直接使用
请打开基于MOCOV2的自监督学习示例，并点击右上角 “ 在DSW中打开” 。
EasyCV自监督训练-MOCOv2
自监督学习（Self-Supervised Learning）能利用大量无标注的数据进行表征学习，然后在特定下游任务上对参数进行微调。通过这样的方式，能够在较少有标注数据上取得优于有监督学习方法的精度。
近年来，自监督学习受到了越来越多的关注，如Yann Lecun也在 AAAI 上讲 Self-Supervised Learning 是未来的大势所趋。在CV领域涌现了如SwAV、MOCO、DINO、MoBY等一系列工作。
本文将介绍如何在pai-dsw基于EasyCV快速使用MOCOV2进行度量模型的训练、微调、推理。
运行环境要求
PAI-Pytorch 1.7/1.8镜像， GPU机型 P100 or V100， 内存 32G
安装依赖包
注：在PAI-DSW docker中无需安装相关依赖，可跳过此1，2步骤， 在本地notebook环境中执行1，2 步骤安装环境
1、获取torch和cuda版本，并根据版本号修改mmcv安装命令，安装对应版本的mmcv和nvidia-dali
import torch import os os.environ['CUDA']='cu' + torch.version.cuda.replace('.', '') os.environ['Torch']='torch'+torch.version.__version__.replace('+PAI', '') !echo $CUDA !echo $Torch
[2023-02-03 11:17:50,705.705 dsw-16577-7f6b8db66d-qvgjd:382 INFO utils.py:30] NOTICE: PAIDEBUGGER is turned off.
cu101 torch1.8.2
# install some python deps ! pip install --upgrade tqdm ! pip install mmcv-full==1.4.4 -f https://download.openmmlab.com/mmcv/dist/cu101/torch1.8.0/index.html ! pip install http://pai-vision-data-hz.oss-cn-zhangjiakou.aliyuncs.com/third_party/nvidia_dali_cuda100-0.25.0-1535750-py3-none-manylinux2014_x86_64.whl
2、安装EasyCV算法包 注：在PAI-DSW docker中预安装了pai-easycv库，可跳过该步骤，若训练测试过程中报错，尝试用下方命令更新easycv版本
#pip install pai-easycv ! echo y | pip uninstall pai-easycv easycv !pip install pai-easycv
3、简单验证
from easycv.apis import *
正式工作
数据准备
自监督训练只需要提供无标注图片即可进行， 你可以下载ImageNet数据，或者使用你自己的图片数据。需要提供一个包含若干图片的文件夹路径p，以及一个文件列表，文件列表中是每个图片相对图片目录p的路径.
图片文件夹结构示例如下, 文件夹路径为./images
images/ ├── 0001.jpg ├── 0002.jpg ├── 0003.jpg |... └── 9999.jpg 文件列表内容示例如下 0001.jpg 0002.jpg 0003.jpg ... 9999.jpg
为了快速走通流程，我们也提供了一个小的示例数据集，执行如下命令下载解压
! wget http://pai-vision-data-hz.oss-cn-zhangjiakou.aliyuncs.com/data/imagenet_raw_demo/imagenet_raw_demo.tar.gz && tar -zxf imagenet_raw_demo.tar.gz
Will not apply HSTS. The HSTS database must be a regular and non-world-writable file. ERROR: could not open HSTS store at '/root/.wget-hsts'. HSTS will be disabled. --2023-02-03 11:20:43-- http://pai-vision-data-hz.oss-cn-zhangjiakou.aliyuncs.com/data/imagenet_raw_demo/imagenet_raw_demo.tar.gz Resolving pai-vision-data-hz.oss-cn-zhangjiakou.aliyuncs.com (pai-vision-data-hz.oss-cn-zhangjiakou.aliyuncs.com)... 39.98.20.13 Connecting to pai-vision-data-hz.oss-cn-zhangjiakou.aliyuncs.com (pai-vision-data-hz.oss-cn-zhangjiakou.aliyuncs.com)|39.98.20.13|:80... connected. HTTP request sent, awaiting response... 200 OK Length: 33109699 (32M) [application/x-gzip] Saving to: ‘imagenet_raw_demo.tar.gz’ imagenet_raw_demo.t 100%[===================>] 31.58M 17.5MB/s in 1.8s 2023-02-03 11:20:45 (17.5 MB/s) - ‘imagenet_raw_demo.tar.gz’ saved [33109699/33109699]
# 重命名文件夹 ! mv imagenet_raw_demo imagenet_raw
训练模型
这个Demo中我们采用mocov2自监督算法训练ResNet50 主干网络， 下载示例配置文件
! rm -rf mocov2_rn50_8xb32_200e_jpg.py ! wget http://pai-vision-data-hz.oss-cn-zhangjiakou.aliyuncs.com/release/doc/easycv/configs/selfsup/mocov2/mocov2_rn50_8xb32_200e_jpg.py
Will not apply HSTS. The HSTS database must be a regular and non-world-writable file. ERROR: could not open HSTS store at '/root/.wget-hsts'. HSTS will be disabled. --2023-02-03 11:39:05-- http://pai-vision-data-hz.oss-cn-zhangjiakou.aliyuncs.com/release/doc/easycv/configs/selfsup/mocov2/mocov2_rn50_8xb32_200e_jpg.py Resolving pai-vision-data-hz.oss-cn-zhangjiakou.aliyuncs.com (pai-vision-data-hz.oss-cn-zhangjiakou.aliyuncs.com)... 39.98.20.13 Connecting to pai-vision-data-hz.oss-cn-zhangjiakou.aliyuncs.com (pai-vision-data-hz.oss-cn-zhangjiakou.aliyuncs.com)|39.98.20.13|:80... connected. HTTP request sent, awaiting response... 200 OK Length: 2182 (2.1K) [text/x-python] Saving to: ‘mocov2_rn50_8xb32_200e_jpg.py’ mocov2_rn50_8xb32_2 100%[===================>] 2.13K --.-KB/s in 0s 2023-02-03 11:39:05 (223 MB/s) - ‘mocov2_rn50_8xb32_200e_jpg.py’ saved [2182/2182]
