Fork版本项目地址：SSD
一、输入标签生成
在数据预处理之后，图片、类别、真实框格式较为原始，不能够直接作为损失函数的输入标签（ssd向前网络只需要图像就行，这里的处理主要需要满足loss的计算），对于一张图片（三维CHW）我们需要如下格式的数据作为损失函数标签：
gclasse： 搜索框对应的真实类别
长度为ssd特征层f的list，每一个元素是一个Tensor，shape为：该层中心点行数×列数×每个中心点包含搜索框数目
gscores： 搜索框和真实框的IOU，gclasses中记录的就是该真实框的类别
长度为ssd特征层f的list，每一个元素是一个Tensor，shape为：该层中心点行数×列数×每个中心点包含搜索框数目
glocalisations： 搜索框相较于真实框位置修正，由于有4个坐标，所以维度多了一维
长度为ssd特征层f的list，每一个元素是一个Tensor，shape为：该层中心点行数×列数×每个中心点包含搜索框数目×4
为了计算出上面标签，我们函数调用如下（train_ssd_network.py）：
# f层个(m,m,k)，f层个(m,m,k,4xywh)，f层个(m,m,k) f层表示提取ssd特征的层的数目
# 0-20数字,方便loss的坐标记录,IOU值
gclasses, glocalisations, gscores =
ssd_net.bboxes_encode(glabels, gbboxes, ssd_anchors)
输入变量都是前几节中的函数输出（train_ssd_network.py）：
ssd_anchors = ssd_net.anchors(ssd_shape) # 调用类方法，创建搜素框
# Pre-processing image, labels and bboxes.
# 'CHW' (n,) (n, 4)
image, glabels, gbboxes =
image_preprocessing_fn(image, glabels, gbboxes,
out_shape=ssd_shape, # (300,300)
data_format=DATA_FORMAT) # 'NCHW'
至此，我们再来看一看该函数如何实现，其处理过程是按照ssd特征层进行划分，首先建立三个list，然后对于每一个特征层计算该层的三个Tensor，最后分别添加进list中（ssd_common.py）：
def tf_ssd_bboxes_encode(labels,
bboxes,
anchors,
num_classes,
no_annotation_label,
ignore_threshold=0.5,
prior_scaling=(0.1, 0.1, 0.2, 0.2),
dtype=tf.float32,
scope='ssd_bboxes_encode'):
with tf.name_scope(scope):
target_labels = []
target_localizations = []
target_scores = []
# anchors_layer: (y, x, h, w)
for i, anchors_layer in enumerate(anchors):
with tf.name_scope('bboxes_encode_block_%i' % i):
# (m,m,k)，xywh(m,m,4k)，(m,m,k)
t_labels, t_loc, t_scores =
tf_ssd_bboxes_encode_layer(labels, bboxes, anchors_layer,
num_classes, no_annotation_label,
ignore_threshold,
prior_scaling, dtype)
target_labels.append(t_labels)
target_localizations.append(t_loc)
target_scores.append(t_scores)
return target_labels, target_localizations, target_scores
每一层处理是重点（ssd_common.py），从这里我们可以更深刻体会到所有框体长度信息归一化的便捷之处——不同层的框体均可以直接和真实框做运算，毕竟它们都是0~1的相对位置：
# 为了有助理解，m表示该层中心点行列数，k为每个中心点对应的框数，n为图像上的目标数
def tf_ssd_bboxes_encode_layer(labels, # (n,)
bboxes, # (n, 4)
anchors_layer, # y(m, m, 1), x(m, m, 1), h(k,), w(k,)
num_classes,
no_annotation_label,
ignore_threshold=0.5,
prior_scaling=(0.1, 0.1, 0.2, 0.2),
dtype=tf.float32):
"""Encode groundtruth labels and bounding boxes using SSD anchors from
one layer.
Arguments:
labels: 1D Tensor(int64) containing groundtruth labels;
bboxes: Nx4 Tensor(float) with bboxes relative coordinates;
anchors_layer: Numpy array with layer anchors;
matching_threshold: Threshold for positive match with groundtruth bboxes;
prior_scaling: Scaling of encoded coordinates.
Return:
(target_labels, target_localizations, target_scores): Target Tensors.
