mmdetection自定义backbone(转)

前面我们零散的了解了mmdetection用到的一些python和PyTorch的知识。

现在我们开始深入算法模型来学习，这个模块我们尝试自定义一个se_resnet50来学习自定义backbone.

在mmdetection/mmdet/models/backbone文件下创建一个senet.py的python文件。

"""
ResNet code gently borrowed from
https://github.com/pytorch/vision/blob/master/torchvision/models/resnet.py
"""
from __future__ import print_function, division, absolute_import
from collections import OrderedDict
from ..registry import BACKBONES
from mmcv.runner import load_checkpoint
import logging
from mmcv.cnn import constant_init, kaiming_init
import math
 
import torch.nn as nn
from torch.utils import model_zoo
 
__all__ = ['SENet', 'senet154', 'se_resnet50', 'se_resnet101', 'se_resnet152',
           'se_resnext50_32x4d', 'se_resnext101_32x4d']
 
pretrained_settings = {
    'senet154': {
        'imagenet': {
            'url': 'http://data.lip6.fr/cadene/pretrainedmodels/senet154-c7b49a05.pth',
            'input_space': 'RGB',
            'input_size': [3, 224, 224],
            'input_range': [0, 1],
            'mean': [0.485, 0.456, 0.406],
            'std': [0.229, 0.224, 0.225],
            'num_classes': 1000
        }
    },
    'se_resnet50': {
        'imagenet': {
            'url': 'http://data.lip6.fr/cadene/pretrainedmodels/se_resnet50-ce0d4300.pth',
            'input_space': 'RGB',
            'input_size': [3, 224, 224],
            'input_range': [0, 1],
            'mean': [0.485, 0.456, 0.406],
            'std': [0.229, 0.224, 0.225],
            'num_classes': 1000
        }
    },
    'se_resnet101': {
        'imagenet': {
            'url': 'http://data.lip6.fr/cadene/pretrainedmodels/se_resnet101-7e38fcc6.pth',
            'input_space': 'RGB',
            'input_size': [3, 224, 224],
            'input_range': [0, 1],
            'mean': [0.485, 0.456, 0.406],
            'std': [0.229, 0.224, 0.225],
            'num_classes': 1000
        }
    },
    'se_resnet152': {
        'imagenet': {
            'url': 'http://data.lip6.fr/cadene/pretrainedmodels/se_resnet152-d17c99b7.pth',
            'input_space': 'RGB',
            'input_size': [3, 224, 224],
            'input_range': [0, 1],
            'mean': [0.485, 0.456, 0.406],
            'std': [0.229, 0.224, 0.225],
            'num_classes': 1000
        }
    },
    'se_resnext50_32x4d': {
        'imagenet': {
            'url': 'http://data.lip6.fr/cadene/pretrainedmodels/se_resnext50_32x4d-a260b3a4.pth',
            'input_space': 'RGB',
            'input_size': [3, 224, 224],
            'input_range': [0, 1],
            'mean': [0.485, 0.456, 0.406],
            'std': [0.229, 0.224, 0.225],
            'num_classes': 1000
        }
    },
    'se_resnext101_32x4d': {
        'imagenet': {
            'url': 'http://data.lip6.fr/cadene/pretrainedmodels/se_resnext101_32x4d-3b2fe3d8.pth',
            'input_space': 'RGB',
            'input_size': [3, 224, 224],
            'input_range': [0, 1],
            'mean': [0.485, 0.456, 0.406],
            'std': [0.229, 0.224, 0.225],
            'num_classes': 1000
        }
    },
}
 
 
class SEModule(nn.Module):
 
    def __init__(self, channels, reduction):
        super(SEModule, self).__init__()
        self.avg_pool = nn.AdaptiveAvgPool2d(1)
        self.fc1 = nn.Conv2d(channels, channels // reduction, kernel_size=1,
                             padding=0)
        self.relu = nn.ReLU(inplace=True)
        self.fc2 = nn.Conv2d(channels // reduction, channels, kernel_size=1,
                             padding=0)
        self.sigmoid = nn.Sigmoid()
 
    def forward(self, x):
        module_input = x
        x = self.avg_pool(x)
        x = self.fc1(x)
        x = self.relu(x)
        x = self.fc2(x)
        x = self.sigmoid(x)
        return module_input * x
 