为了缩短训练时间，打开配置文件 mae_vit_base_patch16_8xb64_1600e.py，修改total_epoch参数为10， 每隔1次迭代打印一次日志。
# runtime settings total_epochs = 10 # log config log_config=dict(interval=1)
正式训练时，建议使用单机8卡配合该配置文件使用，如果要使用单机单卡，建议调小optimizer.lr初始学习率
# 查看easycv安装位置 import easycv print(easycv.__file__)
/home/pai/lib/python3.6/site-packages/easycv/__init__.py
!python -m torch.distributed.launch --nproc_per_node=1 --master_port=29930 \ /home/pai/lib/python3.6/site-packages/easycv/tools/train.py mocov2_rn50_8xb32_200e_jpg.py --work_dir work_dir/selfsup/jpg/rn50_mocov2 --launcher pytorch
使用自监督模型进行特征抽取
模型导出
模型导出会对自监督模型信息裁剪，保留特征抽取必要的backbone和head
# 查看训练产生的pt文件 ! ls work_dir/selfsup/jpg/rn50_mocov2/*.pth
work_dir/selfsup/jpg/rn50_mocov2/epoch_10.pth work_dir/selfsup/jpg/rn50_mocov2/epoch_2.pth work_dir/selfsup/jpg/rn50_mocov2/epoch_4.pth work_dir/selfsup/jpg/rn50_mocov2/epoch_6.pth work_dir/selfsup/jpg/rn50_mocov2/epoch_8.pth
! python -m easycv.tools.export mocov2_rn50_8xb32_200e_jpg.py work_dir/selfsup/jpg/rn50_mocov2/epoch_10.pth work_dir/selfsup/jpg/rn50_mocov2/export.pth
[2023-02-03 11:30:39,553.553 dsw-16577-7f6b8db66d-qvgjd:1615 INFO utils.py:30] NOTICE: PAIDEBUGGER is turned off. mocov2_rn50_8xb32_200e_jpg.py load checkpoint from local path: work_dir/selfsup/jpg/rn50_mocov2/epoch_10.pth
#下载测试图片 ! wget http://pai-vision-data-hz.oss-cn-zhangjiakou.aliyuncs.com/pretrained_models/easycv/product_detection/248347732153_1040.jpg
Will not apply HSTS. The HSTS database must be a regular and non-world-writable file. ERROR: could not open HSTS store at '/root/.wget-hsts'. HSTS will be disabled. --2023-02-03 11:30:53-- http://pai-vision-data-hz.oss-cn-zhangjiakou.aliyuncs.com/pretrained_models/easycv/product_detection/248347732153_1040.jpg Resolving pai-vision-data-hz.oss-cn-zhangjiakou.aliyuncs.com (pai-vision-data-hz.oss-cn-zhangjiakou.aliyuncs.com)... 39.98.20.13 Connecting to pai-vision-data-hz.oss-cn-zhangjiakou.aliyuncs.com (pai-vision-data-hz.oss-cn-zhangjiakou.aliyuncs.com)|39.98.20.13|:80... connected. HTTP request sent, awaiting response... 200 OK Length: 122788 (120K) [image/jpeg] Saving to: ‘248347732153_1040.jpg’ 248347732153_1040.j 100%[===================>] 119.91K --.-KB/s in 0.1s 2023-02-03 11:30:53 (1.06 MB/s) - ‘248347732153_1040.jpg’ saved [122788/122788]
import cv2 from easycv.predictors.feature_extractor import TorchFeatureExtractor # 修改output_ckpt指向 output_ckpt = 'work_dir/selfsup/jpg/rn50_mocov2/export.pth' fe = TorchFeatureExtractor(output_ckpt) img = cv2.imread('248347732153_1040.jpg') img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB) feature = fe.predict([img]) print(feature[0]['feature'].shape) print(feature[0])
load model from init weights load checkpoint from local path: work_dir/selfsup/jpg/rn50_mocov2/export.pth The model and loaded state dict do not match exactly missing keys in source state_dict: head_0.fc_cls.weight, head_0.fc_cls.bias (2048,) {'feature': array([0.60371155, 1.058624 , 0.8255347 , ..., 0.3980614 , 0.8326854 , 0.6025144 ], dtype=float32)}
自监督预训练+ 图像分类finetune
参考EasyCV图像分类的demo， 在训练时加上--load_from 参数，使用自监督预训练的模型权重， 注意这里不需要使用
! python -m easycv.tools.train r50.py --work_dir work_dirs/classification/cifar10/r50 --load_from work_dir/selfsup/jpg/rn50_mocov2/epoch_10.pth
