以resnet50为例浅析经典模型在各个框架的实现

最近要在paddlepaddle上实现mnasnet，要对各大框架实现有所了解。resnet50是比较经典的用于图像分类的网络，其在各大框架均有实现。

我们先来看caffe的prototxt实现，大改要写2300多行代码，不仅写的费劲，而且很容易出错，这里就不贴代码了.

通过Netscopes可视化

当然也有有用python写的，很是简洁

import caffe
from caffe import layers as L
from caffe import params as P


def conv_bn_scale_relu(bottom, num_output=64, kernel_size=3, stride=1, pad=0):
    conv = L.Convolution(bottom, num_output=num_output, kernel_size=kernel_size, stride=stride, pad=pad,
                         param=[dict(lr_mult=1, decay_mult=1), dict(lr_mult=2, decay_mult=0)],
                         weight_filler=dict(type='xavier', std=0.01),
                         bias_filler=dict(type='constant', value=0))
    conv_bn = L.BatchNorm(conv, use_global_stats=False, in_place=True)
    conv_scale = L.Scale(conv, scale_param=dict(bias_term=True), in_place=True)
    conv_relu = L.ReLU(conv, in_place=True)

    return conv, conv_bn, conv_scale, conv_relu


def conv_bn_scale(bottom, num_output=64, kernel_size=3, stride=1, pad=0):
    conv = L.Convolution(bottom, num_output=num_output, kernel_size=kernel_size, stride=stride, pad=pad,
                         param=[dict(lr_mult=1, decay_mult=1), dict(lr_mult=2, decay_mult=0)],
                         weight_filler=dict(type='xavier', std=0.01),
                         bias_filler=dict(type='constant', value=0.2))
    conv_bn = L.BatchNorm(conv, use_global_stats=False, in_place=True)
    conv_scale = L.Scale(conv, scale_param=dict(bias_term=True), in_place=True)

    return conv, conv_bn, conv_scale


def eltwize_relu(bottom1, bottom2):
    residual_eltwise = L.Eltwise(bottom1, bottom2, eltwise_param=dict(operation=1))
    residual_eltwise_relu = L.ReLU(residual_eltwise, in_place=True)

    return residual_eltwise, residual_eltwise_relu


def residual_branch(bottom, base_output=64):
    """
    input:4*base_output x n x n
    output:4*base_output x n x n
    :param base_output: base num_output of branch2
    :param bottom: bottom layer
    :return: layers
    """
    branch2a, branch2a_bn, branch2a_scale, branch2a_relu = \
        conv_bn_scale_relu(bottom, num_output=base_output, kernel_size=1)  # base_output x n x n
    branch2b, branch2b_bn, branch2b_scale, branch2b_relu = \
        conv_bn_scale_relu(branch2a, num_output=base_output, kernel_size=3, pad=1)  # base_output x n x n
    branch2c, branch2c_bn, branch2c_scale = \
        conv_bn_scale(branch2b, num_output=4 * base_output, kernel_size=1)  # 4*base_output x n x n

    residual, residual_relu = \
        eltwize_relu(bottom, branch2c)  # 4*base_output x n x n

    return branch2a, branch2a_bn, branch2a_scale, branch2a_relu, branch2b, branch2b_bn, branch2b_scale, branch2b_relu, \
           branch2c, branch2c_bn, branch2c_scale, residual, residual_relu


def residual_branch_shortcut(bottom, stride=2, base_output=64):
    """

    :param stride: stride
    :param base_output: base num_output of branch2
    :param bottom: bottom layer
    :return: layers
    """
    branch1, branch1_bn, branch1_scale = \
        conv_bn_scale(bottom, num_output=4 * base_output, kernel_size=1, stride=stride)

    branch2a, branch2a_bn, branch2a_scale, branch2a_relu = \
        conv_bn_scale_relu(bottom, num_output=base_output, kernel_size=1, stride=stride)
    branch2b, branch2b_bn, branch2b_scale, branch2b_relu = \
        conv_bn_scale_relu(branch2a, num_output=base_output, kernel_size=3, pad=1)
    branch2c, branch2c_bn, branch2c_scale = \
        conv_bn_scale(branch2b, num_output=4 * base_output, kernel_size=1)

