tensorflow版yolo V1 逐行逐句解读
解读pascal_voc.py
class pascal_voc(object):
def __init__(self, phase, rebuild=False):
self.devkil_path = os.path.join(cfg.PASCAL_PATH, 'VOCdevkit')
self.data_path = os.path.join(self.devkil_path, 'VOC2007')
self.cache_path = cfg.CACHE_PATH
self.batch_size = cfg.BATCH_SIZE
self.image_size = cfg.IMAGE_SIZE
self.cell_size = cfg.CELL_SIZE
self.classes = cfg.CLASSES
self.class_to_ind = dict(zip(self.classes, range(len(self.classes))))
self.flipped = cfg.FLIPPED
self.phase = phase
self.rebuild = rebuild
self.cursor = 0
self.epoch = 1
self.gt_labels = None
self.prepare()
其中,devkil_path:是存放VOCdevkit的路径,
batch_size :这个应该大家都知道,就是返回批次的大小
image_size:就是图像的大小
cell_size:值得是我们把图像分割成多少快,classes是总共多少个类
在看这个类里面的方法把:
def image_read(self, imname, flipped=False):
image = cv2.imread(imname)
image = cv2.resize(image, (self.image_size, self.image_size))
image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB).astype(np.float32)
image = (image / 255.0) * 2.0 - 1.0
if flipped:
image = image[:, ::-1, :]
return image
这个是读取图片,并把图像resieze到image_size,因为opencv读取的bgr格式,同时转化为rgb格式,同时将图像归一化到[1-,1],flipped 是翻转
def load_pascal_annotation(self, index):
"""
Load image and bounding boxes info from XML file in the PASCAL VOC
format.
"""
imname = os.path.join(self.data_path, 'JPEGImages', index + '.jpg')
im = cv2.imread(imname)
h_ratio = 1.0 * self.image_size / im.shape[0]
w_ratio = 1.0 * self.image_size / im.shape[1]
# im = cv2.resize(im, [self.image_size, self.image_size])
label = np.zeros((self.cell_size, self.cell_size, 25))
filename = os.path.join(self.data_path, 'Annotations', index + '.xml')
tree = ET.parse(filename)
objs = tree.findall('object')
for obj in objs:
bbox = obj.find('bndbox')
# Make pixel indexes 0-based
x1 = max(min((float(bbox.find('xmin').text) - 1) * w_ratio, self.image_size - 1), 0)
y1 = max(min((float(bbox.find('ymin').text) - 1) * h_ratio, self.image_size - 1), 0)
x2 = max(min((float(bbox.find('xmax').text) - 1) * w_ratio, self.image_size - 1), 0)
y2 = max(min((float(bbox.find('ymax').text) - 1) * h_ratio, self.image_size - 1), 0)
cls_ind = self.class_to_ind[obj.find('name').text.lower().strip()]
boxes = [(x2 + x1) / 2.0, (y2 + y1) / 2.0, x2 - x1, y2 - y1]
x_ind = int(boxes[0] * self.cell_size / self.image_size)
y_ind = int(boxes[1] * self.cell_size / self.image_size)
if label[y_ind, x_ind, 0] == 1:
continue
label[y_ind, x_ind, 0] = 1
label[y_ind, x_ind, 1:5] = boxes
label[y_ind, x_ind, 5 + cls_ind] = 1
return label, len(objs)
这个地方就是从xml中解析出bbox的坐标,我们看下xml里面的东西长啥样吧:
其实我们就发现,框的信息放在object中的bndbox里面。所以可以看代码里面的ETtree的解析,w_ratio 和h_ratio就是因为我们的图像是resize的,所以,其实我们坐标也是要跟着变得,解析bbox的坐标,里面有一个min,这个是为了控制框不超过image_size,如果你的框超过了图像边界,我们就把它拉回来,外面有一个max,也是为了控制框在图像里面,当你的框坐标小于0的时候,也是把它拉回来,之所以对坐标减去1,是为了让像素也是从0开始,像素的索引从0开始。self.class_to_ind = dict(zip(self.classes, range(len(self.classes)))),这个是让我们的每一个类对应一个数字。boxes就是把左上角坐标和右下角坐标变成框的中心点坐标和长和宽。后面就是制作label,看x_int和y_int 这个是因为我们的输入图像最后是分成了7×7的网络,所以其实我们需要把我们的框的中心放在对应的cell中,所以乘以了7,(其中cell_size=7),也是为了和后面求loss统一,然后label里面的第一个值是表示置信度,第2-5是bbox的坐标,后面表示类别,ont-hot了。
