder
from object_detection.builders import hyperparams_builder
from object_detection.builders import image_resizer_builder
from object_detection.builders import losses_builder
from object_detection.builders import matcher_builder
from object_detection.builders import post_processing_builder
from object_detection.builders import region_similarity_calculator_builder as sim_calc
from object_detection.core import balanced_positive_negative_sampler as sampler
from object_detection.core import post_processing
from object_detection.core import target_assigner
from object_detection.meta_architectures import faster_rcnn_meta_arch
from object_detection.meta_architectures import rfcn_meta_arch
from object_detection.meta_architectures import ssd_meta_arch
from object_detection.models import faster_rcnn_inception_resnet_v2_feature_extractor as frcnn_inc_res
from object_detection.models import faster_rcnn_inception_v2_feature_extractor as frcnn_inc_v2
from object_detection.models import faster_rcnn_nas_feature_extractor as frcnn_nas
from object_detection.models import faster_rcnn_pnas_feature_extractor as frcnn_pnas
from object_detection.models import faster_rcnn_resnet_v1_feature_extractor as frcnn_resnet_v1
from object_detection.models import ssd_resnet_v1_fpn_feature_extractor as ssd_resnet_v1_fpn
from object_detection.models import ssd_resnet_v1_ppn_feature_extractor as ssd_resnet_v1_ppn
from object_detection.models.embedded_ssd_mobilenet_v1_feature_extractor import EmbeddedSSDMobileNetV1FeatureExtractor
from object_detection.models.ssd_inception_v2_feature_extractor import SSDInceptionV2FeatureExtractor
from object_detection.models.ssd_inception_v3_feature_extractor import SSDInceptionV3FeatureExtractor
from object_detection.models.ssd_mobilenet_v1_feature_extractor import SSDMobileNetV1FeatureExtractor
from object_detection.models.ssd_mobilenet_v1_fpn_feature_extractor import SSDMobileNetV1FpnFeatureExtractor
from object_detection.models.ssd_mobilenet_v1_ppn_feature_extractor import SSDMobileNetV1PpnFeatureExtractor
from object_detection.models.ssd_mobilenet_v2_feature_extractor import SSDMobileNetV2FeatureExtractor
from object_detection.models.ssd_mobilenet_v2_fpn_feature_extractor import SSDMobileNetV2FpnFeatureExtractor
from object_detection.models.ssd_mobilenet_v2_keras_feature_extractor import SSDMobileNetV2KerasFeatureExtractor
from object_detection.models.ssd_pnasnet_feature_extractor import SSDPNASNetFeatureExtractor
from object_detection.predictors import rfcn_box_predictor
from object_detection.predictors.heads import mask_head
from object_detection.protos import model_pb2
from object_detection.utils import ops
#SSD及Fasrer-rcnn现有特征提取器map
# A map of names to SSD feature extractors.
SSD_FEATURE_EXTRACTOR_CLASS_MAP = {
'ssd_inception_v2': SSDInceptionV2FeatureExtractor,
'ssd_inception_v3': SSDInceptionV3FeatureExtractor,
'ssd_mobilenet_v1': SSDMobileNetV1FeatureExtractor,
'ssd_mobilenet_v1_fpn': SSDMobileNetV1FpnFeatureExtractor,
'ssd_mobilenet_v1_ppn': SSDMobileNetV1PpnFeatureExtractor,
'ssd_mobilenet_v2': SSDMobileNetV2FeatureExtractor,
'ssd_mobilenet_v2_fpn': SSDMobileNetV2FpnFeatureExtractor,
'ssd_resnet50_v1_fpn': ssd_resnet_v1_fpn.SSDResnet50V1FpnFeatureExtractor,
'ssd_resnet101_v1_fpn': ssd_resnet_v1_fpn.SSDResnet101V1FpnFeatureExtractor,
'ssd_resnet152_v1_fpn': ssd_resnet_v1_fpn.SSDResnet152V1FpnFeatureExtractor,
'ssd_resnet50_v1_ppn': ssd_resnet_v1_ppn.SSDResnet50V1PpnFeatureExtractor,
'ssd_resnet101_v1_ppn':
ssd_resnet_v1_ppn.SSDResnet101V1PpnFeatureExtractor,
'ssd_resnet152_v1_ppn':
