本章继续分析Efficientdet模型，Box与Class回归网络部分。
Box与Class实现代码一样，共5层共用一个卷积层，然后链接单独的BN，接激活函数swish，drop_connect方法用于drop部分卷积。Box与Class主要区别是在最后一层通道数不同，Box通道数为num_anchors*4，Class通道数为num_anchors*(num_classes+1)，这里的1为背景。
结构图如下：
代码如下：
class ClassNet(tf.keras.layers.Layer): """Object class prediction network.""" def __init__(self, num_classes=90, num_anchors=9, num_filters=32, min_level=3, max_level=7, is_training_bn=False, act_type='swish', repeats=4, separable_conv=True, survival_prob=None, strategy=None, data_format='channels_last', name='class_net', **kwargs): """Initialize the ClassNet. Args: num_classes: number of classes. num_anchors: number of anchors. num_filters: number of filters for "intermediate" layers. min_level: minimum level for features. max_level: maximum level for features. is_training_bn: True if we train the BatchNorm. act_type: String of the activation used. repeats: number of intermediate layers. separable_conv: True to use separable_conv instead of conv2D. survival_prob: if a value is set then drop connect will be used. strategy: string to specify training strategy for TPU/GPU/CPU. data_format: string of 'channel_first' or 'channels_last'. name: the name of this layerl. **kwargs: other parameters. """ super().__init__(name=name, **kwargs) self.num_classes = num_classes self.num_anchors = num_anchors self.num_filters = num_filters self.min_level = min_level self.max_level = max_level self.repeats = repeats self.separable_conv = separable_conv self.is_training_bn = is_training_bn self.survival_prob = survival_prob self.act_type = act_type self.strategy = strategy self.data_format = data_format self.conv_ops = [] self.bns = [] if separable_conv: conv2d_layer = functools.partial( tf.keras.layers.SeparableConv2D, depth_multiplier=1, data_format=data_format, pointwise_initializer=tf.initializers.VarianceScaling(), depthwise_initializer=tf.initializers.VarianceScaling()) else: conv2d_layer = functools.partial( tf.keras.layers.Conv2D, data_format=data_format, kernel_initializer=tf.random_normal_initializer(stddev=0.01)) for i in range(self.repeats): # If using SeparableConv2D self.conv_ops.append( conv2d_layer( self.num_filters, kernel_size=3, bias_initializer=tf.zeros_initializer(), activation=None, padding='same', name='class-%d' % i)) bn_per_level = [] for level in range(self.min_level, self.max_level + 1): bn_per_level.append( util_keras.build_batch_norm( is_training_bn=self.is_training_bn, strategy=self.strategy, data_format=self.data_format, name='class-%d-bn-%d' % (i, level), )) self.bns.append(bn_per_level) self.classes = conv2d_layer( num_classes * num_anchors, kernel_size=3, bias_initializer=tf.constant_initializer(-np.log((1 - 0.01) / 0.01)), padding='same', name='class-predict') def call(self, inputs, training, **kwargs): """Call ClassNet.""" class_outputs = [] for level_id in range(0, self.max_level - self.min_level + 1): image = inputs[level_id] for i in range(self.repeats): # 这里有跳层 original_image = image # 所有层经过同一个卷积 image = self.conv_ops[i](image) # 每层单独对应BN image = self.bns[i][level_id](image, training=training) if self.act_type: image = utils.activation_fn(image, self.act_type) if i > 0 and self.survival_prob: # 多于1层时，drop掉后面几层 image = utils.drop_connect(image, training, self.survival_prob) image = image + original_image # 最后输出结果 class_outputs.append(self.classes(image)) return class_outputs class BoxNet(tf.keras.layers.Layer): """Box regression network.""" def __init__(self, num_anchors=9, num_filters=32, min_level=3, max_level=7, is_training_bn=False, act_type='swish', repeats=4, separable_conv=True, survival_prob=None, strategy=None, data_format='channels_last', name='box_net', **kwargs): """Initialize BoxNet. Args: num_anchors: number of anchors used. num_filters: number of filters for "intermediate" layers. min_level: minimum level for features. max_level: maximum level for features. is_training_bn: True if we train the BatchNorm. act_type: String of the activation used. repeats: number of "intermediate" layers. separable_conv: True to use separable_conv instead of conv2D. survival_prob: if a value is set then drop connect will be used. strategy: string to specify training strategy for TPU/GPU/CPU. data_format: string of 'channel_first' or 'channels_last'. name: Name of the layer. **kwargs: other parameters. """ super().__init__(name=name, **kwargs) self.num_anchors = num_anchors self.num_filters = num_filters self.min_level = min_level self.max_level = max_level self.repeats = repeats self.separable_conv = separable_conv self.is_training_bn = is_training_bn self.survival_prob = survival_prob self.act_type = act_type self.strategy = strategy self.data_format = data_format self.conv_ops = [] self.bns = [] for i in range(self.repeats): # If using SeparableConv2D if self.separable_conv: self.conv_ops.append( tf.keras.layers.SeparableConv2D( filters=self.num_filters, depth_multiplier=1, pointwise_initializer=tf.initializers.VarianceScaling(), depthwise_initializer=tf.initializers.VarianceScaling(), data_format=self.data_format, kernel_size=3, activation=None, bias_initializer=tf.zeros_initializer(), padding='same', name='box-%d' % i)) # If using Conv2d else: self.conv_ops.append( tf.keras.layers.Conv2D( filters=self.num_filters, kernel_initializer=tf.random_normal_initializer(stddev=0.01), data_format=self.data_format, kernel_size=3, activation=None, bias_initializer=tf.zeros_initializer(), padding='same', name='box-%d' % i)) bn_per_level = [] for level in range(self.min_level, self.max_level + 1): bn_per_level.append( util_keras.build_batch_norm( is_training_bn=self.is_training_bn, strategy=self.strategy, data_format=self.data_format, name='box-%d-bn-%d' % (i, level))) self.bns.append(bn_per_level) if self.separable_conv: self.boxes = tf.keras.layers.SeparableConv2D( filters=4 * self.num_anchors, depth_multiplier=1, pointwise_initializer=tf.initializers.VarianceScaling(), depthwise_initializer=tf.initializers.VarianceScaling(), data_format=self.data_format, kernel_size=3, activation=None, bias_initializer=tf.zeros_initializer(), padding='same', name='box-predict') else: self.boxes = tf.keras.layers.Conv2D( filters=4 * self.num_anchors, kernel_initializer=tf.random_normal_initializer(stddev=0.01), data_format=self.data_format, kernel_size=3, activation=None, bias_initializer=tf.zeros_initializer(), padding='same', name='box-predict') def call(self, inputs, training): """Call boxnet.""" # 和分类网络一样 box_outputs = [] for level_id in range(0, self.max_level - self.min_level + 1): image = inputs[level_id] for i in range(self.repeats): # 这里有跳层 original_image = image # 所有层经过同一个卷积 image = self.conv_ops[i](image) # 每层单独对应BN image = self.bns[i][level_id](image, training=training) if self.act_type: image = utils.activation_fn(image, self.act_type) if i > 0 and self.survival_prob: # 多于1层时，drop掉后面几层 image = utils.drop_connect(image, training, self.survival_prob) image = image + original_image box_outputs.append(self.boxes(image)) return box_outputs到此本章内容结束，下一章，Loss计算细节分析，敬请关注！！！