def load_labels(self):
cache_file = os.path.join(
self.cache_path, 'pascal_' + self.phase + '_gt_labels.pkl')
if os.path.isfile(cache_file) and not self.rebuild:
print('Loading gt_labels from: ' + cache_file)
with open(cache_file, 'rb') as f:
gt_labels = pickle.load(f)
return gt_labels
print('Processing gt_labels from: ' + self.data_path)
if not os.path.exists(self.cache_path):
os.makedirs(self.cache_path)
if self.phase == 'train':
txtname = os.path.join(
self.data_path, 'ImageSets', 'Main', 'trainval.txt')
else:
txtname = os.path.join(
self.data_path, 'ImageSets', 'Main', 'test.txt')
with open(txtname, 'r') as f:
self.image_index = [x.strip() for x in f.readlines()]
gt_labels = []
for index in self.image_index:
label, num = self.load_pascal_annotation(index)
if num == 0:
continue
imname = os.path.join(self.data_path, 'JPEGImages', index + '.jpg')
gt_labels.append({'imname': imname,
'label': label,
'flipped': False})
print('Saving gt_labels to: ' + cache_file)
with open(cache_file, 'wb') as f:
pickle.dump(gt_labels, f)
return gt_labels
这个就是把所有图像都读取出来,调用我们上面讲的函数,load_pascal_annotation,一个一个做成label,并存在gt_labels里面。同时保存在pickle里面。
def prepare(self):
gt_labels = self.load_labels()
if self.flipped:
print('Appending horizontally-flipped training examples ...')
gt_labels_cp = copy.deepcopy(gt_labels)
for idx in range(len(gt_labels_cp)):
gt_labels_cp[idx]['flipped'] = True
gt_labels_cp[idx]['label'] =\
gt_labels_cp[idx]['label'][:, ::-1, :]
for i in range(self.cell_size):
for j in range(self.cell_size):
if gt_labels_cp[idx]['label'][i, j, 0] == 1:
gt_labels_cp[idx]['label'][i, j, 1] = \
self.image_size - 1 -\
gt_labels_cp[idx]['label'][i, j, 1]
gt_labels += gt_labels_cp
np.random.shuffle(gt_labels)
self.gt_labels = gt_labels
return gt_labels
这个就是调用我们前面的load_label函数,把所有的图像的bbox等信息都提出出来,如果flipped,其实就是数据增强,因为你flipped了,所以对应坐标之类的也要变,因为水平翻转,所以x坐标是要变得,y不变,所以gt_labels_cp[idx]['label'][i, j, 1] = self.image_size - 1 -gt_labels_cp[idx]['label'][i, j, 1] 其实就是求水平反转后的坐标。
最后把做了数据增强的返回,这个地方只是做了水平反转。同时把最终的结果传给gt_labels,self.gt_labels = gt_labels
然后再看,其实__init__也就是初始化里面self.prepare(),有这一句程序,所以当你新建类的对象的时候已经做好了。
def get(self):
images = np.zeros(
(self.batch_size, self.image_size, self.image_size, 3))
labels = np.zeros(
(self.batch_size, self.cell_size, self.cell_size, 25))
count = 0
while count < self.batch_size:
imname = self.gt_labels[self.cursor]['imname']
flipped = self.gt_labels[self.cursor]['flipped']
images[count, :, :, :] = self.image_read(imname, flipped)
labels[count, :, :, :] = self.gt_labels[self.cursor]['label']
count += 1
self.cursor += 1
if self.cursor >= len(self.gt_labels):
np.random.shuffle(self.gt_labels)
self.cursor = 0
self.epoch += 1
return images, labels
这个是我们train程序里面调用的,这里返回一个batch_size的images和labels,这个地方其实类似生成器,因为你不同调用get这个方法,因为这个是类,然后不停返回,如果全部遍历完了,由从头开始,shuffle一下,cursor又变为0了,