ssd_resnet_v1_ppn.SSDResnet152V1PpnFeatureExtractor,
'embedded_ssd_mobilenet_v1': EmbeddedSSDMobileNetV1FeatureExtractor,
'ssd_pnasnet': SSDPNASNetFeatureExtractor,
}
SSD_KERAS_FEATURE_EXTRACTOR_CLASS_MAP = {
'ssd_mobilenet_v2_keras': SSDMobileNetV2KerasFeatureExtractor
}
# A map of names to Faster R-CNN feature extractors.
FASTER_RCNN_FEATURE_EXTRACTOR_CLASS_MAP = {
'faster_rcnn_nas':
frcnn_nas.FasterRCNNNASFeatureExtractor,
'faster_rcnn_pnas':
frcnn_pnas.FasterRCNNPNASFeatureExtractor,
'faster_rcnn_inception_resnet_v2':
frcnn_inc_res.FasterRCNNInceptionResnetV2FeatureExtractor,
'faster_rcnn_inception_v2':
frcnn_inc_v2.FasterRCNNInceptionV2FeatureExtractor,
'faster_rcnn_resnet50':
frcnn_resnet_v1.FasterRCNNResnet50FeatureExtractor,
'faster_rcnn_resnet101':
frcnn_resnet_v1.FasterRCNNResnet101FeatureExtractor,
'faster_rcnn_resnet152':
frcnn_resnet_v1.FasterRCNNResnet152FeatureExtractor,
}
def build(model_config, is_training, add_summaries=True):
"""Builds a DetectionModel based on the model config.
Args:
model_config: A model.proto object containing the config for the desired
DetectionModel.
is_training: True if this model is being built for training purposes.
add_summaries: Whether to add tensorflow summaries in the model graph.
Returns:
DetectionModel based on the config.
Raises:
ValueError: On invalid meta architecture or model.
"""
if not isinstance(model_config, model_pb2.DetectionModel):
raise ValueError('model_config not of type model_pb2.DetectionModel.')
meta_architecture = model_config.WhichOneof('model')
#构建SSD 模型
if meta_architecture == 'ssd':
return _build_ssd_model(model_config.ssd, is_training, add_summaries)
#构建faster_rcnn模型
if meta_architecture == 'faster_rcnn':
return _build_faster_rcnn_model(model_config.faster_rcnn, is_training,
add_summaries)
raise ValueError('Unknown meta architecture: {}'.format(meta_architecture))
def _build_ssd_feature_extractor(feature_extractor_config,
is_training,
freeze_batchnorm,
reuse_weights=None):
"""Builds a ssd_meta_arch.SSDFeatureExtractor based on config.
Args:
feature_extractor_config: A SSDFeatureExtractor proto config from ssd.proto.
is_training: True if this feature extractor is being built for training.
freeze_batchnorm: Whether to freeze batch norm parameters during
training or not. When training with a small batch size (e.g. 1), it is
desirable to freeze batch norm update and use pretrained batch norm
params.
reuse_weights: if the feature extractor should reuse weights.
Returns:
ssd_meta_arch.SSDFeatureExtractor based on config.
Raises:
ValueError: On invalid feature extractor type.
"""
#ssd_mobilenet_v2
feature_type = feature_extractor_config.type
is_keras_extractor = feature_type in SSD_KERAS_FEATURE_EXTRACTOR_CLASS_MAP
#1.0
depth_multiplier = feature_extractor_config.depth_multiplier
#16
min_depth = feature_extractor_config.min_depth
pad_to_multiple = feature_extractor_config.pad_to_multiple
use_explicit_padding = feature_extractor_config.use_explicit_padding
use_depthwise = feature_extractor_config.use_depthwise
if is_keras_extractor:
conv_hyperparams = hyperparams_builder.KerasLayerHyperparams(
feature_extractor_config.conv_hyperparams)
else:
#读入卷积操作的配置"""Builds tf-slim arg_scope for convolution ops based on the config.
conv_hyperparams = hyperparams_builder.build(
feature_extractor_config.conv_hyperparams, is_training)
override_base_feature_extractor_hyperparams = (
feature_extractor_config.override_base_feature_extractor_hyperparams)
if (feature_type not in SSD_FEATURE_EXTRACTOR_CLASS_MAP) and (
not is_keras_extractor):
raise ValueError('Unknown ssd feature_extractor: {}'.format(feature_type))
if is_keras_extractor:
feature_extractor_class = SSD_KERAS_FEATURE_EXTRACTOR_CLASS_MAP[
feature_type]
else:
feature_extractor_class = SSD_FEATURE_EXTRACTOR_CLASS_MAP[feature_type]
kwargs = {
'is_training':
is_training,
'depth_multiplier':
depth_multiplier,
'min_depth':
min_depth,
'pad_to_multiple':
pad_to_multiple,
'use_explicit_padding':
use_explicit_padding,
'use_depthwise':
use_depthwise,
'override_base_feature_extractor_hyperparams':
override_base_feature_extractor_hyperparams
}
if is_keras_extractor:
kwargs.update({
'conv_hyperparams': conv_hyperparams,
'inplace_batchnorm_update': False,
'freeze_batchnorm': freeze_batchnorm
})
else:
kwargs.update({
'conv_hyperparams_fn': conv_hyperparams,
'reuse_weights': reuse_weights,
})
if feature_extractor_config.HasField('fpn'):
kwargs.update({
'fpn_min_level':
feature_extractor_config.fpn.min_level,
'fpn_max_level':
feature_extractor_config.fpn.max_level,
'additional_layer_depth':
feature_extractor_config.fpn.additional_layer_depth,
})
#将配置文件中的参数传入SSD_FEATURE_EXTRACTOR_CLASS_MAP[ssd_mobilenet_v2]-->SSDMobileNetV2FeatureExtractor中
return feature_extractor_class(**kwargs)object_detection/meta_architectures/ssd_meta_arch.py
- 功能：
- 对输入图片进行预处理。
- 输入图片预处理结果，获取特征图列表
class SSDFeatureExtractor(object):
"""SSD Feature Extractor definition."""