 
 
class Bottleneck(nn.Module):
    """
    Base class for bottlenecks that implements `forward()` method.
    """
 
    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 = self.se_module(out) + residual
        out = self.relu(out)
 
        return out
 
 
class SEBottleneck(Bottleneck):
    """
    Bottleneck for SENet154.
    """
    expansion = 4
 
    def __init__(self, inplanes, planes, groups, reduction, stride=1,
                 downsample=None):
        super(SEBottleneck, self).__init__()
        self.conv1 = nn.Conv2d(inplanes, planes * 2, kernel_size=1, bias=False)
        self.bn1 = nn.BatchNorm2d(planes * 2)
        self.conv2 = nn.Conv2d(planes * 2, planes * 4, kernel_size=3,
                               stride=stride, padding=1, groups=groups,
                               bias=False)
        self.bn2 = nn.BatchNorm2d(planes * 4)
        self.conv3 = nn.Conv2d(planes * 4, planes * 4, kernel_size=1,
                               bias=False)
        self.bn3 = nn.BatchNorm2d(planes * 4)
        self.relu = nn.ReLU(inplace=True)
        self.se_module = SEModule(planes * 4, reduction=reduction)
        self.downsample = downsample
        self.stride = stride
 
 
class SEResNetBottleneck(Bottleneck):
    """
    ResNet bottleneck with a Squeeze-and-Excitation module. It follows Caffe
    implementation and uses `stride=stride` in `conv1` and not in `conv2`
    (the latter is used in the torchvision implementation of ResNet).
    """
    expansion = 4
 
    def __init__(self, inplanes, planes, groups, reduction, stride=1,
                 downsample=None):
        super(SEResNetBottleneck, self).__init__()
        self.conv1 = nn.Conv2d(inplanes, planes, kernel_size=1, bias=False,
                               stride=stride)
        self.bn1 = nn.BatchNorm2d(planes)
        self.conv2 = nn.Conv2d(planes, planes, kernel_size=3, padding=1,
                               groups=groups, bias=False)
        self.bn2 = nn.BatchNorm2d(planes)
        self.conv3 = nn.Conv2d(planes, planes * 4, kernel_size=1, bias=False)
        self.bn3 = nn.BatchNorm2d(planes * 4)
        self.relu = nn.ReLU(inplace=True)
        self.se_module = SEModule(planes * 4, reduction=reduction)
        self.downsample = downsample
        self.stride = stride
 
 
class SEResNeXtBottleneck(Bottleneck):
    """
    ResNeXt bottleneck type C with a Squeeze-and-Excitation module.
    """
    expansion = 4
 
    def __init__(self, inplanes, planes, groups, reduction, stride=1,
                 downsample=None, base_width=4):
        super(SEResNeXtBottleneck, self).__init__()
        width = math.floor(planes * (base_width / 64)) * groups
        self.conv1 = nn.Conv2d(inplanes, width, kernel_size=1, bias=False,
                               stride=1)
        self.bn1 = nn.BatchNorm2d(width)
        self.conv2 = nn.Conv2d(width, width, kernel_size=3, stride=stride,
                               padding=1, groups=groups, bias=False)
        self.bn2 = nn.BatchNorm2d(width)
        self.conv3 = nn.Conv2d(width, planes * 4, kernel_size=1, bias=False)
        self.bn3 = nn.BatchNorm2d(planes * 4)
        self.relu = nn.ReLU(inplace=True)
        self.se_module = SEModule(planes * 4, reduction=reduction)
        self.downsample = downsample
        self.stride = stride
 
 
bottleneck_dic = {
    'SEBottleneck': SEBottleneck,
    'SEResNetBottleneck': SEResNetBottleneck,
    'SEResNeXtBottleneck': SEResNeXtBottleneck
}
 