"""
# Anchors coordinates and volume.
yref, xref, href, wref = anchors_layer # y(m, m, 1), x(m, m, 1), h(k,), w(k,)
ymin = yref - href / 2. # (m, m, k)
xmin = xref - wref / 2.
ymax = yref + href / 2.
xmax = xref + wref / 2.
vol_anchors = (xmax - xmin) * (ymax - ymin) # 搜索框面积(m, m, k)
# Initialize tensors...
# 下面各个Tensor矩阵的shape等于中心点坐标矩阵的shape
shape = (yref.shape[0], yref.shape[1], href.size) # (m, m, k)
feat_labels = tf.zeros(shape, dtype=tf.int64) # (m, m, k)
feat_scores = tf.zeros(shape, dtype=dtype)
feat_ymin = tf.zeros(shape, dtype=dtype)
feat_xmin = tf.zeros(shape, dtype=dtype)
feat_ymax = tf.ones(shape, dtype=dtype)
feat_xmax = tf.ones(shape, dtype=dtype)
def jaccard_with_anchors(bbox):
"""Compute jaccard score between a box and the anchors.
"""
int_ymin = tf.maximum(ymin, bbox[0]) # (m, m, k)
int_xmin = tf.maximum(xmin, bbox[1])
int_ymax = tf.minimum(ymax, bbox[2])
int_xmax = tf.minimum(xmax, bbox[3])
h = tf.maximum(int_ymax - int_ymin, 0.)
w = tf.maximum(int_xmax - int_xmin, 0.)
# Volumes.
# 处理搜索框和bbox之间的联系
inter_vol = h * w # 交集面积
union_vol = vol_anchors - inter_vol
+ (bbox[2] - bbox[0]) * (bbox[3] - bbox[1]) # 并集面积
jaccard = tf.div(inter_vol, union_vol) # 交集/并集，即IOU
return jaccard # (m, m, k)
def condition(i, feat_labels, feat_scores,
feat_ymin, feat_xmin, feat_ymax, feat_xmax):
"""Condition: check label index.
"""
r = tf.less(i, tf.shape(labels))
return r[0] # tf.shape(labels)有维度，所以r有维度
def body(i, feat_labels, feat_scores,
feat_ymin, feat_xmin, feat_ymax, feat_xmax):
"""Body: update feature labels, scores and bboxes.
Follow the original SSD paper for that purpose:
- assign values when jaccard > 0.5;
- only update if beat the score of other bboxes.
"""
# Jaccard score.
label = labels[i] # 当前图片上第i个对象的标签
bbox = bboxes[i] # 当前图片上第i个对象的真实框bbox
jaccard = jaccard_with_anchors(bbox) # 当前对象的bbox和当前层的搜索网格IOU，(m, m, k)
# Mask: check threshold + scores + no annotations + num_classes.
mask = tf.greater(jaccard, feat_scores) # 掩码矩阵，IOU大于历史得分的为True，(m, m, k)
# mask = tf.logical_and(mask, tf.greater(jaccard, matching_threshold))
mask = tf.logical_and(mask, feat_scores > -0.5)
mask = tf.logical_and(mask, label < num_classes) # 不太懂，label应该必定小于类别数
imask = tf.cast(mask, tf.int64) # 整形mask
fmask = tf.cast(mask, dtype) # 浮点型mask
# Update values using mask.
# 保证feat_labels存储对应位置得分最大对象标签，feat_scores存储那个得分
# (m, m, k) × 当前类别scalar + (1 - (m, m, k)) × (m, m, k)
# 更新label记录，此时的imask已经保证了True位置当前对像得分高于之前的对象得分，其他位置值不变
feat_labels = imask * label + (1 - imask) * feat_labels
# 更新score记录，mask为True使用本类别IOU，否则不变
feat_scores = tf.where(mask, jaccard, feat_scores)
# 下面四个矩阵存储对应label的真实框坐标
# (m, m, k) × 当前框坐标scalar + (1 - (m, m, k)) × (m, m, k)
feat_ymin = fmask * bbox[0] + (1 - fmask) * feat_ymin
feat_xmin = fmask * bbox[1] + (1 - fmask) * feat_xmin
feat_ymax = fmask * bbox[2] + (1 - fmask) * feat_ymax
feat_xmax = fmask * bbox[3] + (1 - fmask) * feat_xmax
return [i+1, feat_labels, feat_scores,
feat_ymin, feat_xmin, feat_ymax, feat_xmax]
