mmdetection和mmsegmentation无法加载预训练模型的问题

mmdetection和mmsegmentation无法加载预训练模型的问题

Openmmlab无法加载预训练模型的问题

这两天在调试​​mmsegmentation​​​和​​mmdetection​​，可能是因为自己的原因，预训练模型死活加载不了预训练的模型，无法正常的索引到预训练模型的地址，最后通过降低版本的方式成功地加载了预训练模型并跑了起来，具体的流程如下：

解决过程

安装pytorch和torchvision我是30系列的显卡，所以需要的cuda版本需要是11以上。

conda install pytorch==1.9.0 torchvision==0.10.0 cudatoolkit=11.1

安装mmcv-full

pip install mmcv-full==1.3.10 -f clone install -v --disable-pip-version-check --no-cache-dir --global-option="--cpp_ext" --global-option="--cuda_ext"

安装mmdetection，我使用的是​​SwinTransformer/Swin-Transformer-Object-Detection​​这个版本的

git clone install -v -e .

安装mmpycocotools

pip uninstall pycocotoolspip install

测试的代码如下，当时我主要是想测试一下mmdetection在dota数据集上的表现：

from mmcv import Configfrom mmdet.datasets import build_datasetfrom mmdet.models import build_detectorfrom mmdet.apis import train_detectorfrom mmdet.apis import set_random_seedimport os.path as ospimport mmcvimport numpy as npfrom mmdet.datasets.builder import DATASETSfrom mmdet.datasets.custom import CustomDatasetimport warnings# warnings.filterwarnings('ignore')# 目前的解决方案，要不重写一个dataset的类，要不统一都弄成coco的形式。cfg = Config.fromfile('./configs/fcos/fcos_r50_caffe_fpn_gn-head_mstrain_640-800_2x_coco.py')# todo 1. 定义数据集# 目前这个数据有大问题，咱首先得处理coco格式，然后得写个带可视化得api方便查看，奶奶得。cfg.dataset_type = 'CocoDataset' # todo 数据集格式cfg.classes = ('plane', 'baseball-diamond', 'bridge', 'ground-track-field', 'small-vehicle', 'large-vehicle', 'ship', 'tennis-court', 'basketball-court', 'storage-tank', 'soccer-ball-field', 'roundabout', 'harbor', 'swimming-pool', 'helicopter', 'container-crane',) # todo 类名data_images = '/home/lyc/data/scm/remote/dota1.5hbb/PNGImages/images/' # todo 数据集根路径cfg.data.train.ann_file = '/home/lyc/data/scm/remote/dota1.5hbb/dota_train.json' # todo json文件路径cfg.data.val.ann_file = '/home/lyc/data/scm/remote/dota1.5hbb/dota_val.json' # todo 验证集json文件路径cfg.data.test.ann_file = '/home/lyc/data/scm/remote/dota1.5hbb/dota_val.json' # todo 测试集json文件路径cfg.data.train.type = cfg.dataset_typecfg.data.val.type = cfg.dataset_typecfg.data.test.type = cfg.dataset_typecfg.data.train.classes = cfg.classescfg.data.val.classes = cfg.classescfg.data.test.classes = cfg.classescfg.data.train.img_prefix = data_images #cfg.data.val.img_prefix = data_imagescfg.data.test.img_prefix = data_imagescfg.data.samples_per_gpu = 4 # Batch size of a single GPU used in testing 默认是8x2cfg.data.workers_per_gpu = 1 # Worker to pre-fetch data for each single GPU# *************** transform **************train_pipeline = [ dict(type='LoadImageFromFile'), dict(type='LoadAnnotations', with_bbox=True), dict( type='Resize', img_scale=(1024, 1024), # multiscale_mode='value', keep_ratio=True), dict(type='RandomFlip', flip_ratio=0.5), dict( type='Normalize', mean=[102.9801, 115.9465, 122.7717], std=[1.0, 1.0, 1.0], to_rgb=False), 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=(1024, 1024), flip=False, transforms=[ dict(type='Resize', keep_ratio=True), dict(type='RandomFlip'), dict( type='Normalize', mean=[102.9801, 115.9465, 122.7717], std=[1.0, 1.0, 1.0], to_rgb=False), dict(type='Pad', size_divisor=32), dict(type='ImageToTensor', keys=['img']), dict(type='Collect', keys=['img']) ])]cfg.data.train.pipeline = cfg.train_pipelinecfg.data.val.pipeline = cfg.test_pipelinecfg.data.test.pipeline = cfg.test_pipeline# modify num classes of the model in box headcfg.model.bbox_head.num_classes = len(cfg.classes)#cfg.load_from = '../checkpoints/resnet50_caffe-788b5fa3.pth'cfg.work_dir = '../tutorial_exps/2-dota_fcos_1024_backbone'# The original learning rate (LR) is set for 8-GPU training.# We divide it by 8 since we only use one GPU.cfg.optimizer.lr = 0.02 / 8cfg.lr_config.warmup = Nonecfg.log_config.interval = 10# Change the evaluation metric since we use customized dataset.# cfg.evaluation.metric = 'mAP'cfg.evaluation.metric = 'bbox'cfg.evaluation.save_best = 'bbox_mAP'# We can set the evaluation interval to reduce the evaluation timescfg.evaluation.interval = 1# We can set the checkpoint saving interval to reduce the storage costcfg.checkpoint_config.interval = 12# Set seed thus the results are more reproduciblecfg.seed = 0set_random_seed(0, deterministic=False)# cfg.gpu_ids = range(1)cfg.gpu_ids = (0,)# We can initialize the logger for training and have a look# at the final config used for trainingprint(f'Config:\n{cfg.pretty_text}')# 保存模型的各种参数（一定要记得嗷）cfg.dump(F'{cfg.work_dir}/customformat_fcos.py')# 训练主要进程# Build datasetdatasets = [build_dataset(cfg.data.train)]print(cfg.data.train)print(datasets[0])print(datasets[0].CLASSES)# Build the detectormodel = build_detector( cfg.model, train_cfg=cfg.get('train_cfg'), test_cfg=cfg.get('test_cfg'))print("数据集加载完毕！")# Add an attribute for visualization conveniencemodel.CLASSES = datasets[0].CLASSES# Create work_dirmmcv.mkdir_or_exist(osp.abspath(cfg.work_dir))train_detector(model, datasets, cfg, distributed=False, validate=True)

