mmdetection的安裝與個人數據的訓練

第一部分：mmdetection的安裝

1. 首先查看一下gcc的版本，有的服務器gcc版本較低，導致後續安裝出現問題（大部分無此問題）

gcc --version

官網v1.0rc0 (27/07/2019)版本要求GCC: 4.9/5.3/5.4/7.3

2.創建基本運行環境

注：這裏使用anaconda管理項目環境，不建議使用Anaconda3-2019.03-Linux-x86_64.sh版本的anaconda,個人實驗證明這個版本的anaconda在進行conda安裝某些包時存在bug,需要更新conda.使用 Anaconda3-2019.07-Linux-x86_64.sh版本的anaconda時一切順利。(mmdetection安裝時可以不注意這個，不影響)

conda create -n open-mmlab python=3.7 -y
conda activate open-mmlab
conda install cython

安裝pytorch官方鏈接:https://pytorch.org/

有的時候因網速過慢無法使用官方方式直接安裝，這裏提供了一個使用清華鏡像的安裝方式：

### 設置清華源鏡像
 conda config --add channels https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/free/
 conda config --add channels https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main/
 conda config --set show_channel_urls yes
### 設置pytorch鏡像
 conda config --add channels https://mirrors.tuna.tsinghua.edu.cn/anaconda/cloud/pytorch/
 conda config --add channels https://mirrors.tuna.tsinghua.edu.cn/anaconda/cloud/peterjc123/

添加鏡像後將官方安裝的命令-c pytorch去掉，例如：

conda install pytorch torchvision cudatoolkit=9.0

3.安裝mmdetection

安裝mmcv:

git clone https://github.com/open-mmlab/mmcv.git
cd mmcv
pip install -e .

安裝mmdetection:

git clone https://github.com/open-mmlab/mmdetection.git
cd mmdetection
python setup.py develop # or "pip install -v -e ."

安裝mmdetection時若出現警告：cc1plus: warning: command line option ‘-Wstrict-prototypes’ is valid for C/ObjC but not for C++，按照https://note.qidong.name/2018/03/setup-warning-strict-prototypes/ 方式可以解決，當然你也可以忽略。

安裝mmdetection時若出現此錯誤：libSM.so.6: cannot open shared，按照https://blog.csdn.net/fanzonghao/article/details/85112476。方式可以解決，注意把cuda-8改成你的cuda版本。

apt-get install tk-dev python-tk
sudo ldconfig /usr/local/cuda-9.0/lib64

第二部分：使用mmdetection訓練自己的數據

1.mmdetection提控了一個CustomDataset數據格式用於訓練自己的數據，首先附上自己csv格式數據轉CustomDataset格式的代碼：

'''
CSV annotation file transferred to mmdetection annotation file

@author: hobart
'''

# -*- coding: UTF-8 -*-
import os
import cv2
import numpy as np
import pandas as pd


def read_csv(ROOT_DIR, FILE_NAME):
    ''' Read the CSV file , put the image csv in data.

    Args:
        ROOT_DIR (str): csv file root_path
        FILE_NAME (str): csv file name

    Return (dict):
        a dictionary contains ...
        {"name1" : [[[x1],[y1],[x2],[y2]]],
        "name2" : [[[x1],[y1],[x2],[y2]],[[x1],[y1],[x2],[y2]]]
        }
    '''
    anns_train = pd.read_csv(os.path.join(ROOT_DIR, FILE_NAME))
    data = {}
    for _, row in anns_train.iterrows():
        if row["name"] in data and row["label"] == "1":
            data[row["name"]].append([row["x1"], row["y1"], row["x2"], row["y2"]])
        else:
            if row["label"] == "1":
                data[row["name"]] = [[row["x1"], row["y1"], row["x2"], row["y2"]]]
            else:
                data[row["name"]] = []

    return data


def generate_dataset(ROOT_DIR, annotations):
    '''

    Args (type):
        ROOT_DIR (str)
        annotations (dictionary):
        {"name1" : [[[x1],[y1],[x2],[y2]]],
        "name2" : [[[x1],[y1],[x2],[y2]],[[x1],[y1],[x2],[y2]]]
        }
    Return (type):
        dataset (list):
        [
            {
                'filename': 'a.jpg',
                'width': 1280,
                'height': 720,
                'ann': {
                    'bboxes': <np.ndarray> (n, 4),
                    'labels': <np.ndarray> (n, ),
                    'bboxes_ignore': <np.ndarray> (k, 4),
                    'labels_ignore': <np.ndarray> (k, ) (optional field)
                }
            },
            ...
        ]
    '''
    dataset = []
    for record in annotations:
        image_path = os.path.join(ROOT_DIR, record)
        image_name = os.path.basename(image_path)
        img = cv2.imread(image_path)
        height, width = img.shape[:2]

