PolarMask代码相关

0、环境说明

• cd PolarMask/
• conda create --name polarmask python=3.7
• conda activate polarmask
• pip install torch==1.2.0
• pip install torchvision==0.4.0
• pip install Cython
• python setup.py develop
• pip install mmcv
• pip install prompt-toolkit==2.0.10
• pip install Polygon3

机器环境：nvidia-418+cuda 10.0+cudnn 7.6.3

可参考Ubuntu 安装显卡驱动 +cuda +cudnn

1、config设置

配置文件的设置：

# model settings
model = dict(
    type='PolarMask',
    pretrained='open-mmlab://resnet50_caffe',
    backbone=dict(
        type='ResNet',
        depth=50,
        num_stages=4,
        out_indices=(0, 1, 2, 3),
        frozen_stages=1,
        norm_cfg=dict(type='BN', requires_grad=False),
        style='caffe'),
    neck=dict(
        type='FPN',
        in_channels=[256, 512, 1024, 2048],
        out_channels=256,
        start_level=1,
        add_extra_convs=True,
        extra_convs_on_inputs=False,  # use P5
        num_outs=5,
        relu_before_extra_convs=True),
    bbox_head=dict(
        type='PolarMask_Head',
        num_classes=81,
        in_channels=256,
        stacked_convs=4,
        feat_channels=256,
        strides=[8, 16, 32, 64, 128],
        loss_cls=dict(
            type='FocalLoss',
            use_sigmoid=True,
            gamma=2.0,
            alpha=0.25,
            loss_weight=1.0),
        loss_bbox=dict(type='IoULoss', loss_weight=1.0),
        loss_centerness=dict(
            type='CrossEntropyLoss', use_sigmoid=True, loss_weight=1.0)))
# training and testing settings
train_cfg = dict(
    assigner=dict(
        type='MaxIoUAssigner',
        pos_iou_thr=0.5,
        neg_iou_thr=0.4,
        min_pos_iou=0,
        ignore_iof_thr=-1),
    allowed_border=-1,
    pos_weight=-1,
    debug=False)
test_cfg = dict(
    nms_pre=1000,
    min_bbox_size=0,
    score_thr=0.05,
    nms=dict(type='nms', iou_thr=0.5),
    max_per_img=100)
# dataset settings
dataset_type = 'Coco_Seg_Dataset'
data_root = '/media/dell/Elements/dataset/MSCOCO/2017/' # 数据路径
img_norm_cfg = dict(
    mean=[102.9801, 115.9465, 122.7717], std=[1.0, 1.0, 1.0], to_rgb=False)
data = dict(
    imgs_per_gpu=2,
    workers_per_gpu=5,
    train=dict(
        type=dataset_type,
        ann_file=data_root + 'annotations/instances_train2017.json',
        img_prefix=data_root + 'train2017/',
        img_scale=(1280, 768),
        img_norm_cfg=img_norm_cfg,
        # size_divisor=0,
        flip_ratio=0.5,
        with_mask=True,
        with_crowd=False,
        with_label=True,
        resize_keep_ratio=False),
    val=dict(
        type=dataset_type,
        ann_file=data_root + 'annotations/instances_val2017.json',
        img_prefix=data_root + 'val2017/',
        img_scale=(1280, 768),
        img_norm_cfg=img_norm_cfg,
        # size_divisor=0,
        flip_ratio=0,
        with_mask=False,
        with_crowd=False,
        with_label=True,
        resize_keep_ratio=False),
    test=dict(
        type=dataset_type,
        ann_file=data_root + 'annotations/instances_val2017.json',
        img_prefix=data_root + 'val2017/',
        img_scale=(1280, 768),
        img_norm_cfg=img_norm_cfg,
        size_divisor=32,
        flip_ratio=0,
        with_mask=False,
        with_crowd=False,
        with_label=False,
        resize_keep_ratio=False,
        test_mode=True))
# optimizer
lr_ratio = 1

optimizer = dict(
    type='SGD',
    lr=0.01 / 8 * lr_ratio,
    momentum=0.9,
    weight_decay=0.0001,
    paramwise_options=dict(bias_lr_mult=2., bias_decay_mult=0.))
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=1.0 / 3 / lr_ratio,
    step=[8, 11])
checkpoint_config = dict(interval=1)
# yapf:disable
log_config = dict(
    interval=10,
    hooks=[
        dict(type='TextLoggerHook'),
        # dict(type='TensorboardLoggerHook')
    ])
# yapf:enable
# runtime settings
total_epochs = 12
device_ids = range(2) # 由几张卡决定
dist_params = dict(backend='nccl')
log_level = 'INFO'
work_dir = './work_dirs/trash'
load_from = None
resume_from = None
workflow = [('train', 1)]

2、训练语句

双GPU

python ./tools/train.py configs/polarmask/4gpu/polar_768_1x_r50.py --gpus 2 --launcher pytorch --work_dir ./tools/work_dirs/polar_768_1x_r50_4gpu --resume_from ./tools/work_dirs/polar_768_1x_r50_4gpu/epoch_2.pth