！！！成功-权重文件

附上第一轮的结果，好像不会太离谱了

# 改之前DONE (t=8.48s). Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.005 Average Precision (AP) @[ IoU=0.50 | area= all | maxDets=1000 ] = 0.020 Average Precision (AP) @[ IoU=0.75 | area= all | maxDets=1000 ] = 0.001 Average Precision (AP) @[ IoU=0.50:0.95 | area= small | maxDets=1000 ] = 0.001 Average Precision (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=1000 ] = 0.005 Average Precision (AP) @[ IoU=0.50:0.95 | area= large | maxDets=1000 ] = 0.009 Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.035 Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets=300 ] = 0.035 Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets=1000 ] = 0.035 Average Recall (AR) @[ IoU=0.50:0.95 | area= small | maxDets=1000 ] = 0.010 Average Recall (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=1000 ] = 0.030# 改了之后Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.076Average Precision (AP) @[ IoU=0.50 | area= all | maxDets=1000 ] = 0.187Average Precision (AP) @[ IoU=0.75 | area= all | maxDets=1000 ] = 0.049Average Precision (AP) @[ IoU=0.50:0.95 | area= small | maxDets=1000 ] = 0.008Average Precision (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=1000 ] = 0.092Average Precision (AP) @[ IoU=0.50:0.95 | area= large | maxDets=1000 ] = 0.101Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.167Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets=300 ] = 0.167Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets=1000 ] = 0.167Average Recall (AR) @[ IoU=0.50:0.95 | area= small | maxDets=1000 ] = 0.032Average Recall (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=1000 ] = 0.182Average Recall (AR) @[ IoU=0.50:0.95 | area= large | maxDets=1000 ] = 0.246

但是有新的bug，后面在解决，应该是配置文件的问题

附录

swintransformer挺牛的，大家可以自己试试看

​​microsoft/Swin-Transformer: This is an official implementation for “Swin Transformer: Hierarchical Vision Transformer using Shifted Windows”. (github.com)​​

最后附上mmdetection和mmsegmnetation的对照表。

版权声明：本文内容由网络用户投稿，版权归原作者所有，本站不拥有其著作权，亦不承担相应法律责任。如果您发现本站中有涉嫌抄袭或描述失实的内容，请联系我们jiasou666@gmail.com 处理，核实后本网站将在24小时内删除侵权内容。