        if len(annotations[record]) > 0:
            labels = [1] * len(annotations[record])
            labels = np.array(labels)
            bboxes = annotations[record]
            bboxes = np.array(bboxes)
        else:
            bboxes = np.zeros((0, 4))
            labels = np.zeros((0,))
        bboxes_ignore = np.zeros((0, 4))
        labels_ignore = np.zeros((0,))
        annotation = {
            'filename': image_name,
            'width': height,
            'height': width,
            'ann': {
                'bboxes': bboxes.astype(np.float32),
                'labels': labels.astype(np.int64),
                'bboxes_ignore': bboxes_ignore.astype(np.float32),
                'labels_ignore': labels_ignore.astype(np.int64)
            }
        }
        dataset.append(annotation)

    return dataset


def save_pkl(dataset, ROOT_DIR, OUTPUT_DIR, OUTPUT_FILENAME):

    folder = os.path.join(ROOT_DIR, OUTPUT_DIR)
    if not os.path.isdir(folder):
        os.makedirs(folder)
    out_file = os.path.join(ROOT_DIR, OUTPUT_DIR, OUTPUT_FILENAME)

    import mmcv
    mmcv.dump(dataset, out_file)


def main():
    # parameter definition
    ROOT_DIR = "csv/" # The root of all file
    FILE_NAME = "test.csv"
    OUTPUT_DIR = "pkl/"
    OUTPUT_FILENAME = "val.pkl"

    annotations = read_csv(ROOT_DIR, FILE_NAME)
    dataset = generate_dataset(ROOT_DIR, annotations)
    save_pkl(dataset, ROOT_DIR, OUTPUT_DIR, OUTPUT_FILENAME)


if __name__ == "__main__":
    main()
    print("do")

2.修改configs文件下的配置文件：

這裏以retinanet_r50_fpn_1x.py爲例。主要修改dataset_type以及所用的訓練集驗證集測試集的標註數據以及圖像的地址，還有算法分類個數（目標類別書+背景類）。附上自己的配置文件

# model settings
model = dict(
    type='RetinaNet',
    pretrained='modelzoo://resnet50',
    backbone=dict(
        type='ResNet',
        depth=50,
        num_stages=4,
        out_indices=(0, 1, 2, 3),
        frozen_stages=1,
        style='pytorch'),
    neck=dict(
        type='FPN',
        in_channels=[256, 512, 1024, 2048],
        out_channels=256,
        start_level=1,
        add_extra_convs=True,
        num_outs=5),
    bbox_head=dict(
        type='RetinaHead',
        num_classes=2,
        in_channels=256,
        stacked_convs=4,
        feat_channels=256,
        octave_base_scale=4,
        scales_per_octave=3,
        anchor_ratios=[0.5, 1.0, 2.0],
        anchor_strides=[8, 16, 32, 64, 128],
        target_means=[.0, .0, .0, .0],
        target_stds=[1.0, 1.0, 1.0, 1.0],
        loss_cls=dict(
            type='FocalLoss',
            use_sigmoid=True,
            gamma=2.0,
            alpha=0.25,
            loss_weight=1.0),
        loss_bbox=dict(type='SmoothL1Loss', beta=0.11, 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 = "CustomDataset"#'CocoDataset'
data_root = 'data'
img_norm_cfg = dict(
    mean=[123.675, 116.28, 103.53], std=[58.395, 57.12, 57.375], to_rgb=True)
data = dict(
    imgs_per_gpu=2,
    workers_per_gpu=2,
    train=dict(
        type=dataset_type,
        ann_file=data_root + 'pkl/txt.pkl',
        img_prefix=data_root + 'picture/train/',
        img_scale=(1333, 800),
        img_norm_cfg=img_norm_cfg,
        size_divisor=32,
        flip_ratio=0.5,
        with_mask=False,
        with_crowd=False,
        with_label=True),
    val=dict(
        type=dataset_type,
        ann_file=data_root + 'pkl/val.pkl',
        img_prefix=data_root + 'picture/val/',
        img_scale=(1333, 800),
        img_norm_cfg=img_norm_cfg,
        size_divisor=32,
        flip_ratio=0,
        with_mask=False,
        with_crowd=False,
        with_label=True),
    test=dict(
        type=dataset_type,
        ann_file=data_root + 'pkl/test.pkl',
        img_prefix=data_root + 'picture/test/',
        img_scale=(1333, 800),
        img_norm_cfg=img_norm_cfg,
        size_divisor=32,
        flip_ratio=0,
        with_mask=False,
        with_crowd=False,
        with_label=False,
        test_mode=True))
# 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',
    warmup='linear',
    warmup_iters=500,
    warmup_ratio=1.0 / 3,
    step=[8, 11])
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 = 12
device_ids = range(8)
dist_params = dict(backend='nccl')
log_level = 'INFO'
work_dir = './work_dirs/retinanet_r50_fpn_1x'
load_from = None
resume_from = None
workflow = [('train', 1)]

注意，如果實在docker上進行訓練，若出現以下錯誤：ERROR: Unexpected bus error encountered in worker. This might be caused by insufficient shared memory (shm)

解決：在服務器上的docker中運行訓練代碼時，batch size設置得過大，shared memory不夠（因爲docker限制了shm）.解決方法是，將Dataloader的num_workers設置爲0.

3.訓練

python tools/train.py ${CONFIG_FILE}