    residual, residual_relu = \
        eltwize_relu(branch1, branch2c)  # 4*base_output x n x n

    return branch1, branch1_bn, branch1_scale, branch2a, branch2a_bn, branch2a_scale, branch2a_relu, branch2b, \
           branch2b_bn, branch2b_scale, branch2b_relu, branch2c, branch2c_bn, branch2c_scale, residual, residual_relu


branch_shortcut_string = 'n.res(stage)a_branch1, n.res(stage)a_branch1_bn, n.res(stage)a_branch1_scale, \
        n.res(stage)a_branch2a, n.res(stage)a_branch2a_bn, n.res(stage)a_branch2a_scale, n.res(stage)a_branch2a_relu, \
        n.res(stage)a_branch2b, n.res(stage)a_branch2b_bn, n.res(stage)a_branch2b_scale, n.res(stage)a_branch2b_relu, \
        n.res(stage)a_branch2c, n.res(stage)a_branch2c_bn, n.res(stage)a_branch2c_scale, n.res(stage)a, n.res(stage)a_relu = \
            residual_branch_shortcut((bottom), stride=(stride), base_output=(num))'

branch_string = 'n.res(stage)b(order)_branch2a, n.res(stage)b(order)_branch2a_bn, n.res(stage)b(order)_branch2a_scale, \
        n.res(stage)b(order)_branch2a_relu, n.res(stage)b(order)_branch2b, n.res(stage)b(order)_branch2b_bn, \
        n.res(stage)b(order)_branch2b_scale, n.res(stage)b(order)_branch2b_relu, n.res(stage)b(order)_branch2c, \
        n.res(stage)b(order)_branch2c_bn, n.res(stage)b(order)_branch2c_scale, n.res(stage)b(order), n.res(stage)b(order)_relu = \
            residual_branch((bottom), base_output=(num))'


class ResNet(object):
    def __init__(self, lmdb_train, lmdb_test, num_output):
        self.train_data = lmdb_train
        self.test_data = lmdb_test
        self.classifier_num = num_output

    def resnet_layers_proto(self, batch_size, phase='TRAIN', stages=(3, 4, 6, 3)):
        """

        :param batch_size: the batch_size of train and test phase
        :param phase: TRAIN or TEST
        :param stages: the num of layers = 2 + 3*sum(stages), layers would better be chosen from [50, 101, 152]
                       {every stage is composed of 1 residual_branch_shortcut module and stage[i]-1 residual_branch
                       modules, each module consists of 3 conv layers}
                        (3, 4, 6, 3) for 50 layers; (3, 4, 23, 3) for 101 layers; (3, 8, 36, 3) for 152 layers
        """
        n = caffe.NetSpec()
        if phase == 'TRAIN':
            source_data = self.train_data
            mirror = True
        else:
            source_data = self.test_data
            mirror = False
        n.data, n.label = L.Data(source=source_data, backend=P.Data.LMDB, batch_size=batch_size, ntop=2,
                                 transform_param=dict(crop_size=224, mean_value=[104, 117, 123], mirror=mirror))

        n.conv1, n.conv1_bn, n.conv1_scale, n.conv1_relu = \
            conv_bn_scale_relu(n.data, num_output=64, kernel_size=7, stride=2, pad=3)  # 64x112x112
        n.pool1 = L.Pooling(n.conv1, kernel_size=3, stride=2, pool=P.Pooling.MAX)  # 64x56x56

        for num in xrange(len(stages)):  # num = 0, 1, 2, 3
            for i in xrange(stages[num]):
                if i == 0:
                    stage_string = branch_shortcut_string
                    bottom_string = ['n.pool1', 'n.res2b%s' % str(stages[0] - 1), 'n.res3b%s' % str(stages[1] - 1),
                                     'n.res4b%s' % str(stages[2] - 1)][num]
                else:
                    stage_string = branch_string
                    if i == 1:
                        bottom_string = 'n.res%sa' % str(num + 2)
                    else:
                        bottom_string = 'n.res%sb%s' % (str(num + 2), str(i - 1))
                exec (stage_string.replace('(stage)', str(num + 2)).replace('(bottom)', bottom_string).
                      replace('(num)', str(2 ** num * 64)).replace('(order)', str(i)).
                      replace('(stride)', str(int(num > 0) + 1)))