# 构造器中的参数大多用于构建模型的过程中
def __init__(self,
is_training,
depth_multiplier,
min_depth,
pad_to_multiple,
conv_hyperparams_fn,
reuse_weights=None,
use_explicit_padding=False,
use_depthwise=False,
override_base_feature_extractor_hyperparams=False):
# 忽略赋值语句...
# 用于预处理图片
# 输入 [batch, height, width, channels] 的float数据
# 输出 preprocessed_inputs, true_image_shapes
# preprocessed_inputs 为 [batch, height, width, channels] 的float数据
# true_image_shapes 为 [batch, 3] 的int32数据，
@abstractmethod
def preprocess(self, resized_inputs):
pass
# 用途提取特征图
# 以之前 preprocess 函数的输出作为输入，即 [batch, height, width, channels] 的float数据
# 输出 tensor 列表，每个tensor的shape是 [batch, height_i, width_i, depth_i]
@abstractmethod
def extract_features(self, preprocessed_inputs):
raise NotImplementedError以models/ssd_mobilenet_v2_feature_extractor.py为例，SSDMobileNetV2FeatureExtractor继承了父类ssd_meta_arch.SSDFeatureExtractor
"""SSDFeatureExtractor for MobilenetV2 features."""
import tensorflow as tf
from object_detection.meta_architectures import ssd_meta_arch
from object_detection.models import feature_map_generators
from object_detection.utils import context_manager
from object_detection.utils import ops
from object_detection.utils import shape_utils
from nets.mobilenet import mobilenet
from nets.mobilenet import mobilenet_v2
slim = tf.contrib.slim
#继承了父类ssd_meta_arch.SSDFeatureExtractor
class SSDMobileNetV2FeatureExtractor(ssd_meta_arch.SSDFeatureExtractor):
"""SSD Feature Extractor using MobilenetV2 features."""
def __init__(self,
is_training,
depth_multiplier,
min_depth,
pad_to_multiple,
conv_hyperparams_fn,
reuse_weights=None,
use_explicit_padding=False,
use_depthwise=False,
override_base_feature_extractor_hyperparams=False):
"""MobileNetV2 Feature Extractor for SSD Models.
Mobilenet v2 (experimental), designed by sandler@. More details can be found
in //knowledge/cerebra/brain/compression/mobilenet/mobilenet_experimental.py
Args:
is_training: whether the network is in training mode.
depth_multiplier: float depth multiplier for feature extractor.
min_depth: minimum feature extractor depth.
pad_to_multiple: the nearest multiple to zero pad the input height and
width dimensions to.
conv_hyperparams_fn: A function to construct tf slim arg_scope for conv2d
and separable_conv2d ops in the layers that are added on top of the
base feature extractor.
reuse_weights: Whether to reuse variables. Default is None.
use_explicit_padding: Whether to use explicit padding when extracting
features. Default is False.
use_depthwise: Whether to use depthwise convolutions. Default is False.
override_base_feature_extractor_hyperparams: Whether to override
hyperparameters of the base feature extractor with the one from
`conv_hyperparams_fn`.
"""
super(SSDMobileNetV2FeatureExtractor, self).__init__(
is_training=is_training,
depth_multiplier=depth_multiplier,
min_depth=min_depth,
pad_to_multiple=pad_to_multiple,
conv_hyperparams_fn=conv_hyperparams_fn,
reuse_weights=reuse_weights,
use_explicit_padding=use_explicit_padding,
use_depthwise=use_depthwise,
override_base_feature_extractor_hyperparams=
override_base_feature_extractor_hyperparams)
#归一化0-255-->-1,1
def preprocess(self, resized_inputs):
"""SSD preprocessing.
Maps pixel values to the range [-1, 1].
Args:
resized_inputs: a [batch, height, width, channels] float tensor
representing a batch of images.
Returns:
preprocessed_inputs: a [batch, height, width, channels] float tensor
representing a batch of images.
"""
return (2.0 / 255.0) * resized_inputs - 1.0
#从预处理的输入中提取特征图
def extract_features(self, preprocessed_inputs):
"""Extract features from preprocessed inputs.