@BACKBONES.register_module
class SENet(nn.Module):
 
    def __init__(self, block, layers, groups, reduction, dropout_p=0.2,
                 inplanes=128, input_3x3=True, downsample_kernel_size=3,
                 downsample_padding=1,norm_eval=True, frozen_stages=-1, zero_init_residual=True, num_classes=1000):
        """
        Parameters
        ----------
        block (nn.Module): Bottleneck class.
            - For SENet154: SEBottleneck
            - For SE-ResNet models: SEResNetBottleneck
            - For SE-ResNeXt models:  SEResNeXtBottleneck
        layers (list of ints): Number of residual blocks for 4 layers of the
            network (layer1...layer4).
        groups (int): Number of groups for the 3x3 convolution in each
            bottleneck block.
            - For SENet154: 64
            - For SE-ResNet models: 1
            - For SE-ResNeXt models:  32
        reduction (int): Reduction ratio for Squeeze-and-Excitation modules.
            - For all models: 16
        dropout_p (float or None): Drop probability for the Dropout layer.
            If `None` the Dropout layer is not used.
            - For SENet154: 0.2
            - For SE-ResNet models: None
            - For SE-ResNeXt models: None
        inplanes (int):  Number of input channels for layer1.
            - For SENet154: 128
            - For SE-ResNet models: 64
            - For SE-ResNeXt models: 64
        input_3x3 (bool): If `True`, use three 3x3 convolutions instead of
            a single 7x7 convolution in layer0.
            - For SENet154: True
            - For SE-ResNet models: False
            - For SE-ResNeXt models: False
        downsample_kernel_size (int): Kernel size for downsampling convolutions
            in layer2, layer3 and layer4.
            - For SENet154: 3
            - For SE-ResNet models: 1
            - For SE-ResNeXt models: 1
        downsample_padding (int): Padding for downsampling convolutions in
            layer2, layer3 and layer4.
            - For SENet154: 1
            - For SE-ResNet models: 0
            - For SE-ResNeXt models: 0
        num_classes (int): Number of outputs in `last_linear` layer.
            - For all models: 1000
        """
        super(SENet, self).__init__()
        block = bottleneck_dic[block]
        self.inplanes = inplanes
        self.frozen_stages = frozen_stages
        self.norm_eval = norm_eval
        self.zero_init_residual = zero_init_residual
        if input_3x3:
            layer0_modules = [
                ('conv1', nn.Conv2d(3, 64, 3, stride=2, padding=1,
                                    bias=False)),
                ('bn1', nn.BatchNorm2d(64)),
                ('relu1', nn.ReLU(inplace=True)),
                ('conv2', nn.Conv2d(64, 64, 3, stride=1, padding=1,
                                    bias=False)),
                ('bn2', nn.BatchNorm2d(64)),
                ('relu2', nn.ReLU(inplace=True)),
                ('conv3', nn.Conv2d(64, inplanes, 3, stride=1, padding=1,
                                    bias=False)),
                ('bn3', nn.BatchNorm2d(inplanes)),
                ('relu3', nn.ReLU(inplace=True)),
            ]
        else:
            layer0_modules = [
                ('conv1', nn.Conv2d(3, inplanes, kernel_size=7, stride=2,
                                    padding=3, bias=False)),
                ('bn1', nn.BatchNorm2d(inplanes)),
                ('relu1', nn.ReLU(inplace=True)),
            ]
        # To preserve compatibility with Caffe weights `ceil_mode=True`
        # is used instead of `padding=1`.
        layer0_modules.append(('pool', nn.MaxPool2d(3, stride=2,
                                                    ceil_mode=True)))
        self.layer0 = nn.Sequential(OrderedDict(layer0_modules))
        self.layer1 = self._make_layer(
            block,
            planes=64,
            blocks=layers[0],
            groups=groups,
            reduction=reduction,
            downsample_kernel_size=1,
            downsample_padding=0
        )
        self.layer2 = self._make_layer(
            block,
            planes=128,
            blocks=layers[1],
            stride=2,
            groups=groups,
            reduction=reduction,
            downsample_kernel_size=downsample_kernel_size,
            downsample_padding=downsample_padding
        )
        self.layer3 = self._make_layer(
            block,
            planes=256,
            blocks=layers[2],
            stride=2,
            groups=groups,
            reduction=reduction,
            downsample_kernel_size=downsample_kernel_size,
            downsample_padding=downsample_padding
        )
        self.layer4 = self._make_layer(
            block,
            planes=512,
            blocks=layers[3],
            stride=2,
            groups=groups,
            reduction=reduction,
            downsample_kernel_size=downsample_kernel_size,
            downsample_padding=downsample_padding
        )
        # self.avg_pool = nn.AvgPool2d(7, stride=1)
        # self.dropout = nn.Dropout(dropout_p) if dropout_p is not None else None
        # self.last_linear = nn.Linear(512 * block.expansion, num_classes)
        self._freeze_stages()
 