        exec 'n.pool5 = L.Pooling((bottom), pool=P.Pooling.AVE, global_pooling=True)'.\
            replace('(bottom)', 'n.res5b%s' % str(stages[3] - 1))
        n.classifier = L.InnerProduct(n.pool5, num_output=self.classifier_num,
                                      param=[dict(lr_mult=1, decay_mult=1), dict(lr_mult=2, decay_mult=0)],
                                      weight_filler=dict(type='xavier'),
                                      bias_filler=dict(type='constant', value=0))
        n.loss = L.SoftmaxWithLoss(n.classifier, n.label)
        if phase == 'TRAIN':
            pass
        else:
            n.accuracy_top1 = L.Accuracy(n.classifier, n.label, include=dict(phase=1))
            n.accuracy_top5 = L.Accuracy(n.classifier, n.label, include=dict(phase=1),
                                         accuracy_param=dict(top_k=5))

        return n.to_proto()

再来看pytorch的，已然少了很多，把很多公共的操作封装了起来，不足之处在于对新手不友好，不太能一下就看出意思

import torch.nn as nn
import torch.utils.model_zoo as model_zoo


__all__ = ['ResNet', 'resnet18', 'resnet34', 'resnet50', 'resnet101',
           'resnet152']


model_urls = {
    'resnet18': 'https://download.pytorch.org/models/resnet18-5c106cde.pth',
    'resnet34': 'https://download.pytorch.org/models/resnet34-333f7ec4.pth',
    'resnet50': 'https://download.pytorch.org/models/resnet50-19c8e357.pth',
    'resnet101': 'https://download.pytorch.org/models/resnet101-5d3b4d8f.pth',
    'resnet152': 'https://download.pytorch.org/models/resnet152-b121ed2d.pth',
}


def conv3x3(in_planes, out_planes, stride=1):
    """3x3 convolution with padding"""
    return nn.Conv2d(in_planes, out_planes, kernel_size=3, stride=stride,
                     padding=1, bias=False)


def conv1x1(in_planes, out_planes, stride=1):
    """1x1 convolution"""
    return nn.Conv2d(in_planes, out_planes, kernel_size=1, stride=stride, bias=False)


class BasicBlock(nn.Module):
    expansion = 1

    def __init__(self, inplanes, planes, stride=1, downsample=None):
        super(BasicBlock, self).__init__()
        self.conv1 = conv3x3(inplanes, planes, stride)
        self.bn1 = nn.BatchNorm2d(planes)
        self.relu = nn.ReLU(inplace=True)
        self.conv2 = conv3x3(planes, planes)
        self.bn2 = nn.BatchNorm2d(planes)
        self.downsample = downsample
        self.stride = stride

    def forward(self, x):
        identity = x

        out = self.conv1(x)
        out = self.bn1(out)
        out = self.relu(out)

        out = self.conv2(out)
        out = self.bn2(out)

        if self.downsample is not None:
            identity = self.downsample(x)

        out += identity
        out = self.relu(out)

        return out


class Bottleneck(nn.Module):
    expansion = 4

    def __init__(self, inplanes, planes, stride=1, downsample=None):
        super(Bottleneck, self).__init__()
        self.conv1 = conv1x1(inplanes, planes)
        self.bn1 = nn.BatchNorm2d(planes)
        self.conv2 = conv3x3(planes, planes, stride)
        self.bn2 = nn.BatchNorm2d(planes)
        self.conv3 = conv1x1(planes, planes * self.expansion)
        self.bn3 = nn.BatchNorm2d(planes * self.expansion)
        self.relu = nn.ReLU(inplace=True)
        self.downsample = downsample
        self.stride = stride

    def forward(self, x):
        identity = x

        out = self.conv1(x)
        out = self.bn1(out)
        out = self.relu(out)

        out = self.conv2(out)
        out = self.bn2(out)
        out = self.relu(out)

        out = self.conv3(out)
        out = self.bn3(out)

