模型結構

def downsample_basic_block(x, planes, stride):
    out = F.avg_pool3d(x, kernel_size=1, stride=stride)
    zero_pads = torch.Tensor(
        out.size(0), planes - out.size(1), out.size(2), out.size(3),
        out.size(4)).zero_()
    if isinstance(out.data, torch.cuda.FloatTensor):
        zero_pads = zero_pads.cuda()

    out
 
 = Variable(torch.cat([out.data, zero_pads], dim=1))

    return out


class ResNeXtBottleneck(nn.Module):
    expansion = 2

    def __init__(self, inplanes, planes, cardinality, stride=1,
                 downsample=None,conv3d_bias=True):
        super(ResNeXtBottleneck, self).__init__()
        mid_planes = cardinality * int(planes / 32
 
)
        self.conv1 = nn.Conv3d(inplanes, mid_planes, kernel_size=1, bias=conv3d_bias)
        self.bn1 = nn.BatchNorm3d(mid_planes)
        self.conv2 = nn.Conv3d(
            mid_planes,
            mid_planes,
            kernel_size=3,
            stride=stride,
            padding=1,
            groups=cardinality,
            bias=conv3d_bias)
        self
 
.bn2 = nn.BatchNorm3d(mid_planes)
        self.conv3 = nn.Conv3d(
            mid_planes, planes * self.expansion, kernel_size=1, bias=conv3d_bias)
        self.bn3 = nn.BatchNorm3d(planes * self.expansion)
        self.relu = nn.ReLU(inplace=True)
        self.downsample = downsample
        self.stride = stride

    def forward(self, x):
        residual = 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:
            residual = self.downsample(x)

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

        return out


class ResNeXt(nn.Module):

    def __init__(self,
                 block,
                 layers,
                 sample_size,
                 sample_duration,
                 shortcut_type='B',
                 cardinality=32,
                 num_classes=400,
                 conv3d_bias=True):
        self.conv3d_bias = conv3d_bias
        self.inplanes = 64
        super(ResNeXt, self).__init__()
        self.conv1 = nn.Conv3d(
            3,
            64,
            kernel_size=7,
            stride=2,
            padding=3,
            bias=conv3d_bias)
        self.bn1 = nn.BatchNorm3d(64)
        self.relu = nn.ReLU(inplace=True)
        self.maxpool = nn.MaxPool3d(kernel_size=(3, 3, 3), stride=2, padding=1)
        self.layer1 = self._make_layer(block, 128, layers[0], shortcut_type,
                                       cardinality)
        self.layer2 = self._make_layer(
            block, 256, layers[1], shortcut_type, cardinality, stride=2)
        self.layer3 = self._make_layer(
            block, 512, layers[2], shortcut_type, cardinality, stride=2)
        self.layer4 = self._make_layer(
            block, 1024, layers[3], shortcut_type, cardinality, stride=2)
        last_duration = int(math.ceil(sample_duration / 16))
        last_size = int(math.ceil(sample_size / 32))
        self.avgpool = nn.AvgPool3d(
            (last_duration, last_size, last_size), stride=1)
        self.fc = nn.Linear(cardinality * 32 * block.expansion, num_classes)

        for m in self.modules():
            if isinstance(m, nn.Conv3d):
                m.weight = nn.init.kaiming_normal_(m.weight, mode='fan_out')
            elif isinstance(m, nn.BatchNorm3d):
                m.weight.data.fill_(1)
                m.bias.data.zero_()

    def _make_layer(self,
                    block,
                    planes,
                    blocks,
                    shortcut_type,
                    cardinality,
                    stride=1):
        downsample = None
        if stride != 1 or self.inplanes != planes * block.expansion:
            if shortcut_type == 'A':
                downsample = partial(
                    downsample_basic_block,
                    planes=planes * block.expansion,
                    stride=stride)
            else:
                downsample = nn.Sequential(
                    nn.Conv3d(
                        self.inplanes,
                        planes * block.expansion,
                        kernel_size=1,
                        stride=stride,
                        bias=self.conv3d_bias), nn.BatchNorm3d(planes * block.expansion))

        layers = []
        layers.append(
            block(self.inplanes, planes, cardinality, stride, downsample, conv3d_bias=self.conv3d_bias))
        self.inplanes = planes * block.expansion
        for i in range(1, blocks):
            layers.append(block(self.inplanes, planes, cardinality, conv3d_bias=self.conv3d_bias))

        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

    pretrain = torch.load(opt.pretrain_path)
    assert opt.arch == pretrain['arch']
    from collections import OrderedDict
    new_state_dict = OrderedDict()
    for k, v in pretrain['state_dict'].items():
        print(k)
        name = k[7:]
        print(name) # remove `module.`
        new_state_dict[name] = v

程式結果

這裡之所以需要pretrain['state_dict']而不是直接使用model.load_state_dict(torch.load(opt.pretrain_path))是因為儲存模型的時候不但儲存了引數，還有周期，結構等資訊。

        states = {
            'epoch': epoch + 1,
            'arch': opt.arch,
            'state_dict': model.state_dict(),
            'optimizer': optimizer.state_dict(),
        }
        torch.save(states, save_file_path)

name = k[7:]去掉每一個引數名的前七個字元，因為下載的預訓練模型是在torch.nn.DataParallel 分散式下訓練的，而我只有單卡，所以需要去掉引數名前面的module，再load。

optimizer.state_dict() 有state 和param_groups 兩個key，其中param_groups的value如下所示。

    for k, v in model.named_parameters():
        print(k)

    for i in model.parameters():
        print(i)

print(model.fc)

返回

Linear(in_features=2048, out_features=90, bias=True)

print(model.layer4)

總結：首先構建計算圖，返回model。如果只想要一致的學習率，只需要再optimizer的第一個引數裡寫model.parameters()。model.parameters（）應該是個有順序的字典，此時len(optimizer.param_groups)等於1。如果想分別設定學習率等引數，可以如下設定，因為model.parameters()字典有順序，所以這裡列表新增的{‘params’: v, ‘lr’: 0.0}都沒有引數名，剩下的沒有定義的引數如momentum就按照optimizer引數設定的給每個都分配

    for i in range(opt.ft_begin_index, 5):
        ft_module_names.append('layer{}'.format(i))
    ft_module_names.append('fc')
    parameters = []
    for k, v in model.parameters():
        print(k)
        for ft_module in ft_module_names:
            if ft_module in k:
                parameters.append({'params': v})
                break
        else:
            parameters.append({'params': v, 'lr': 0.0})