    def _freeze_stages(self):
        if self.frozen_stages >= 0:
            for m in [self.layer0]:
                m.eval()
                for param in m.parameters():
                    param.requires_grad = False
        for i in range(1, self.frozen_stages + 1):
            m = getattr(self, 'layer{}'.format(i))
            m.eval()
            for param in m.parameters():
                param.requires_grad = False
 
    def init_weights(self, pretrained=None):
        if isinstance(pretrained, str):
            logger = logging.getLogger()
            load_checkpoint(self, pretrained, strict=False, logger=logger)
        elif pretrained is None:
            for m in self.modules():
                if isinstance(m, nn.Conv2d):
                    kaiming_init(m)
                elif isinstance(m, nn.BatchNorm2d):
                    constant_init(m, 1)
                if self.zero_init_residual:
                    for m in self.modules():
                        if isinstance(m, Bottleneck):
                            constant_init(m.bn3, 0)
        else:
            raise TypeError('pretrained must be a str or None')
 
    def _make_layer(self, block, planes, blocks, groups, reduction, stride=1,
                    downsample_kernel_size=1, downsample_padding=0):
        downsample = None
        if stride != 1 or self.inplanes != planes * block.expansion:
            downsample = nn.Sequential(
                nn.Conv2d(self.inplanes, planes * block.expansion,
                          kernel_size=downsample_kernel_size, stride=stride,
                          padding=downsample_padding, bias=False),
                nn.BatchNorm2d(planes * block.expansion),
            )
 
        layers = []
        layers.append(block(self.inplanes, planes, groups, reduction, stride,
                            downsample))
        self.inplanes = planes * block.expansion
        for i in range(1, blocks):
            layers.append(block(self.inplanes, planes, groups, reduction))
 
        return nn.Sequential(*layers)
 
    def features(self, x):
        outputs = []
        x = self.layer0(x)
        x = self.layer1(x)
        outputs.append(x)
        x = self.layer2(x)
        outputs.append(x)
        x = self.layer3(x)
        outputs.append(x)
        x = self.layer4(x)
        outputs.append(x)
        return x, outputs
    '''    
    def logits(self, x):
        x = self.avg_pool(x)
        if self.dropout is not None:
            x = self.dropout(x)
        x = x.view(x.size(0), -1)
        x = self.last_linear(x)
        return x
    '''
    def forward(self, x):
        x, outputs = self.features(x)
        # x = self.logits(x)
        return outputs  # x
 