Args:
preprocessed_inputs: a [batch, height, width, channels] float tensor
representing a batch of images.
Returns:
feature_maps: a list of tensors where the ith tensor has shape
[batch, height_i, width_i, depth_i]
"""
#检查预处理输入的最小尺寸
preprocessed_inputs = shape_utils.check_min_image_dim(
33, preprocessed_inputs)
#指定从MobileNetV2中衍生出来的特征层
'''
‘from_layer’：从mobilenet中提取的namescop的名称，如果为空则表示新生成的。其value为list，长度为6，也就是6个不同尺度的featuremap。
‘layer_depth’:表示channel的深度，为-1，表示从原有的net中继承。
'use_depthwise':表示要使用深度可分离卷积，可降低头部计算的参数量，ssdlite会用到此
‘use_explicit_padding’:如果使能，选择valid pading，并在valid padding之前先做一次fixed padding，
其目的是为了让经过卷积后的size与使用same padding的大小一致。'
'''
feature_map_layout = {
'from_layer': ['layer_15/expansion_output', 'layer_19', '', '', '', ''],
'layer_depth': [-1, -1, 512, 256, 256, 128],
'use_depthwise': self._use_depthwise,
'use_explicit_padding': self._use_explicit_padding,
}
with tf.variable_scope('MobilenetV2', reuse=self._reuse_weights) as scope:
with slim.arg_scope(
mobilenet_v2.training_scope(is_training=None, bn_decay=0.9997)), \
slim.arg_scope(
[mobilenet.depth_multiplier], min_depth=self._min_depth):
with (slim.arg_scope(self._conv_hyperparams_fn())
if self._override_base_feature_extractor_hyperparams else
context_manager.IdentityContextManager()):
#从backbone中获得特征图，指定最后的卷积层
_, image_features = mobilenet_v2.mobilenet_base(
ops.pad_to_multiple(preprocessed_inputs, self._pad_to_multiple),
final_endpoint='layer_19',
depth_multiplier=self._depth_multiplier,
use_explicit_padding=self._use_explicit_padding,
scope=scope)
#通过feature_map_layout中指定的信息去构建多个特征层
'''
multi_resolution_feature_maps：
1）如果配置中提供了层名称，则将该层作为特征图。
2）如果层名称保留为空字符串，则使用卷积生成新的层构造新的特征图，S=2,K=3,导致空间分辨率
降低2倍。
'''
with slim.arg_scope(self._conv_hyperparams_fn()):
feature_maps = feature_map_generators.multi_resolution_feature_maps(
feature_map_layout=feature_map_layout,
depth_multiplier=self._depth_multiplier,
min_depth=self._min_depth,
insert_1x1_conv=True,
image_features=image_features)
return feature_maps.values()
object_detection/models/feature_map_generators.py.py
"""Functions to generate a list of feature maps based on image features.
Provides several feature map generators that can be used to build object
detection feature extractors.
Object detection feature extractors usually are built by stacking two components
- A base feature extractor such as Inception V3 and a feature map generator.
Feature map generators build on the base feature extractors and produce a list
of final feature maps.
"""
import collections
import functools
import tensorflow as tf
from object_detection.utils import ops
slim = tf.contrib.slim
def get_depth_fn(depth_multiplier, min_depth):
"""Builds a callable to compute depth (output channels) of conv filters.
Args:
depth_multiplier: a multiplier for the nominal depth.
min_depth: a lower bound on the depth of filters.
Returns:
A callable that takes in a nominal depth and returns the depth to use.
"""
def multiply_depth(depth):
new_depth = int(depth * depth_multiplier)
return max(new_depth, min_depth)
return multiply_depth
class KerasMultiResolutionFeatureMaps(tf.keras.Model):
"""Generates multi resolution feature maps from input image features.
A Keras model that generates multi-scale feature maps for detection as in the
SSD papers by Liu et al: https://arxiv.org/pdf/1512.02325v2.pdf, See Sec 2.1.
More specifically, when called on inputs it performs the following two tasks:
1) If a layer name is provided in the configuration, returns that layer as a
feature map.
2) If a layer name is left as an empty string, constructs a new feature map
based on the spatial shape and depth configuration. Note that the current
implementation only supports generating new layers using convolution of
stride 2 resulting in a spatial resolution reduction by a factor of 2.
By default convolution kernel size is set to 3, and it can be customized
by caller.
An example of the configuration for Inception V3:
{
'from_layer': ['Mixed_5d', 'Mixed_6e', 'Mixed_7c', '', '', ''],
'layer_depth': [-1, -1, -1, 512, 256, 128]
}
When this feature generator object is called on input image_features:
Args:
image_features: A dictionary of handles to activation tensors from the
base feature extractor.
Returns:
feature_maps: an OrderedDict mapping keys (feature map names) to
tensors where each tensor has shape [batch, height_i, width_i, depth_i].