        if self.downsample is not None:
            identity = self.downsample(x)

        out += identity
        out = self.relu(out)

        return out


class ResNet(nn.Module):

    def __init__(self, block, layers, num_classes=1000, zero_init_residual=False):
        super(ResNet, self).__init__()
        self.inplanes = 64
        self.conv1 = nn.Conv2d(3, 64, kernel_size=7, stride=2, padding=3,
                               bias=False)
        self.bn1 = nn.BatchNorm2d(64)
        self.relu = nn.ReLU(inplace=True)
        self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
        self.layer1 = self._make_layer(block, 64, layers[0])
        self.layer2 = self._make_layer(block, 128, layers[1], stride=2)
        self.layer3 = self._make_layer(block, 256, layers[2], stride=2)
        self.layer4 = self._make_layer(block, 512, layers[3], stride=2)
        self.avgpool = nn.AdaptiveAvgPool2d((1, 1))
        self.fc = nn.Linear(512 * block.expansion, num_classes)

        for m in self.modules():
            if isinstance(m, nn.Conv2d):
                nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu')
            elif isinstance(m, nn.BatchNorm2d):
                nn.init.constant_(m.weight, 1)
                nn.init.constant_(m.bias, 0)

        # Zero-initialize the last BN in each residual branch,
        # so that the residual branch starts with zeros, and each residual block behaves like an identity.
        # This improves the model by 0.2~0.3% according to https://arxiv.org/abs/1706.02677
        if zero_init_residual:
            for m in self.modules():
                if isinstance(m, Bottleneck):
                    nn.init.constant_(m.bn3.weight, 0)
                elif isinstance(m, BasicBlock):
                    nn.init.constant_(m.bn2.weight, 0)

    def _make_layer(self, block, planes, blocks, stride=1):
        downsample = None
        if stride != 1 or self.inplanes != planes * block.expansion:
            downsample = nn.Sequential(
                conv1x1(self.inplanes, planes * block.expansion, stride),
                nn.BatchNorm2d(planes * block.expansion),
            )

        layers = []
        layers.append(block(self.inplanes, planes, stride, downsample))
        self.inplanes = planes * block.expansion
        for _ in range(1, blocks):
            layers.append(block(self.inplanes, planes))

        return nn.Sequential(*layers)

    def forward(self, x):
        x = self.conv1(x)
        x = self.bn1(x)
        x = self.relu(x)
        x = self.maxpool(x)

        x = self.layer1(x)
        x = self.layer2(x)
        x = self.layer3(x)
        x = self.layer4(x)

        x = self.avgpool(x)
        x = x.view(x.size(0), -1)
        x = self.fc(x)

        return x


def resnet18(pretrained=False, **kwargs):
    """Constructs a ResNet-18 model.

    Args:
        pretrained (bool): If True, returns a model pre-trained on ImageNet
    """
    model = ResNet(BasicBlock, [2, 2, 2, 2], **kwargs)
    if pretrained:
        model.load_state_dict(model_zoo.load_url(model_urls['resnet18']))
    return model


def resnet34(pretrained=False, **kwargs):
    """Constructs a ResNet-34 model.

    Args:
        pretrained (bool): If True, returns a model pre-trained on ImageNet
    """
    model = ResNet(BasicBlock, [3, 4, 6, 3], **kwargs)
    if pretrained:
        model.load_state_dict(model_zoo.load_url(model_urls['resnet34']))
    return model


def resnet50(pretrained=False, **kwargs):
    """Constructs a ResNet-50 model.

    Args:
        pretrained (bool): If True, returns a model pre-trained on ImageNet
    """
    model = ResNet(Bottleneck, [3, 4, 6, 3], **kwargs)
    if pretrained:
        model.load_state_dict(model_zoo.load_url(model_urls['resnet50']))
    return model


def resnet101(pretrained=False, **kwargs):
    """Constructs a ResNet-101 model.

    Args:
        pretrained (bool): If True, returns a model pre-trained on ImageNet
    """
    model = ResNet(Bottleneck, [3, 4, 23, 3], **kwargs)
    if pretrained:
        model.load_state_dict(model_zoo.load_url(model_urls['resnet101']))
    return model


def resnet152(pretrained=False, **kwargs):
    """Constructs a ResNet-152 model.

