solo2试跑报错记录

跑自己训练集前，注意自己的gpu个数和选用的训练文件对应！

1.solov2训练自己数据集

1.1先跑通demo

demo跑通说明环境配置没有问题

出现UserWarning: This overload of nonzero is deprecated，说torch的nonzero要用个as_type的参数（nonzero(*, bool as_tuple)）

跳到相应文件下把inds = inds.nonzero改成inds = inds.nonzero(as_type=false)就行

1.2整理自己的数据集

1.2.1数据集格式

labelme标注后转为coco格式（annotations下放labelme转coco后的合并json文件，train和val里面放的是你用来训练的原图!），路径如下

1.2.2添加上自己的数据集

路径：SOLO-master/mmdet/core/evaluation/class_names.py
路径：SOLO-master/mmdet/datasets/coco.py

这里是一个关于数据集的配置说明，它用了另一种数据集配置的方式，效果一样(后文配置跟随它的)

1.2.3修改配置文件

按照自己的需求选择配置文件，我用的是SOLO-master\configs\solov2\solov2_r50_fpn_8gpu_3x.py，对应的权重文件是SOLO-master\checkpoints\SOLOv2_R50_3x.pth（要自己下载放到对应路径，solo项目里给出了）

100

101

102

103

104

105

106

107

108

109

110

111

112

113

114

115

116

117

118

119

120

121

122

123

124

125

126

127

128

129

130

131

132

133

134

135

136

137

138

139

140

141

142

143

# model settings
model = dict(
    type='SOLOv2',
    pretrained='torchvision://resnet50',
    backbone=dict(
        type='ResNet',
        depth=50,
        num_stages=4,
        out_indices=(0, 1, 2, 3), # C2, C3, C4, C5
        frozen_stages=1,
        style='pytorch'),
    neck=dict(
        type='FPN',
        in_channels=[256, 512, 1024, 2048],
        out_channels=256,
        start_level=0,
        num_outs=5),
    bbox_head=dict(
        type='SOLOv2Head',
        num_classes=2,
        in_channels=256,
        stacked_convs=4,
        seg_feat_channels=512,
        strides=[8, 8, 16, 32, 32],
        scale_ranges=((1, 96), (48, 192), (96, 384), (192, 768), (384, 2048)),
        sigma=0.2,
        num_grids=[40, 36, 24, 16, 12],
        ins_out_channels=256,
        loss_ins=dict(
            type='DiceLoss',
            use_sigmoid=True,
            loss_weight=3.0),
        loss_cate=dict(
            type='FocalLoss',
            use_sigmoid=True,
            gamma=2.0,
            alpha=0.25,
            loss_weight=1.0)),
    mask_feat_head=dict(
            type='MaskFeatHead',
            in_channels=256,
            out_channels=128,
            start_level=0,
            end_level=3,
            num_classes=256,
            norm_cfg=dict(type='GN', num_groups=32, requires_grad=True)),
    )
 
# training and testing settings
train_cfg = dict()
test_cfg = dict(
    nms_pre=500,
    score_thr=0.1,
    mask_thr=0.5,
    update_thr=0.05,
    kernel='gaussian',  # gaussian/linear
    sigma=2.0,
    max_per_img=100)
 
# dataset settings
dataset_type = 'my_data'
data_root = 'D:/view/things/SOLO-master/mmdet/datasets/coco/'
# data_root = 'mmdet/datasets/coco/'# 自己的数据集路径
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, with_mask=True),
    dict(type='Resize',
         img_scale=[(548,192),],# 自己要训练的图片大小
         # img_scale=[(1333, 800), (1333, 768), (1333, 736),
         #            (1333, 704), (1333, 672), (1333, 640)],
         multiscale_mode='value',
         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', 'gt_masks']),
]
test_pipeline = [
    dict(type='LoadImageFromFile'),
    dict(
        type='MultiScaleFlipAug',
        img_scale = (548,192),
        # img_scale=(1333, 800),
        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=dataset_type,
        ann_file=data_root + 'annotations/train.json',
        img_prefix=data_root + 'train/',
        pipeline=train_pipeline),
    val=dict(
        type=dataset_type,
        ann_file=data_root + 'annotations/val.json',
        img_prefix=data_root + 'val/',
        pipeline=test_pipeline),
    test=dict(
        type=dataset_type,
        ann_file=data_root + 'annotations/val.json',
        img_prefix=data_root + 'val/',
        pipeline=test_pipeline))
 
# optimizer
optimizer = dict(type='SGD', lr=0.00125, 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',
    warmup='linear',
    warmup_iters=500,
    warmup_ratio=0.01,
    step=[27, 33])
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 = 36
device_ids = range(8)
dist_params = dict(backend='nccl')
log_level = 'INFO'
work_dir = './work_dirs/solov2_release_r50_fpn_8gpu_3x'
load_from = None
resume_from = None
workflow = [('train', 1)]