"""
def __init__(self,
feature_map_layout,
depth_multiplier,
min_depth,
insert_1x1_conv,
is_training,
conv_hyperparams,
freeze_batchnorm,
name=None):
"""Constructor.
Args:
feature_map_layout: Dictionary of specifications for the feature map
layouts in the following format (Inception V2/V3 respectively):
{
'from_layer': ['Mixed_3c', 'Mixed_4c', 'Mixed_5c', '', '', ''],
'layer_depth': [-1, -1, -1, 512, 256, 128]
}
or
{
'from_layer': ['Mixed_5d', 'Mixed_6e', 'Mixed_7c', '', '', ''],
'layer_depth': [-1, -1, -1, 512, 256, 128]
}
If 'from_layer' is specified, the specified feature map is directly used
as a box predictor layer, and the layer_depth is directly infered from
the feature map (instead of using the provided 'layer_depth' parameter).
In this case, our convention is to set 'layer_depth' to -1 for clarity.
Otherwise, if 'from_layer' is an empty string, then the box predictor
layer will be built from the previous layer using convolution
operations. Note that the current implementation only supports
generating new layers using convolutions of stride 2 (resulting in a
spatial resolution reduction by a factor of 2), and will be extended to
a more flexible design. Convolution kernel size is set to 3 by default,
and can be customized by 'conv_kernel_size' parameter (similarily,
'conv_kernel_size' should be set to -1 if 'from_layer' is specified).
The created convolution operation will be a normal 2D convolution by
default, and a depthwise convolution followed by 1x1 convolution if
'use_depthwise' is set to True.
depth_multiplier: Depth multiplier for convolutional layers.
min_depth: Minimum depth for convolutional layers.
insert_1x1_conv: A boolean indicating whether an additional 1x1
convolution should be inserted before shrinking the feature map.
is_training: Indicates whether the feature generator is in training mode.
conv_hyperparams: A `hyperparams_builder.KerasLayerHyperparams` object
containing hyperparameters for convolution ops.
freeze_batchnorm: Bool. Whether to freeze batch norm parameters during
training or not. When training with a small batch size (e.g. 1), it is
desirable to freeze batch norm update and use pretrained batch norm
params.
name: A string name scope to assign to the model. If 'None', Keras
will auto-generate one from the class name.
"""
super(KerasMultiResolutionFeatureMaps, self).__init__(name=name)
self.feature_map_layout = feature_map_layout
self.convolutions = []
depth_fn = get_depth_fn(depth_multiplier, min_depth)
base_from_layer = ''
use_explicit_padding = False
if 'use_explicit_padding' in feature_map_layout:
use_explicit_padding = feature_map_layout['use_explicit_padding']
use_depthwise = False
if 'use_depthwise' in feature_map_layout:
use_depthwise = feature_map_layout['use_depthwise']
for index, from_layer in enumerate(feature_map_layout['from_layer']):
net = []
layer_depth = feature_map_layout['layer_depth'][index]
conv_kernel_size = 3
if 'conv_kernel_size' in feature_map_layout:
conv_kernel_size = feature_map_layout['conv_kernel_size'][index]
if from_layer:
base_from_layer = from_layer
else:
if insert_1x1_conv:
layer_name = '{}_1_Conv2d_{}_1x1_{}'.format(
base_from_layer, index, depth_fn(layer_depth / 2))
net.append(tf.keras.layers.Conv2D(depth_fn(layer_depth / 2),
[1, 1],
padding='SAME',
strides=1,
name=layer_name + '_conv',
**conv_hyperparams.params()))
net.append(
conv_hyperparams.build_batch_norm(
training=(is_training and not freeze_batchnorm),
name=layer_name + '_batchnorm'))
net.append(
conv_hyperparams.build_activation_layer(
name=layer_name))
layer_name = '{}_2_Conv2d_{}_{}x{}_s2_{}'.format(
base_from_layer, index, conv_kernel_size, conv_kernel_size,
depth_fn(layer_depth))
stride = 2
padding = 'SAME'
if use_explicit_padding:
padding = 'VALID'
# We define this function here while capturing the value of
# conv_kernel_size, to avoid holding a reference to the loop variable
# conv_kernel_size inside of a lambda function
def fixed_padding(features, kernel_size=conv_kernel_size):
return ops.fixed_padding(features, kernel_size)
net.append(tf.keras.layers.Lambda(fixed_padding))
# TODO(rathodv): Add some utilities to simplify the creation of
# Depthwise & non-depthwise convolutions w/ normalization & activations
if use_depthwise:
net.append(tf.keras.layers.DepthwiseConv2D(
[conv_kernel_size, conv_kernel_size],
depth_multiplier=1,
padding=padding,
strides=stride,
name=layer_name + '_depthwise_conv',
**conv_hyperparams.params()))
net.append(
conv_hyperparams.build_batch_norm(
training=(is_training and not freeze_batchnorm),
name=layer_name + '_depthwise_batchnorm'))
net.append(
conv_hyperparams.build_activation_layer(
name=layer_name + '_depthwise'))
net.append(tf.keras.layers.Conv2D(depth_fn(layer_depth), [1, 1],
padding='SAME',
strides=1,
name=layer_name + '_conv',
**conv_hyperparams.params()))
net.append(
conv_hyperparams.build_batch_norm(
training=(is_training and not freeze_batchnorm),
name=layer_name + '_batchnorm'))
net.append(
conv_hyperparams.build_activation_layer(
name=layer_name))
else:
net.append(tf.keras.layers.Conv2D(
depth_fn(layer_depth),
[conv_kernel_size, conv_kernel_size],
padding=padding,
strides=stride,
name=layer_name + '_conv',
**conv_hyperparams.params()))
net.append(
conv_hyperparams.build_batch_norm(
training=(is_training and not freeze_batchnorm),
name=layer_name + '_batchnorm'))
net.append(
conv_hyperparams.build_activation_layer(
name=layer_name))
# Until certain bugs are fixed in checkpointable lists,
# this net must be appended only once it's been filled with layers
self.convolutions.append(net)
def call(self, image_features):
"""Generate the multi-resolution feature maps.
