2021SC@SDUSC
骨干网络
代码位置：ppocr->modeling->backbones->det_resnet_vd.py
网络退化问题：网络深度增加时，网络准确度出现饱和，甚至出现下降。
ResNet网络是参考了VGG19网络，在其基础上进行了修改，并通过短路机制加入了残差单元，如下图所示。变化主要体现在ResNet直接使用stride=2的卷积做下采样，并且用global average pool层替换了全连接层。ResNet的一个重要设计原则是：当feature map大小降低一半时，feature map的数量增加一倍，这保持了网络层的复杂度。从图5中可以看到，ResNet相比普通网络每两层间增加了短路机制，这就形成了残差学习，其中虚线表示feature map数量发生了改变。图5展示的34-layer的ResNet，还可以构建更深的网络如表1所示。从表中可以看到，对于18-layer和34-layer的ResNet，其进行的两层间的残差学习，当网络更深时，其进行的是三层间的残差学习，三层卷积核分别是1x1，3x3和1x1，一个值得注意的是隐含层的feature map数量是比较小的，并且是输出feature map数量的1/4。
class ResNet(nn.Layer):
def __init__(self, in_channels=3, layers=50, **kwargs):
super(ResNet, self).__init__()
self.layers = layers
supported_layers = [18, 34, 50, 101, 152, 200]
assert layers in supported_layers, \
"supported layers are {} but input layer is {}".format(
supported_layers, layers)
if layers == 18:
depth = [2, 2, 2, 2]
elif layers == 34 or layers == 50:
depth = [3, 4, 6, 3]
elif layers == 101:
depth = [3, 4, 23, 3]
elif layers == 152:
depth = [3, 8, 36, 3]
elif layers == 200:
depth = [3, 12, 48, 3]
num_channels = [64, 256, 512,
1024] if layers >= 50 else [64, 64, 128, 256]
num_filters = [64, 128, 256, 512]
self.conv1_1 = ConvBNLayer(
in_channels=in_channels,
out_channels=32,
kernel_size=3,
stride=2,
act='relu',
name="conv1_1")
self.conv1_2 = ConvBNLayer(
in_channels=32,
out_channels=32,
kernel_size=3,
stride=1,
act='relu',
name="conv1_2")
self.conv1_3 = ConvBNLayer(
in_channels=32,
out_channels=64,
kernel_size=3,
stride=1,
act='relu',
name="conv1_3")
self.pool2d_max = nn.MaxPool2D(kernel_size=3, stride=2, padding=1)
self.stages = []
self.out_channels = []
if layers >= 50:
for block in range(len(depth)):
block_list = []
shortcut = False
for i in range(depth[block]):
if layers in [101, 152] and block == 2:
if i == 0:
conv_name = "res" + str(block + 2) + "a"
else:
conv_name = "res" + str(block + 2) + "b" + str(i)
else:
conv_name = "res" + str(block + 2) + chr(97 + i)
bottleneck_block = self.add_sublayer(
'bb_%d_%d' % (block, i),
BottleneckBlock(
in_channels=num_channels[block]
if i == 0 else num_filters[block] * 4,
out_channels=num_filters[block],
stride=2 if i == 0 and block != 0 else 1,
shortcut=shortcut,
if_first=block == i == 0,
name=conv_name))
shortcut = True
block_list.append(bottleneck_block)
self.out_channels.append(num_filters[block] * 4)
self.stages.append(nn.Sequential(*block_list))
else:
for block in range(len(depth)):
block_list = []
shortcut = False
for i in range(depth[block]):
conv_name = "res" + str(block + 2) + chr(97 + i)
basic_block = self.add_sublayer(
'bb_%d_%d' % (block, i),
BasicBlock(
in_channels=num_channels[block]
if i == 0 else num_filters[block],
out_channels=num_filters[block],
stride=2 if i == 0 and block != 0 else 1,
shortcut=shortcut,
if_first=block == i == 0,
name=conv_name))
shortcut = True
block_list.append(basic_block)
self.out_channels.append(num_filters[block])
self.stages.append(nn.Sequential(*block_list))
ResNet是一种残差网络，咱们可以把它理解为一个子网络，这个子网络经过堆叠可以构成一个很深的网络。
上表是Resnet不同的结构，上表一共提出了5中深度的ResNet，分别是18，34，50，101和152，首先看表的最左侧，我们发现所有的网络都分成5部分，分别是：conv1，conv2_x，conv3_x，conv4_x，conv5_x，之后的其他论文也会专门用这个称呼指代ResNet50或者101的每部分。 例如：101-layer那列，101-layer指的是101层网络，首先有个输入7x7x64的卷积，然后经过3 + 4 + 23 + 3 = 33个building block，每个block为3层，所以有33 x 3 = 99层，最后有个fc层(用于分类)，所以1 + 99 + 1 = 101层，确实有101层网络； 注：101层网络仅仅指卷积或者全连接层，而激活层或者Pooling层并没有计算在内；我们关注50-layer和101-layer这两列，可以发现，它们唯一的不同在于conv4_x，ResNet50有6个block，而ResNet101有23个block，两者之间差了17个block，也就是17 x 3 = 51层。