修改配置文件并训练

天天学Java,好久没跑过实验了，记录下，防止日后踩坑

1. 将mmdet.models.utils.se_layer.SELayer注册为plugin
   注意修改下参数channels为in_channels，否则解析参数会对不上，出现unexpected args channels的参数解析错误

# Copyright (c) OpenMMLab. All rights reserved.
import mmcv
import torch.nn as nn
from mmcv.cnn import ConvModule
from mmcv.runner import BaseModule
from mmcv.cnn import PLUGIN_LAYERS


@PLUGIN_LAYERS.register_module()
class SELayer(BaseModule):
	def __init__(self,
	                 in_channels,  # 对应resnet.py中的参数设置修改下
	                 ratio=16,
	                 conv_cfg=None,
	                 act_cfg=(dict(type='ReLU'), dict(type='Sigmoid')),
	                 init_cfg=None):
	        super(SELayer, self).__init__(init_cfg)

2. 修改SSD配置文件，并训练
   此处我是参照ssdlite_mobilenetv2_scratch_600e_coco.py进行修改的

_base_ = [
    '../_base_/datasets/coco_detection.py', '../_base_/default_runtime.py'
]

model = dict(
    type='SingleStageDetector',
    backbone=dict(
        type='ResNet',
        depth=50,
        num_stages=4,
        # 哪几个stage的结果作为输出，从0开始
        out_indices=(1, 2, 3),
        frozen_stages=1,
        norm_cfg=dict(type='BN', requires_grad=True),
        norm_eval=True,
        style='pytorch',
        # 按照自己需求配置plugin
        plugins=[dict(cfg=dict(type='SELayer'),stages=(True, True, True, True),position='after_conv3'),],
        # 使用预训练权重
        init_cfg=dict(type='Pretrained', checkpoint='torchvision://resnet50')),
        # norm_cfg=dict(type='BN', eps=0.001, momentum=0.03),
        # init_cfg=dict(type='TruncNormal', layer='Conv2d', std=0.03)),
    neck=dict(
        type='SSDNeck',
        # 对应backbone的输出，默认resnet最后三个stage out_channels=[512, 1024, 2048]
        in_channels=(512, 1024, 2048),
        # 对应SSD的6层输出,前三层为什么这么设置，见补充说明SSD-NECK结构
        out_channels=(512, 1024, 2048, 256, 256, 256),
        # 小坑：这里设置注意下，如果没设置对可能出现得到的先验框数不匹配(tensor.shape不匹配)
        level_strides=(2, 2, 2),
        level_paddings=(1, 1, 0),
        l2_norm_scale=None,
        use_depthwise=True,
        norm_cfg=dict(type='BN', eps=0.001, momentum=0.03),
        act_cfg=dict(type='ReLU6'),
        init_cfg=dict(type='TruncNormal', layer='Conv2d', std=0.03)),
    bbox_head=dict(
        type='SSDHead',
        in_channels=(512, 1024, 2048, 256, 256, 256),
        # 按照自己的数据集配置
        num_classes=1,
        use_depthwise=True,
        norm_cfg=dict(type='BN', eps=0.001, momentum=0.03),
        act_cfg=dict(type='ReLU6'),
        init_cfg=dict(type='Normal', layer='Conv2d', std=0.001),

        # 默认的SSD300 检测框设置
        anchor_generator=dict(
            type='SSDAnchorGenerator',
            scale_major=False,
            strides=[8, 16, 32, 64, 100, 300],
            ratios=[[2], [2, 3], [2, 3], [2, 3], [2], [2]]),
        bbox_coder=dict(
            type='DeltaXYWHBBoxCoder',
            target_means=[.0, .0, .0, .0],
            target_stds=[0.1, 0.1, 0.2, 0.2])),
    # model training and testing settings
    train_cfg=dict(
        assigner=dict(
            type='MaxIoUAssigner',
            pos_iou_thr=0.5,
            neg_iou_thr=0.5,
            min_pos_iou=0.,
            ignore_iof_thr=-1,
            gt_max_assign_all=False),
        smoothl1_beta=1.,
        allowed_border=-1,
        pos_weight=-1,
        neg_pos_ratio=3,
        debug=False),
    test_cfg=dict(
        nms_pre=1000,
        nms=dict(type='nms', iou_threshold=0.45),
        min_bbox_size=0,
        score_thr=0.02,
        max_per_img=200))
cudnn_benchmark = True