    def train(self, mode=True):
        super(SENet, self).train(mode)
        self._freeze_stages()
        if mode and self.norm_eval:
            for m in self.modules():
                # trick: eval have effect on BatchNorm only
                if isinstance(m, (nn.BatchNorm2d)):
                    m.eval()
 
def initialize_pretrained_model(model, num_classes, settings):
    assert num_classes == settings['num_classes'], \
        'num_classes should be {}, but is {}'.format(
            settings['num_classes'], num_classes)
    model.load_state_dict(model_zoo.load_url(settings['url']))
    model.input_space = settings['input_space']
    model.input_size = settings['input_size']
    model.input_range = settings['input_range']
    model.mean = settings['mean']
    model.std = settings['std']
 
 
def senet154(num_classes=1000, pretrained='imagenet'):
    model = SENet(SEBottleneck, [3, 8, 36, 3], groups=64, reduction=16,
                  dropout_p=0.2, num_classes=num_classes)
    if pretrained is not None:
        settings = pretrained_settings['senet154'][pretrained]
        initialize_pretrained_model(model, num_classes, settings)
    return model
 
 
def se_resnet50(num_classes=1000, pretrained='imagenet'):
    model = SENet(SEResNetBottleneck, [3, 4, 6, 3], groups=1, reduction=16,
                  dropout_p=None, inplanes=64, input_3x3=False,
                  downsample_kernel_size=1, downsample_padding=0,
                  num_classes=num_classes)
    if pretrained is not None:
        settings = pretrained_settings['se_resnet50'][pretrained]
        initialize_pretrained_model(model, num_classes, settings)
    return model
 
 
def se_resnet101(num_classes=1000, pretrained='imagenet'):
    model = SENet(SEResNetBottleneck, [3, 4, 23, 3], groups=1, reduction=16,
                  dropout_p=None, inplanes=64, input_3x3=False,
                  downsample_kernel_size=1, downsample_padding=0,
                  num_classes=num_classes)
    if pretrained is not None:
        settings = pretrained_settings['se_resnet101'][pretrained]
        initialize_pretrained_model(model, num_classes, settings)
    return model
 
 
def se_resnet152(num_classes=1000, pretrained='imagenet'):
    model = SENet(SEResNetBottleneck, [3, 8, 36, 3], groups=1, reduction=16,
                  dropout_p=None, inplanes=64, input_3x3=False,
                  downsample_kernel_size=1, downsample_padding=0,
                  num_classes=num_classes)
    if pretrained is not None:
        settings = pretrained_settings['se_resnet152'][pretrained]
        initialize_pretrained_model(model, num_classes, settings)
    return model
 
 
def se_resnext50_32x4d(num_classes=1000, pretrained='imagenet'):
    model = SENet(SEResNeXtBottleneck, [3, 4, 6, 3], groups=32, reduction=16,
                  dropout_p=None, inplanes=64, input_3x3=False,
                  downsample_kernel_size=1, downsample_padding=0,
                  num_classes=num_classes)
    if pretrained is not None:
        settings = pretrained_settings['se_resnext50_32x4d'][pretrained]
        initialize_pretrained_model(model, num_classes, settings)
    return model
 
 
def se_resnext101_32x4d(num_classes=1000, pretrained='imagenet'):
    model = SENet(SEResNeXtBottleneck, [3, 4, 23, 3], groups=32, reduction=16,
                  dropout_p=None, inplanes=64, input_3x3=False,
                  downsample_kernel_size=1, downsample_padding=0,
                  num_classes=num_classes)
    if pretrained is not None:
        settings = pretrained_settings['se_resnext101_32x4d'][pretrained]
        initialize_pretrained_model(model, num_classes, settings)
    return model
上面就是senet.py的内容了。

主要是下面这里，自定义好类后，需要进行注册。

@BACKBONES.register_module
class SENet(nn.Module):
还有__init__.py文件里也需要添加相应的模块

from .hrnet import HRNet
from .resnet import ResNet, make_res_layer
from .resnext import ResNeXt
from .ssd_vgg import SSDVGG
from .senet import SENet
 