    Args:
        pretrained (bool): If True, returns a model pre-trained on ImageNet
    """
    model = ResNet(Bottleneck, [3, 8, 36, 3], **kwargs)
    if pretrained:
        model.load_state_dict(model_zoo.load_url(model_urls['resnet152']))
    return model

最后是paddlepaddle的

from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import paddle
import paddle.fluid as fluid
import math

__all__ = ["ResNet", "ResNet50", "ResNet101", "ResNet152"]

train_parameters = {
    "input_size": [3, 224, 224],
    "input_mean": [0.485, 0.456, 0.406],
    "input_std": [0.229, 0.224, 0.225],
    "learning_strategy": {
        "name": "piecewise_decay",
        "batch_size": 256,
        "epochs": [30, 60, 90],
        "steps": [0.1, 0.01, 0.001, 0.0001]
    }
}


class ResNet():
    def __init__(self, layers=50):
        self.params = train_parameters
        self.layers = layers

    def net(self, input, class_dim=1000):
        layers = self.layers
        supported_layers = [50, 101, 152]
        assert layers in supported_layers, \
            "supported layers are {} but input layer is {}".format(supported_layers, layers)

        if layers == 50:
            depth = [3, 4, 6, 3]
        elif layers == 101:
            depth = [3, 4, 23, 3]
        elif layers == 152:
            depth = [3, 8, 36, 3]
        num_filters = [64, 128, 256, 512]

        conv = self.conv_bn_layer(
            input=input, num_filters=64, filter_size=7, stride=2, act='relu')
        conv = fluid.layers.pool2d(
            input=conv,
            pool_size=3,
            pool_stride=2,
            pool_padding=1,
            pool_type='max')

        for block in range(len(depth)):
            for i in range(depth[block]):
                conv = self.bottleneck_block(
                    input=conv,
                    num_filters=num_filters[block],
                    stride=2 if i == 0 and block != 0 else 1)

        pool = fluid.layers.pool2d(
            input=conv, pool_size=7, pool_type='avg', global_pooling=True)
        stdv = 1.0 / math.sqrt(pool.shape[1] * 1.0)
        out = fluid.layers.fc(input=pool,
                              size=class_dim,
                              act='softmax',
                              param_attr=fluid.param_attr.ParamAttr(
                                  initializer=fluid.initializer.Uniform(-stdv,
                                                                        stdv)))
        return out

    def conv_bn_layer(self,
                      input,
                      num_filters,
                      filter_size,
                      stride=1,
                      groups=1,
                      act=None):
        conv = fluid.layers.conv2d(
            input=input,
            num_filters=num_filters,
            filter_size=filter_size,
            stride=stride,
            padding=(filter_size - 1) // 2,
            groups=groups,
            act=None,
            bias_attr=False)
        return fluid.layers.batch_norm(input=conv, act=act)

    def shortcut(self, input, ch_out, stride):
        ch_in = input.shape[1]
        if ch_in != ch_out or stride != 1:
            return self.conv_bn_layer(input, ch_out, 1, stride)
        else:
            return input

    def bottleneck_block(self, input, num_filters, stride):
        conv0 = self.conv_bn_layer(
            input=input, num_filters=num_filters, filter_size=1, act='relu')
        conv1 = self.conv_bn_layer(
            input=conv0,
            num_filters=num_filters,
            filter_size=3,
            stride=stride,
            act='relu')
        conv2 = self.conv_bn_layer(
            input=conv1, num_filters=num_filters * 4, filter_size=1, act=None)

        short = self.shortcut(input, num_filters * 4, stride)

        return fluid.layers.elementwise_add(x=short, y=conv2, act='relu')


def ResNet50():
    model = ResNet(layers=50)
    return model


def ResNet101():
    model = ResNet(layers=101)
    return model


def ResNet152():
    model = ResNet(layers=152)
    return model

当然，上面存在的主要问题是没对结构做清晰的说明以及实验案例，好在PaddlePaddle-ResNet-ImageClassification给我们准备了全套方案。

最后附上简化版结构图