Executed when calling the `.__call__` method on input.
Args:
image_features: A dictionary of handles to activation tensors from the
base feature extractor.
Returns:
feature_maps: an OrderedDict mapping keys (feature map names) to
tensors where each tensor has shape [batch, height_i, width_i, depth_i].
"""
feature_maps = []
feature_map_keys = []
for index, from_layer in enumerate(self.feature_map_layout['from_layer']):
if from_layer:
feature_map = image_features[from_layer]
feature_map_keys.append(from_layer)
else:
feature_map = feature_maps[-1]
for layer in self.convolutions[index]:
feature_map = layer(feature_map)
layer_name = self.convolutions[index][-1].name
feature_map_keys.append(layer_name)
feature_maps.append(feature_map)
return collections.OrderedDict(
[(x, y) for (x, y) in zip(feature_map_keys, feature_maps)])
def multi_resolution_feature_maps(feature_map_layout, depth_multiplier,
min_depth, insert_1x1_conv, image_features,
pool_residual=False):
"""Generates multi resolution feature maps from input image features.
Generates multi-scale feature maps for detection as in the SSD papers by
Liu et al: https://arxiv.org/pdf/1512.02325v2.pdf, See Sec 2.1.
More specifically, it performs the following two tasks:
1) If a layer name is provided in the configuration, returns that layer as a
feature map.
2) If a layer name is left as an empty string, constructs a new feature map
based on the spatial shape and depth configuration. Note that the current
implementation only supports generating new layers using convolution of
stride 2 resulting in a spatial resolution reduction by a factor of 2.
By default convolution kernel size is set to 3, and it can be customized
by caller.
An example of the configuration for Inception V3:
{
'from_layer': ['Mixed_5d', 'Mixed_6e', 'Mixed_7c', '', '', ''],
'layer_depth': [-1, -1, -1, 512, 256, 128]
}
Args:
feature_map_layout: Dictionary of specifications for the feature map
layouts in the following format (Inception V2/V3 respectively):
{
'from_layer': ['Mixed_3c', 'Mixed_4c', 'Mixed_5c', '', '', ''],
'layer_depth': [-1, -1, -1, 512, 256, 128]
}
or
{
'from_layer': ['Mixed_5d', 'Mixed_6e', 'Mixed_7c', '', '', ''],
'layer_depth': [-1, -1, -1, 512, 256, 128]
}
If 'from_layer' is specified, the specified feature map is directly used
as a box predictor layer, and the layer_depth is directly infered from the
feature map (instead of using the provided 'layer_depth' parameter). In
this case, our convention is to set 'layer_depth' to -1 for clarity.
Otherwise, if 'from_layer' is an empty string, then the box predictor
layer will be built from the previous layer using convolution operations.
Note that the current implementation only supports generating new layers
using convolutions of stride 2 (resulting in a spatial resolution
reduction by a factor of 2), and will be extended to a more flexible
design. Convolution kernel size is set to 3 by default, and can be
customized by 'conv_kernel_size' parameter (similarily, 'conv_kernel_size'
should be set to -1 if 'from_layer' is specified). The created convolution
operation will be a normal 2D convolution by default, and a depthwise
convolution followed by 1x1 convolution if 'use_depthwise' is set to True.
depth_multiplier: Depth multiplier for convolutional layers.
min_depth: Minimum depth for convolutional layers.
insert_1x1_conv: A boolean indicating whether an additional 1x1 convolution
should be inserted before shrinking the feature map.
image_features: A dictionary of handles to activation tensors from the
base feature extractor.
pool_residual: Whether to add an average pooling layer followed by a
residual connection between subsequent feature maps when the channel
depth match. For example, with option 'layer_depth': [-1, 512, 256, 256],
a pooling and residual layer is added between the third and forth feature
map. This option is better used with Weight Shared Convolution Box
Predictor when all feature maps have the same channel depth to encourage
more consistent features across multi-scale feature maps.