# Main loop definition.
# 对当前图像上每一个目标进行循环
i = 0
(i,
feat_labels, feat_scores,
feat_ymin, feat_xmin,
feat_ymax, feat_xmax) = tf.while_loop(condition, body,
[i,
feat_labels, feat_scores,
feat_ymin, feat_xmin,
feat_ymax, feat_xmax])
# Transform to center / size.
# 这里的y、x、h、w指的是对应位置所属真实框的相关属性
feat_cy = (feat_ymax + feat_ymin) / 2.
feat_cx = (feat_xmax + feat_xmin) / 2.
feat_h = feat_ymax - feat_ymin
feat_w = feat_xmax - feat_xmin
# Encode features.
# prior_scaling: [0.1, 0.1, 0.2, 0.2]，放缩意义不明
# ((m, m, k) - (m, m, 1)) / (k,) * 10
# 以搜索网格中心点为参考，真实框中心的偏移，单位长度为网格hw
feat_cy = (feat_cy - yref) / href / prior_scaling[0]
feat_cx = (feat_cx - xref) / wref / prior_scaling[1]
# log((m, m, k) / (m, m, 1)) * 5
# 真实框宽高/搜索网格宽高，取对
feat_h = tf.log(feat_h / href) / prior_scaling[2]
feat_w = tf.log(feat_w / wref) / prior_scaling[3]
# Use SSD ordering: x / y / w / h instead of ours.(m, m, k, 4)
feat_localizations = tf.stack([feat_cx, feat_cy, feat_w, feat_h], axis=-1) # -1会扩维，故有4
return feat_labels, feat_localizations, feat_scores
可以看到（最后几行），feat_localizations用于位置修正记录，其中存储的并不是直接的搜索框和真实框的差，而是按照loss函数所需要的格式进行存储，但是进行prior_scaling处理的意义不明，不过直观来看对loss函数不构成负面影响（损失函数值依旧是搜索框等于真实框最佳）。
二、处理为batch
生成batch数据队列
截止到目前，我们的数据都是对单张图片而言，需要将之整理为batch size的Tensor，不过有点小麻烦，就是我们的数据以list包含Tensor为主，维度扩充需要一点小技巧（tf_utils.py）：
def reshape_list(l, shape=None):
"""Reshape list of (list): 1D to 2D or the other way around.
Args:
l: List or List of list.
shape: 1D or 2D shape.
Return
Reshaped list.
"""
r = []
if shape is None:
# Flatten everything.
for a in l:
if isinstance(a, (list, tuple)):
r = r + list(a)
else:
r.append(a)
else:
# Reshape to list of list.
i = 0
for s in shape:
if s == 1:
r.append(l[i])
else:
r.append(l[i:i+s])
i += s
return r
这个函数可以将list1：[Tensor11, [Tensor21, Tensor22, ……], [Ten31, Tensor32, ……], ……]和list2：[Tensor1, Tensor2, ……]这样的形式相互转换，需要的就是记录下list1中各子list长度，单个Tensor记为1（train_ssd_network.py）：
batch_shape = [1] + [len(ssd_anchors)] * 3 # (1,f层,f层,f层)
# Training batches and queue.
r = tf.train.batch( # 图片，中心点类别，真实框坐标，得分
tf_utils.reshape_list([image, gclasses, glocalisations, gscores]),
batch_size=FLAGS.batch_size, # 32
num_threads=FLAGS.num_preprocessing_threads,
capacity=5 * FLAGS.batch_size)
b_image, b_gclasses, b_glocalisations, b_gscores =
tf_utils.reshape_list(r, batch_shape)