__all__ = ['SENet','ResNet', 'make_res_layer', 'ResNeXt', 'SSDVGG', 'HRNet']
最后在mmdetection/configs/pascal_voc目录下创建配置文件faster_rcnn_ser50_fpn_1x_voc0712.py。

# model settings
model = dict(
    type='FasterRCNN',
    #pretrained='torchvision://resnet50',
    pretrained='/home/yyf/mmdetection/checkpoints/se_resnet50-ce0d4300.pth',
    backbone=dict(
        type='SENet',
    block='SEResNetBottleneck',
    layers=[3, 4, 6, 3],
    groups=1,
    reduction=16,
    dropout_p=None,
    inplanes=64,
    input_3x3=False,
    downsample_kernel_size=1,
    downsample_padding=0,
    frozen_stages=1
        #depth=50,
        #num_stages=4,
        #out_indices=(0, 1, 2, 3),
        #style='pytorch'
    ),
    neck=dict(
        type='FPN',
        in_channels=[256, 512, 1024, 2048],
        out_channels=256,
        num_outs=5),
    rpn_head=dict(
        type='RPNHead',
        in_channels=256,
        feat_channels=256,
        anchor_scales=[8],
        anchor_ratios=[0.5, 1.0, 2.0],
        anchor_strides=[4, 8, 16, 32, 64],
        target_means=[.0, .0, .0, .0],
        target_stds=[1.0, 1.0, 1.0, 1.0],
        loss_cls=dict(
            type='CrossEntropyLoss', use_sigmoid=True, loss_weight=1.0),
        loss_bbox=dict(type='SmoothL1Loss', beta=1.0 / 9.0, loss_weight=1.0)),
    bbox_roi_extractor=dict(
        type='SingleRoIExtractor',
        roi_layer=dict(type='RoIAlign', out_size=7, sample_num=2),
        out_channels=256,
        featmap_strides=[4, 8, 16, 32]),
    bbox_head=dict(
        type='SharedFCBBoxHead',
        num_fcs=2,
        in_channels=256,
        fc_out_channels=1024,
        roi_feat_size=7,
        num_classes=21,
        target_means=[0., 0., 0., 0.],
        target_stds=[0.1, 0.1, 0.2, 0.2],
        reg_class_agnostic=False,
        loss_cls=dict(
            type='CrossEntropyLoss', use_sigmoid=False, loss_weight=1.0),
        loss_bbox=dict(type='SmoothL1Loss', beta=1.0, loss_weight=1.0)))
# model training and testing settings
train_cfg = dict(
    rpn=dict(
        assigner=dict(
            type='MaxIoUAssigner',
            pos_iou_thr=0.7,
            neg_iou_thr=0.3,
            min_pos_iou=0.3,
            ignore_iof_thr=-1),
        sampler=dict(
            type='RandomSampler',
            num=256,
            pos_fraction=0.5,
            neg_pos_ub=-1,
            add_gt_as_proposals=False),
        allowed_border=0,
        pos_weight=-1,
        debug=False),
    rpn_proposal=dict(
        nms_across_levels=False,
        nms_pre=2000,
        nms_post=2000,
        max_num=2000,
        nms_thr=0.7,
        min_bbox_size=0),
    rcnn=dict(
        assigner=dict(
            type='MaxIoUAssigner',
            pos_iou_thr=0.5,
            neg_iou_thr=0.5,
            min_pos_iou=0.5,
            ignore_iof_thr=-1),
        sampler=dict(
            type='RandomSampler',
            num=512,
            pos_fraction=0.25,
            neg_pos_ub=-1,
            add_gt_as_proposals=True),
        pos_weight=-1,
        debug=False))
test_cfg = dict(
    rpn=dict(
        nms_across_levels=False,