Returns:
feature_maps: an OrderedDict mapping keys (feature map names) to
tensors where each tensor has shape [batch, height_i, width_i, depth_i].
Raises:
ValueError: if the number entries in 'from_layer' and
'layer_depth' do not match.
ValueError: if the generated layer does not have the same resolution
as specified.
"""
depth_fn = get_depth_fn(depth_multiplier, min_depth)
feature_map_keys = []
feature_maps = []
base_from_layer = ''
use_explicit_padding = False
if 'use_explicit_padding' in feature_map_layout:
use_explicit_padding = feature_map_layout['use_explicit_padding']
use_depthwise = False
if 'use_depthwise' in feature_map_layout:
use_depthwise = feature_map_layout['use_depthwise']
for index, from_layer in enumerate(feature_map_layout['from_layer']):
layer_depth = feature_map_layout['layer_depth'][index]
conv_kernel_size = 3
if 'conv_kernel_size' in feature_map_layout:
conv_kernel_size = feature_map_layout['conv_kernel_size'][index]
if from_layer:
feature_map = image_features[from_layer]
base_from_layer = from_layer
feature_map_keys.append(from_layer)
else:
pre_layer = feature_maps[-1]
pre_layer_depth = pre_layer.get_shape().as_list()[3]
intermediate_layer = pre_layer
if insert_1x1_conv:
layer_name = '{}_1_Conv2d_{}_1x1_{}'.format(
base_from_layer, index, depth_fn(layer_depth / 2))
intermediate_layer = slim.conv2d(
pre_layer,
depth_fn(layer_depth / 2), [1, 1],
padding='SAME',
stride=1,
scope=layer_name)
layer_name = '{}_2_Conv2d_{}_{}x{}_s2_{}'.format(
base_from_layer, index, conv_kernel_size, conv_kernel_size,
depth_fn(layer_depth))
stride = 2
padding = 'SAME'
if use_explicit_padding:
padding = 'VALID'
intermediate_layer = ops.fixed_padding(
intermediate_layer, conv_kernel_size)
if use_depthwise:
feature_map = slim.separable_conv2d(
intermediate_layer,
None, [conv_kernel_size, conv_kernel_size],
depth_multiplier=1,
padding=padding,
stride=stride,
scope=layer_name + '_depthwise')
feature_map = slim.conv2d(
feature_map,
depth_fn(layer_depth), [1, 1],
padding='SAME',
stride=1,
scope=layer_name)
if pool_residual and pre_layer_depth == depth_fn(layer_depth):
feature_map += slim.avg_pool2d(
pre_layer, [3, 3],
padding='SAME',
stride=2,
scope=layer_name + '_pool')
else:
feature_map = slim.conv2d(
intermediate_layer,
depth_fn(layer_depth), [conv_kernel_size, conv_kernel_size],
padding=padding,
stride=stride,
scope=layer_name)
feature_map_keys.append(layer_name)
feature_maps.append(feature_map)
return collections.OrderedDict(
[(x, y) for (x, y) in zip(feature_map_keys, feature_maps)])
def fpn_top_down_feature_maps(image_features,
depth,
use_depthwise=False,
use_explicit_padding=False,
scope=None):
"""Generates `top-down` feature maps for Feature Pyramid Networks.
See https://arxiv.org/abs/1612.03144 for details.
Args:
image_features: list of tuples of (tensor_name, image_feature_tensor).
Spatial resolutions of succesive tensors must reduce exactly by a factor
of 2.
depth: depth of output feature maps.
use_depthwise: whether to use depthwise separable conv instead of regular
conv.
use_explicit_padding: whether to use explicit padding.
scope: A scope name to wrap this op under.
Returns:
feature_maps: an OrderedDict mapping keys (feature map names) to
tensors where each tensor has shape [batch, height_i, width_i, depth_i].