# Intermediate queueing: unique batch computation pipeline for all
# GPUs running the training.
batch_queue = slim.prefetch_queue.prefetch_queue(
tf_utils.reshape_list([b_image, b_gclasses, b_glocalisations, b_gscores]),
capacity=2 * deploy_config.num_clones)
由于tf.train.batch接收输入格式为[Tensor1, Tensor2, ……]，所以要先使用上面函数处理输入，使单张图片的标签数据变化为batch size的标签数据，再将标签数据格式变换回来（实际就是把list1化为list2后给其中每一个Tensor加了一个维度，再变换回list1的格式），最后将batch size的Tensor创建队列，不过没必要这么麻烦，实际上像下面这么做也不会报错，省略了来回折腾Tensor的过程……
batch_shape = [1] + [len(ssd_anchors)] * 3 # (1,f层,f层,f层)
r = tf.train.batch( # 图片，中心点类别，真实框坐标，得分
tf_utils.reshape_list([image, gclasses, glocalisations, gscores]),
batch_size=FLAGS.batch_size, # 32
num_threads=FLAGS.num_preprocessing_threads,
capacity=5 * FLAGS.batch_size)
# Intermediate queueing: unique batch computation pipeline for all
# GPUs running the training.
batch_queue = slim.prefetch_queue.prefetch_queue(
r, # <-----输入格式实际上并不需要调整
capacity=2 * deploy_config.num_clones)
获取batch数据队列
# Dequeue batch.
b_image, b_gclasses, b_glocalisations, b_gscores =
tf_utils.reshape_list(batch_queue.dequeue(), batch_shape) # 重整list
出队后整理一下list格式即可，此时获取的数据格式如下（vgg_300为例）：
<tf.Tensor 'batch:0' shape=(32, 3, 300, 300) dtype=float32>
[<tf.Tensor 'batch:1' shape=(32, 38, 38, 4) dtype=int64>,
<tf.Tensor 'batch:2' shape=(32, 19, 19, 6) dtype=int64>,
<tf.Tensor 'batch:3' shape=(32, 10, 10, 6) dtype=int64>,
<tf.Tensor 'batch:4' shape=(32, 5, 5, 6) dtype=int64>,
<tf.Tensor 'batch:5' shape=(32, 3, 3, 4) dtype=int64>,
<tf.Tensor 'batch:6' shape=(32, 1, 1, 4) dtype=int64>]
[<tf.Tensor 'batch:7' shape=(32, 38, 38, 4, 4) dtype=float32>,
<tf.Tensor 'batch:8' shape=(32, 19, 19, 6, 4) dtype=float32>,
<tf.Tensor 'batch:9' shape=(32, 10, 10, 6, 4) dtype=float32>,
<tf.Tensor 'batch:10' shape=(32, 5, 5, 6, 4) dtype=float32>,
<tf.Tensor 'batch:11' shape=(32, 3, 3, 4, 4) dtype=float32>,
<tf.Tensor 'batch:12' shape=(32, 1, 1, 4, 4) dtype=float32>]
[<tf.Tensor 'batch:13' shape=(32, 38, 38, 4) dtype=float32>,
<tf.Tensor 'batch:14' shape=(32, 19, 19, 6) dtype=float32>,
<tf.Tensor 'batch:15' shape=(32, 10, 10, 6) dtype=float32>,
<tf.Tensor 'batch:16' shape=(32, 5, 5, 6) dtype=float32>,
<tf.Tensor 'batch:17' shape=(32, 3, 3, 4) dtype=float32>,
<tf.Tensor 'batch:18' shape=(32, 1, 1, 4) dtype=float32>]
此时的数据格式已经符合loss函数和网络输入要求，运行即可：
# Construct SSD network.
# 这个实例方法会返回之前定义的函数ssd_arg_scope（允许修改两个参数）
arg_scope = ssd_net.arg_scope(weight_decay=FLAGS.weight_decay,
data_format=DATA_FORMAT)
with slim.arg_scope(arg_scope):
# predictions: (BS, H, W, 4, 21)
# localisations: (BS, H, W, 4, 4)
# logits: (BS, H, W, 4, 21)
predictions, localisations, logits, end_points =
ssd_net.net(b_image, is_training=True)
# Add loss function.
ssd_net.losses(logits, localisations,
b_gclasses, b_glocalisations, b_gscores,
match_threshold=FLAGS.match_threshold, # .5
negative_ratio=FLAGS.negative_ratio, # 3
alpha=FLAGS.loss_alpha, # 1
label_smoothing=FLAGS.label_smoothing) # .0
正向传播函数会获取相关的节点，损失函数则会将函数值添加到loss collection中。