单GPU

python ./tools/train.py configs/polarmask/4gpu/polar_768_1x_r50.py --gpus 1 --launcher pytorch --work_dir ./tools/work_dirs/polar_768_1x_r50_4gpu

看出差别了吗？

3、测试语句

python ./tools/test.py configs/polarmask/4gpu/polar_768_1x_r50.py  ./tools/work_dirs/polar_768_1x_r50_4gpu/latest.pth --launcher pytorch --out work_dirs/trash/res.pkl --eval segm

4、出现问题

1、ModuleNotFoundError: No module named 'mmcv.cnn.weight_init'
解决：pip install mmcv==0.5.9

2、ImportError: cannot import name 'AsyncGenerator'
解决：pip install prompt-toolkit==2.0.10

3、ImportError: cannot import name 'get_dist_info'
解决：
将/PolarMask/mmdet/datasets/loader/sampler.py的第六行的from mmcv.runner.utils import get_dist_infoto改成from mmcv.runner import get_dist_info

from mmcv.runner import get_dist_info

4、anaconda3/lib/python3.6/sitepackages/torch/lib/libtorch_python.so: undefined symbol: PySlice_Unpack
解决：将python3.6.0升级更高（3.6.2及以上）

5、KeyError: 'RANK'
解决：train.py增加一个args.launcher = 'none'

# init distributed env first, since logger depends on the dist info.
    args.launcher = 'none'
    if args.launcher == 'none':
        distributed = False
    else:
        distributed = True
        init_dist(args.launcher, **cfg.dist_params)

5、可视化

找到demo文件夹下的visualize.ipynb，或者利用下面代码：

from mmdet.apis import init_detector, inference_detector, show_result_pyplot
import mmcv
import os

config_file = '../configs/polarmask/4gpu/polar_768_1x_r50.py'
# 模型路径
checkpoint_file = '../tools/tools/work_dirs/polar_768_1x_r50_4gpu/epoch_7.pth'
model = init_detector(config_file, checkpoint_file, device='cuda:0')

# 可视化COCO数据集的test数据集的结果
# img_path = '/media/dell/Elements/dataset/MSCOCO/2017/test2017'
# img_names = os.listdir(img_path)
# for name in img_names:
#     img = os.path.join(img_path, name)
#     result = inference_detector(model, img)
#     show_result_pyplot(img, result, model.CLASSES, score_thr=-1)

# 随便一张图
img = '1.jpg'
result = inference_detector(model, img)
show_result_pyplot(img, result, model.CLASSES, score_thr=0.3)

6、代码结构

网络结构：backbone+neck+box_head，代码见mmdet/models/detectors/polarmask.py，列出主要代码

class PolarMask(SingleStageDetector):
    def __init__(self,
                 backbone,
                 neck,
                 bbox_head,
                 train_cfg=None,
                 test_cfg=None,
                 pretrained=None):
        super(PolarMask, self).__init__(backbone, neck, bbox_head, train_cfg,
                                   test_cfg, pretrained)

必选参数为backbone、neck、bbox_head，这里的参数均可以在configs/polarmask/4gpu中的文件中找到，通过tools/train/py将参数传入。

bbox_head文件：实现掩膜生成的主要文件，位置在mmdet/models/anchor_heads/polarmask_head.py。查看__init__代码,相关参数见注释：

class PolarMask_Head(nn.Module):
    def __init__(self,
                 num_classes, # 数据集类别数
                 in_channels, # 经过neck网络的特征输出层数
                 feat_channels=256, # head输出的特征层数
                 stacked_convs=4, # head包含的基础网络个数
                 strides=(4, 8, 16, 32, 64), 
                 regress_ranges=((-1, 64), (64, 128), (128, 256), (256, 512),
                                 (512, INF)),
                 use_dcn=False,
                 mask_nms=False, 
                 # 以下定义loss，可以自动完成backforward
                 loss_cls=dict(
                     type='FocalLoss',
                     use_sigmoid=True,
                     gamma=2.0,
                     alpha=0.25,
                     loss_weight=1.0),
                 loss_bbox=dict(type='IoULoss', loss_weight=1.0),
                 loss_mask=dict(type='MaskIOULoss'),
                 loss_centerness=dict(
                     type='CrossEntropyLoss',
                  use_sigmoid=True,
               loss_weight=1.0),
        conv_cfg=None,
       norm_cfg=dict(type='GN', num_groups=32, requires_grad=True)):
       super(PolarMask_Head, self).__init__()

PolarMask_head的函数主要有以下：

除了上面的函数外，还有一个distance2mask，就是根据论文中下式转成mask：

数据集：mmdet/datasets/coco_seg.py
主要的函数为prepare_train_img函数，整体结构为：

Loss以及optimizer：在第一步中，通过增加dict完成。

训练网络：利用tools/train.py将参数传入，逐步调用即可。

验证网络：利用tools/test.py将参数传入，逐步调用即可。