"""
with tf.name_scope(scope, 'top_down'):
num_levels = len(image_features)
output_feature_maps_list = []
output_feature_map_keys = []
padding = 'VALID' if use_explicit_padding else 'SAME'
kernel_size = 3
with slim.arg_scope(
[slim.conv2d, slim.separable_conv2d], padding=padding, stride=1):
top_down = slim.conv2d(
image_features[-1][1],
depth, [1, 1], activation_fn=None, normalizer_fn=None,
scope='projection_%d' % num_levels)
output_feature_maps_list.append(top_down)
output_feature_map_keys.append(
'top_down_%s' % image_features[-1][0])
for level in reversed(range(num_levels - 1)):
top_down = ops.nearest_neighbor_upsampling(top_down, 2)
residual = slim.conv2d(
image_features[level][1], depth, [1, 1],
activation_fn=None, normalizer_fn=None,
scope='projection_%d' % (level + 1))
if use_explicit_padding:
# slice top_down to the same shape as residual
residual_shape = tf.shape(residual)
top_down = top_down[:, :residual_shape[1], :residual_shape[2], :]
top_down += residual
if use_depthwise:
conv_op = functools.partial(slim.separable_conv2d, depth_multiplier=1)
else:
conv_op = slim.conv2d
if use_explicit_padding:
top_down = ops.fixed_padding(top_down, kernel_size)
output_feature_maps_list.append(conv_op(
top_down,
depth, [kernel_size, kernel_size],
scope='smoothing_%d' % (level + 1)))
output_feature_map_keys.append('top_down_%s' % image_features[level][0])
return collections.OrderedDict(reversed(
list(zip(output_feature_map_keys, output_feature_maps_list))))
def pooling_pyramid_feature_maps(base_feature_map_depth, num_layers,
image_features, replace_pool_with_conv=False):
"""Generates pooling pyramid feature maps.
The pooling pyramid feature maps is motivated by
multi_resolution_feature_maps. The main difference are that it is simpler and
reduces the number of free parameters.
More specifically:
- Instead of using convolutions to shrink the feature map, it uses max
pooling, therefore totally gets rid of the parameters in convolution.
- By pooling feature from larger map up to a single cell, it generates
features in the same feature space.
- Instead of independently making box predictions from individual maps, it
shares the same classifier across different feature maps, therefore reduces
the "mis-calibration" across different scales.
See go/ppn-detection for more details.
Args:
base_feature_map_depth: Depth of the base feature before the max pooling.
num_layers: Number of layers used to make predictions. They are pooled
from the base feature.
image_features: A dictionary of handles to activation tensors from the
feature extractor.
replace_pool_with_conv: Whether or not to replace pooling operations with
convolutions in the PPN. Default is False.
Returns:
feature_maps: an OrderedDict mapping keys (feature map names) to
tensors where each tensor has shape [batch, height_i, width_i, depth_i].
Raises:
ValueError: image_features does not contain exactly one entry
"""
if len(image_features) != 1:
raise ValueError('image_features should be a dictionary of length 1.')
image_features = image_features[image_features.keys()[0]]
feature_map_keys = []
feature_maps = []
feature_map_key = 'Base_Conv2d_1x1_%d' % base_feature_map_depth
if base_feature_map_depth > 0:
image_features = slim.conv2d(
image_features,
base_feature_map_depth,
[1, 1], # kernel size
padding='SAME', stride=1, scope=feature_map_key)
# Add a 1x1 max-pooling node (a no op node) immediately after the conv2d for
# TPU v1 compatibility. Without the following dummy op, TPU runtime
# compiler will combine the convolution with one max-pooling below into a
# single cycle, so getting the conv2d feature becomes impossible.
image_features = slim.max_pool2d(
image_features, [1, 1], padding='SAME', stride=1, scope=feature_map_key)
feature_map_keys.append(feature_map_key)
feature_maps.append(image_features)
feature_map = image_features
if replace_pool_with_conv:
with slim.arg_scope([slim.conv2d], padding='SAME', stride=2):
for i in range(num_layers - 1):
feature_map_key = 'Conv2d_{}_3x3_s2_{}'.format(i,
base_feature_map_depth)
feature_map = slim.conv2d(
feature_map, base_feature_map_depth, [3, 3], scope=feature_map_key)
feature_map_keys.append(feature_map_key)
feature_maps.append(feature_map)
else:
with slim.arg_scope([slim.max_pool2d], padding='SAME', stride=2):
for i in range(num_layers - 1):
feature_map_key = 'MaxPool2d_%d_2x2' % i
feature_map = slim.max_pool2d(
feature_map, [2, 2], padding='SAME', scope=feature_map_key)
feature_map_keys.append(feature_map_key)
feature_maps.append(feature_map)
return collections.OrderedDict(
[(x, y) for (x, y) in zip(feature_map_keys, feature_maps)])
(300*300)sdd_mobilenetv2输出的feature map 结构
| namescope | channel depth | feature map size |
| layer_15/expansion_output | 512 | 19*19 |
| layer_19 | 1024 | 10*10 |
| layer_19_2_Conv2d_2_3x3_s2_512 | 512 | 5*5 |
| layer_19_2_Conv2d_2_3x3_s2_256 | 256 | 3*3 |
| layer_19_2_Conv2d_2_3x3_s2_256 | 256 | 2*2 |
| layer_19_2_Conv2d_2_3x3_s2_281 | 128 | 1*1 |
参考：https://blog.csdn.net/AECHO1/article/details/81196498