解读yolo_net.py 文件
def build_network(self,
images,
num_outputs,
alpha,
keep_prob=0.5,
is_training=True,
scope='yolo'):
with tf.variable_scope(scope):
with slim.arg_scope(
[slim.conv2d, slim.fully_connected],
activation_fn=leaky_relu(alpha),
weights_regularizer=slim.l2_regularizer(0.0005),
weights_initializer=tf.truncated_normal_initializer(0.0, 0.01)
):
net = tf.pad(
images, np.array([[0, 0], [3, 3], [3, 3], [0, 0]]),
name='pad_1')
net = slim.conv2d(
net, 64, 7, 2, padding='VALID', scope='conv_2')
net = slim.max_pool2d(net, 2, padding='SAME', scope='pool_3')
net = slim.conv2d(net, 192, 3, scope='conv_4')
net = slim.max_pool2d(net, 2, padding='SAME', scope='pool_5')
net = slim.conv2d(net, 128, 1, scope='conv_6')
net = slim.conv2d(net, 256, 3, scope='conv_7')
net = slim.conv2d(net, 256, 1, scope='conv_8')
net = slim.conv2d(net, 512, 3, scope='conv_9')
net = slim.max_pool2d(net, 2, padding='SAME', scope='pool_10')
net = slim.conv2d(net, 256, 1, scope='conv_11')
net = slim.conv2d(net, 512, 3, scope='conv_12')
net = slim.conv2d(net, 256, 1, scope='conv_13')
net = slim.conv2d(net, 512, 3, scope='conv_14')
net = slim.conv2d(net, 256, 1, scope='conv_15')
net = slim.conv2d(net, 512, 3, scope='conv_16')
net = slim.conv2d(net, 256, 1, scope='conv_17')
net = slim.conv2d(net, 512, 3, scope='conv_18')
net = slim.conv2d(net, 512, 1, scope='conv_19')
net = slim.conv2d(net, 1024, 3, scope='conv_20')
net = slim.max_pool2d(net, 2, padding='SAME', scope='pool_21')
net = slim.conv2d(net, 512, 1, scope='conv_22')
net = slim.conv2d(net, 1024, 3, scope='conv_23')
net = slim.conv2d(net, 512, 1, scope='conv_24')
net = slim.conv2d(net, 1024, 3, scope='conv_25')
net = slim.conv2d(net, 1024, 3, scope='conv_26')
net = tf.pad(
net, np.array([[0, 0], [1, 1], [1, 1], [0, 0]]),
name='pad_27')
net = slim.conv2d(
net, 1024, 3, 2, padding='VALID', scope='conv_28')
net = slim.conv2d(net, 1024, 3, scope='conv_29')
net = slim.conv2d(net, 1024, 3, scope='conv_30')
net = tf.transpose(net, [0, 3, 1, 2], name='trans_31')
net = slim.flatten(net, scope='flat_32')
net = slim.fully_connected(net, 512, scope='fc_33')
net = slim.fully_connected(net, 4096, scope='fc_34')
net = slim.dropout(
net, keep_prob=keep_prob, is_training=is_training,
scope='dropout_35')
net = slim.fully_connected(
net, num_outputs, activation_fn=None, scope='fc_36')
return net
def leaky_relu(alpha):
def op(inputs):
return tf.nn.leaky_relu(inputs, alpha=alpha, name='leaky_relu')
return op
网络的搭建比较简单,只是这里将下为啥有pad,这个是填充补数,保证最后卷积层最后的输出特征图大小是7×7,才需要pad操作。
def calc_iou(self, boxes1, boxes2, scope='iou'):
"""calculate ious
Args:
boxes1: 5-D tensor [BATCH_SIZE, CELL_SIZE, CELL_SIZE, BOXES_PER_CELL, 4] ====> (x_center, y_center, w, h)
boxes2: 5-D tensor [BATCH_SIZE, CELL_SIZE, CELL_SIZE, BOXES_PER_CELL, 4] ===> (x_center, y_center, w, h)
Return:
iou: 4-D tensor [BATCH_SIZE, CELL_SIZE, CELL_SIZE, BOXES_PER_CELL]