# 以下和ssdlite_mobilenetv2_scratch_600e_coco.py设置差不多
dataset_type = 'CocoDataset'
data_root = '/home/COCO/'
img_norm_cfg = dict(mean=[123.675, 116.28, 103.53], std=[1, 1, 1], to_rgb=True)
train_pipeline = [
    dict(type='LoadImageFromFile', to_float32=True),
    dict(type='LoadAnnotations', with_bbox=True),
    dict(
        type='PhotoMetricDistortion',
        brightness_delta=32,
        contrast_range=(0.5, 1.5),
        saturation_range=(0.5, 1.5),
        hue_delta=18),
    dict(
        type='Expand',
        mean=img_norm_cfg['mean'],
        to_rgb=img_norm_cfg['to_rgb'],
        ratio_range=(1, 4)),
    dict(
        type='MinIoURandomCrop',
        min_ious=(0.1, 0.3, 0.5, 0.7, 0.9),
        min_crop_size=0.3),
    dict(type='Resize', img_scale=(300, 300), keep_ratio=False),
    dict(type='Normalize', **img_norm_cfg),
    dict(type='RandomFlip', flip_ratio=0.5),
    dict(type='DefaultFormatBundle'),
    dict(type='Collect', keys=['img', 'gt_bboxes', 'gt_labels']),
]
test_pipeline = [
    dict(type='LoadImageFromFile'),
    dict(
        type='MultiScaleFlipAug',
        img_scale=(300, 300),
        flip=False,
        transforms=[
            dict(type='Resize', keep_ratio=False),
            dict(type='Normalize', **img_norm_cfg),
            dict(type='ImageToTensor', keys=['img']),
            dict(type='Collect', keys=['img']),
        ])
]
data = dict(
    samples_per_gpu=24,
    workers_per_gpu=4,
    train=dict(
        _delete_=True,
        type='RepeatDataset',
        times=5,
        dataset=dict(
            type=dataset_type,
            ann_file=data_root + 'annotations/train2017.json',
            img_prefix=data_root + 'images/',
            pipeline=train_pipeline)),
    val=dict(pipeline=test_pipeline),
    test=dict(pipeline=test_pipeline))
# optimizer这里根据自己的实际情况设置，可以参照mmd官方文档学习率设置规则
optimizer = dict(type='SGD', lr=0.002, momentum=0.9, weight_decay=5.0e-4)
optimizer_config = dict(grad_clip=None)

# learning policy
lr_config = dict(
    policy='CosineAnnealing',
    warmup='linear',
    warmup_iters=500,
    warmup_ratio=0.001,
    min_lr=0)
runner = dict(type='EpochBasedRunner', max_epochs=60)
# Avoid evaluation and saving weights too frequently
evaluation = dict(interval=5, metric='bbox')
checkpoint_config = dict(interval=5)
custom_hooks = [
    dict(type='NumClassCheckHook'),
    dict(type='CheckInvalidLossHook', interval=50, priority='VERY_LOW')
]

3. 实验收获与注意事项
   a. 实际训练中从0开始训练，加入SE好像反而map轻微掉点了，根据官方issue，似乎backbone需要通过预训练权重初始化才能有相对好的效果，不明白为什么
   b. 在采用cos学习率学习之后还要再用小学习率finetune下

补充说明

1. mmd resnet中plugin可以设置的参数
   参照ResNet源码的注释

    def make_stage_plugins(self, plugins, stage_idx):
        """
        Examples:
            >>> plugins=[
            ...     dict(cfg=dict(type='xxx', arg1='xxx'),
            ...          stages=(False, True, True, True),
            ...          position='after_conv2'),
            ...     dict(cfg=dict(type='yyy'),
            ...          stages=(True, True, True, True),
            ...          position='after_conv3'),
            ...     dict(cfg=dict(type='zzz', postfix='1'),
            ...          stages=(True, True, True, True),
            ...          position='after_conv3'),
            ...     dict(cfg=dict(type='zzz', postfix='2'),
            ...          stages=(True, True, True, True),
            ...          position='after_conv3')
            ... ]
            >>> self = ResNet(depth=18)
            >>> stage_plugins = self.make_stage_plugins(plugins, 0)
            >>> assert len(stage_plugins) == 3
       """

plugins的配置的字典分别需要配置:
1）cfg, 例如以下我自己编写的SELayer按以下配置会解析参数in_channels, map_size,在第一个stage的conv3后添加

dict(cfg=dict(type='SELayer', map_size='150'),stages=(True, False, False, False),position='after_conv3')

2）stages(分别作用ResNet的哪几个stage)，
3）position(作用于conv1/2/3之后)

2. ssd_neck
   参照MMD官方的知乎文章，SSD的项目代码重构了，通过SSD-NECK取代了之前一直在用的ssd_vgg，详细结构见：MMD官方文章传送门