        nms_pre=1000,
        nms_post=1000,
        max_num=1000,
        nms_thr=0.7,
        min_bbox_size=0),
    rcnn=dict(
        score_thr=0.05, nms=dict(type='nms', iou_thr=0.5), max_per_img=100)
    # soft-nms is also supported for rcnn testing
    # e.g., nms=dict(type='soft_nms', iou_thr=0.5, min_score=0.05)
)
# dataset settings
dataset_type = 'VOCDataset'
data_root = 'data/VOCdevkit/'
img_norm_cfg = dict(
    mean=[123.675, 116.28, 103.53], std=[58.395, 57.12, 57.375], to_rgb=True)
train_pipeline = [
    dict(type='LoadImageFromFile'),
    dict(type='LoadAnnotations', with_bbox=True),
    dict(type='Resize', img_scale=(1000, 600), keep_ratio=True),
    dict(type='RandomFlip', flip_ratio=0.5),
    dict(type='Normalize', **img_norm_cfg),
    dict(type='Pad', size_divisor=32),
    dict(type='DefaultFormatBundle'),
    dict(type='Collect', keys=['img', 'gt_bboxes', 'gt_labels']),
]
test_pipeline = [
    dict(type='LoadImageFromFile'),
    dict(
        type='MultiScaleFlipAug',
        img_scale=(1000, 600),
        flip=False,
        transforms=[
            dict(type='Resize', keep_ratio=True),
            dict(type='RandomFlip'),
            dict(type='Normalize', **img_norm_cfg),
            dict(type='Pad', size_divisor=32),
            dict(type='ImageToTensor', keys=['img']),
            dict(type='Collect', keys=['img']),
        ])
]
data = dict(
    imgs_per_gpu=2,
    workers_per_gpu=2,
    train=dict(
        type='RepeatDataset',
        times=3,
        dataset=dict(
            type=dataset_type,
            ann_file=[
                data_root + 'VOC2007/ImageSets/Main/trainval.txt',
                data_root + 'VOC2012/ImageSets/Main/trainval.txt'
            ],
            img_prefix=[data_root + 'VOC2007/', data_root + 'VOC2012/'],
            pipeline=train_pipeline)),
    val=dict(
        type=dataset_type,
        ann_file=data_root + 'VOC2007/ImageSets/Main/test.txt',
        img_prefix=data_root + 'VOC2007/',
        pipeline=test_pipeline),
    test=dict(
        type=dataset_type,
        ann_file=data_root + 'VOC2007/ImageSets/Main/test.txt',
        img_prefix=data_root + 'VOC2007/',
        pipeline=test_pipeline))
# optimizer
optimizer = dict(type='SGD', lr=0.01, momentum=0.9, weight_decay=0.0001)
optimizer_config = dict(grad_clip=dict(max_norm=35, norm_type=2))
# learning policy
lr_config = dict(policy='step', step=[3])  # actual epoch = 3 * 3 = 9
checkpoint_config = dict(interval=1)
# yapf:disable
log_config = dict(
    interval=50,
    hooks=[
        dict(type='TextLoggerHook'),
        # dict(type='TensorboardLoggerHook')
    ])
# yapf:enable
# runtime settings
total_epochs = 4  # actual epoch = 4 * 3 = 12
dist_params = dict(backend='nccl')
log_level = 'INFO'
work_dir = './work_dirs/faster_rcnn_ser50_fpn_1x_voc0712'
load_from = None
resume_from = None
workflow = [('train', 1)]
上面都做好后，还要重新运行python setup.py develop重新编译安装。

完成后就可以愉快的玩耍了。

参考内容：参考
————————————————
版权声明：本文为CSDN博主「Haku_yyf」的原创文章，遵循 CC 4.0 BY-SA 版权协议，转载请附上原文出处链接及本声明。
原文链接：https://blog.csdn.net/haku_yyf/article/details/102504235