"""
with tf.variable_scope(scope):
# transform (x_center, y_center, w, h) to (x1, y1, x2, y2)
boxes1_t = tf.stack([boxes1[..., 0] - boxes1[..., 2] / 2.0,
boxes1[..., 1] - boxes1[..., 3] / 2.0,
boxes1[..., 0] + boxes1[..., 2] / 2.0,
boxes1[..., 1] + boxes1[..., 3] / 2.0],
axis=-1)
boxes2_t = tf.stack([boxes2[..., 0] - boxes2[..., 2] / 2.0,
boxes2[..., 1] - boxes2[..., 3] / 2.0,
boxes2[..., 0] + boxes2[..., 2] / 2.0,
boxes2[..., 1] + boxes2[..., 3] / 2.0],
axis=-1)
# calculate the left up point & right down point
lu = tf.maximum(boxes1_t[..., :2], boxes2_t[..., :2])
rd = tf.minimum(boxes1_t[..., 2:], boxes2_t[..., 2:])
# intersection
intersection = tf.maximum(0.0, rd - lu)
inter_square = intersection[..., 0] * intersection[..., 1]
# calculate the boxs1 square and boxs2 square
square1 = boxes1[..., 2] * boxes1[..., 3]
square2 = boxes2[..., 2] * boxes2[..., 3]
union_square = tf.maximum(square1 + square2 - inter_square, 1e-10)
return tf.clip_by_value(inter_square / union_square, 0.0, 1.0)
这个是计算IOU,思路就是把原始的中心点坐标和长和宽,转换成矩形框左上角和右下角的两个点的坐标,tf.stack是将数组/tensor堆叠起来,boxes_t就是把以前的中心点坐标和长和宽转换成了左上角和右下角的两个点的坐标。lu和rd就是分别求两个框相交的矩形的左上角的坐标和右下角的坐标,因为对于左上角,选择的是x和y较大的,而右下角是选择较小的,可以想想两个矩形框相交是不是这中情况。intersection这个就是求相交矩形的长和宽,所以有rd-ru,相当于x1-x2和y1-y2,之所以外面还要加一个tf.maximum是因为删除那些不合理的框,比如两个框没交集,就会出现左上角坐标比右下角还大。inter_square这个就是求面积了,就是长乘以宽。后面的square1和square2这个就是求面积了,因为之前是中心点坐标和长和宽,所以这里直接用长和宽,union_square就是就两个框的交面积,因为如果两个框的面积相加,那就会重复了相交的部分,所以减去相交的部分,外面有个tf.maximum这个就是保证相交面积不为0,因为后面要做分母。最后有一个tf.clip_by_value,这个是将如果你的交并比大于1,那么就让它等于1,如果小于0,那么就让他变为0,因为交并比在0-1之间。
解读loss_function
def loss_layer(self, predicts, labels, scope='loss_layer'):
with tf.variable_scope(scope):
predict_classes = tf.reshape(
predicts[:, :self.boundary1],
[self.batch_size, self.cell_size, self.cell_size, self.num_class])
predict_scales = tf.reshape(
predicts[:, self.boundary1:self.boundary2],
[self.batch_size, self.cell_size, self.cell_size, self.boxes_per_cell])
predict_boxes = tf.reshape(
predicts[:, self.boundary2:],
[self.batch_size, self.cell_size, self.cell_size, self.boxes_per_cell, 4])
response = tf.reshape(
labels[..., 0],
[self.batch_size, self.cell_size, self.cell_size, 1])
boxes = tf.reshape(
labels[..., 1:5],
[self.batch_size, self.cell_size, self.cell_size, 1, 4])
boxes = tf.tile(
boxes, [1, 1, 1, self.boxes_per_cell, 1]) / self.image_size
classes = labels[..., 5:]
offset = tf.reshape(
tf.constant(self.offset, dtype=tf.float32),
[1, self.cell_size, self.cell_size, self.boxes_per_cell])
offset = tf.tile(offset, [self.batch_size, 1, 1, 1])
offset_tran = tf.transpose(offset, (0, 2, 1, 3))
predict_boxes_tran = tf.stack(
[(predict_boxes[..., 0] + offset) / self.cell_size,
(predict_boxes[..., 1] + offset_tran) / self.cell_size,
tf.square(predict_boxes[..., 2]),
tf.square(predict_boxes[..., 3])], axis=-1)
iou_predict_truth = self.calc_iou(predict_boxes_tran, boxes)
