从ResNet101到ResNet50

一直用VGG训练，几天前想看下ResNet的效果如何，因为SSD源码中有python实现的ResNet网络结构实现代码，包含ResNet101和ResNet152，直接拿ResNet101来训练，GTX1060配置，batchsize竟然只降到2才跑的起来，果然一直收敛不了。看了下model_libs.py里面的实现代码：

def ResNet101Body(net, from_layer, use_pool5=True, use_dilation_conv5=False, **bn_param):
    conv_prefix = ''
    conv_postfix = ''
    bn_prefix = 'bn_'
    bn_postfix = ''
    scale_prefix = 'scale_'
    scale_postfix = ''
    ConvBNLayer(net, from_layer, 'conv1', use_bn=True, use_relu=True,
        num_output=64, kernel_size=7, pad=3, stride=2,
        conv_prefix=conv_prefix, conv_postfix=conv_postfix,
        bn_prefix=bn_prefix, bn_postfix=bn_postfix,
        scale_prefix=scale_prefix, scale_postfix=scale_postfix, **bn_param)

    net.pool1 = L.Pooling(net.conv1, pool=P.Pooling.MAX, kernel_size=3, stride=2)

    ResBody(net, 'pool1', '2a', out2a=64, out2b=64, out2c=256, stride=1, use_branch1=True, **bn_param)
    ResBody(net, 'res2a', '2b', out2a=64, out2b=64, out2c=256, stride=1, use_branch1=False, **bn_param)
    ResBody(net, 'res2b', '2c', out2a=64, out2b=64, out2c=256, stride=1, use_branch1=False, **bn_param)

    ResBody(net, 'res2c', '3a', out2a=128, out2b=128, out2c=512, stride=2, use_branch1=True, **bn_param)

    from_layer = 'res3a'
    for i in xrange(1, 4):
      block_name = '3b{}'.format(i)
      ResBody(net, from_layer, block_name, out2a=128, out2b=128, out2c=512, stride=1, use_branch1=False, **bn_param)
      from_layer = 'res{}'.format(block_name)

    ResBody(net, from_layer, '4a', out2a=256, out2b=256, out2c=1024, stride=2, use_branch1=True, **bn_param)

    from_layer = 'res4a'
    for i in xrange(1, 23):
      block_name = '4b{}'.format(i)
      ResBody(net, from_layer, block_name, out2a=256, out2b=256, out2c=1024, stride=1, use_branch1=False, **bn_param)
      from_layer = 'res{}'.format(block_name)

    stride = 2
    dilation = 1
    if use_dilation_conv5:
      stride = 1
      dilation = 2

    ResBody(net, from_layer, '5a', out2a=512, out2b=512, out2c=2048, stride=stride, use_branch1=True, dilation=dilation, **bn_param)
    ResBody(net, 'res5a', '5b', out2a=512, out2b=512, out2c=2048, stride=1, use_branch1=False, dilation=dilation, **bn_param)
    ResBody(net, 'res5b', '5c', out2a=512, out2b=512, out2c=2048, stride=1, use_branch1=False, dilation=dilation, **bn_param)

    if use_pool5:
      net.pool5 = L.Pooling(net.res5c, pool=P.Pooling.AVE, global_pooling=True)

    return net

RenNet152Body为：

def ResNet152Body(net, from_layer, use_pool5=True, use_dilation_conv5=False, **bn_param):
    conv_prefix = ''
    conv_postfix = ''
    bn_prefix = 'bn_'
    bn_postfix = ''
    scale_prefix = 'scale_'
    scale_postfix = ''
    ConvBNLayer(net, from_layer, 'conv1', use_bn=True, use_relu=True,
        num_output=64, kernel_size=7, pad=3, stride=2,
        conv_prefix=conv_prefix, conv_postfix=conv_postfix,
        bn_prefix=bn_prefix, bn_postfix=bn_postfix,
        scale_prefix=scale_prefix, scale_postfix=scale_postfix, **bn_param)

    net.pool1 = L.Pooling(net.conv1, pool=P.Pooling.MAX, kernel_size=3, stride=2)

    ResBody(net, 'pool1', '2a', out2a=64, out2b=64, out2c=256, stride=1, use_branch1=True, **bn_param)
    ResBody(net, 'res2a', '2b', out2a=64, out2b=64, out2c=256, stride=1, use_branch1=False, **bn_param)
    ResBody(net, 'res2b', '2c', out2a=64, out2b=64, out2c=256, stride=1, use_branch1=False, **bn_param)

    ResBody(net, 'res2c', '3a', out2a=128, out2b=128, out2c=512, stride=2, use_branch1=True, **bn_param)

    from_layer = 'res3a'
    for i in xrange(1, 8):
      block_name = '3b{}'.format(i)
      ResBody(net, from_layer, block_name, out2a=128, out2b=128, out2c=512, stride=1, use_branch1=False, **bn_param)
      from_layer = 'res{}'.format(block_name)

    ResBody(net, from_layer, '4a', out2a=256, out2b=256, out2c=1024, stride=2, use_branch1=True, **bn_param)

    from_layer = 'res4a'
    for i in xrange(1, 36):
      block_name = '4b{}'.format(i)
      ResBody(net, from_layer, block_name, out2a=256, out2b=256, out2c=1024, stride=1, use_branch1=False, **bn_param)
      from_layer = 'res{}'.format(block_name)

    stride = 2
    dilation = 1
    if use_dilation_conv5:
      stride = 1
      dilation = 2

    ResBody(net, from_layer, '5a', out2a=512, out2b=512, out2c=2048, stride=stride, use_branch1=True, dilation=dilation, **bn_param)
    ResBody(net, 'res5a', '5b', out2a=512, out2b=512, out2c=2048, stride=1, use_branch1=False, dilation=dilation, **bn_param)
    ResBody(net, 'res5b', '5c', out2a=512, out2b=512, out2c=2048, stride=1, use_branch1=False, dilation=dilation, **bn_param)

    if use_pool5:
      net.pool5 = L.Pooling(net.res5c, pool=P.Pooling.AVE, global_pooling=True)

    return net

其中每次调用ResBody，当use_brabch1 = True，会创建4个卷积层，当use_brabch1 = False时，创建3个卷积层。ResNet101Body和ResNet152Body的区别在于两个for循环的次数不一样，101层和152层差的51层就是这里体现的，所以现在要创建ResNet50Body就容易多了。根据网上下载的模型对应的ResNet_50_train_val.prototxt，对上面代码进行修改即可。50层，batchsize=4，训练马上收敛。当然训练方式多种，可用直接利用已有ResNet_50_train_val.prototxt进行训练。