# calculate I tensor [BATCH_SIZE, CELL_SIZE, CELL_SIZE, BOXES_PER_CELL]
object_mask = tf.reduce_max(iou_predict_truth, 3, keep_dims=True)
object_mask = tf.cast(
(iou_predict_truth >= object_mask), tf.float32) * response
# calculate no_I tensor [CELL_SIZE, CELL_SIZE, BOXES_PER_CELL]
noobject_mask = tf.ones_like(
object_mask, dtype=tf.float32) - object_mask
boxes_tran = tf.stack(
[boxes[..., 0] * self.cell_size - offset,
boxes[..., 1] * self.cell_size - offset_tran,
tf.sqrt(boxes[..., 2]),
tf.sqrt(boxes[..., 3])], axis=-1)
# class_loss
class_delta = response * (predict_classes - classes)
class_loss = tf.reduce_mean(
tf.reduce_sum(tf.square(class_delta), axis=[1, 2, 3]),
name='class_loss') * self.class_scale
# object_loss
object_delta = object_mask * (predict_scales - iou_predict_truth)
object_loss = tf.reduce_mean(
tf.reduce_sum(tf.square(object_delta), axis=[1, 2, 3]),
name='object_loss') * self.object_scale
# noobject_loss
noobject_delta = noobject_mask * predict_scales
noobject_loss = tf.reduce_mean(
tf.reduce_sum(tf.square(noobject_delta), axis=[1, 2, 3]),
name='noobject_loss') * self.noobject_scale
# coord_loss
coord_mask = tf.expand_dims(object_mask, 4)
boxes_delta = coord_mask * (predict_boxes - boxes_tran)
coord_loss = tf.reduce_mean(
tf.reduce_sum(tf.square(boxes_delta), axis=[1, 2, 3, 4]),
name='coord_loss') * self.coord_scale
tf.losses.add_loss(class_loss)
tf.losses.add_loss(object_loss)
tf.losses.add_loss(noobject_loss)
tf.losses.add_loss(coord_loss)
tf.summary.scalar('class_loss', class_loss)
tf.summary.scalar('object_loss', object_loss)
tf.summary.scalar('noobject_loss', noobject_loss)
tf.summary.scalar('coord_loss', coord_loss)
tf.summary.histogram('boxes_delta_x', boxes_delta[..., 0])
tf.summary.histogram('boxes_delta_y', boxes_delta[..., 1])
tf.summary.histogram('boxes_delta_w', boxes_delta[..., 2])
tf.summary.histogram('boxes_delta_h', boxes_delta[..., 3])
tf.summary.histogram('iou', iou_predict_truth)
这一部分可能是最难理解的,因为大神用了很多技巧,妙用,predict_classes这个就是把预测的结果中预测类别的提出来,显然这个地方前boundary1都是预测类别的,predict_scales这个是预测置信度的,predict_boxes预测框的。所以这三个就是把预测结果不同部分分开,提取出来。后面的response是提取label中的置信度,表示这个地方是否有框,boxes是提取框的,这个地方用到了tf.tile这个的化,其实就是一个复制操作,后面表示每个维度复制的次数,这个地方只是对轴3复制两次,因为我们知道一个cell_size负责两个框的预测。为了对应。classes就是类别了,这个地方的类别是one-hot编码的。后面有一个offset,这个其实是构造的,是为了将每个cell的坐标对齐,之前我们将读取文件的时候也提过,为啥要乘以cell_size,这个地方也是这样,就是后一个框比前一个框要多加1,比如我们预测了cell_size的每个中心点坐标,那么我们这个中心点落在第几个cell_size就对应坐标要加几,这个用法比较巧妙,构造了这样一个数组,让他们呢对应位置相加。predict_boxes_tran也就是用对应x和y加上了对应cell_size的值,这个地方为啥长和宽为啥要平方,我觉得因为它预测的是长和宽的平方根,为啥预测长和宽需要开方,因为原论文中作者说了,是为了较少大框对于偏移一点和小框偏移一点的误差应该不一样,应该是小框偏移一点误差更加大。iou_predict_truth求交并比。
后面有一个object_mask=tf.reduce_max(iou_predict_truth, 3, keep_dims=True),这个是求道每个框交并比最大的那个,因为我们知道,每个框只负责预测一个目标。所以object_mask就是有目标的,noobject_mask就是没目标的,使用tf.onr_like,使得全部为1,再减去有目标的,也就是有目标的对应坐标为1,这样一减,就变为没有的了。boxes_tran 这个就是把之前的坐标换回来,长和宽开方,后面求loss就方便。
后面就是计算loss,loss计算和原文的loss是一样的,
看懂上面,那么这个损失函数就会写了,就是和这个图上的损失函数一样的。
这次先讲到这里,下次我们分析train和test文件