Example of Mask2Former Image Segmentation Based on EasyCV

Introduction: EasyCV is an all in one visual algorithm modeling tool based on Python, with self-monitoring learning and Transformer technology as the core, and contains SOTA algorithms for image classification, measurement learning, target detection, pose recognition and other visual tasks. This paper will introduce the training and prediction of image segmentation models using EasyCV.

EasyCV Image Segmentation - Mask2Fformer

Mask2Former is a unified Transformer based image segmentation architecture, which can achieve semantic segmentation, instance segmentation and panoramic segmentation tasks under the same model structure.

This paper will introduce how to use EasyCV to train image segmentation models based on Mask2Fformer, and how to use the trained models to predict image segmentation

Operating environment requirements

PAI Python image or native Python 1.8++environment GPU machine with more than 32G memory

Install dependent packages

Note: There is no need to install dependencies in PAI-DSW docker. You can skip this part and execute it in the local notebook environment

1. First, install pytorch and the corresponding version of torchvision, and support Python version 1.8 or above

# install pytorch and torch vision

! conda install --yes pytorch==1.10.0 torchvision==0.11.0 -c pytorch

2. Get the torch and cuda versions, and install the corresponding versions of mmcv and 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 version: $CUDA"

! echo "pytorch version: $Torch"

# install some python deps

! pip install mmdet

! pip install mmcv-full==1.4.4 -f https://download.openmmlab.com/mmcv/dist/ ${CUDA}/${Torch}/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

3. Install EasyCV algorithm package

pip install pai-easycv

4. Simple verification

from easycv. apis import *

5. Install deformable_ attention

import easycv

print(easycv.__file__)

#Enter the easycv installation directory to compile deformable_ attention

! cd /home/pai/lib/python3.6/site-packages/easycv/thirdparty/deformable_ attention && python setup. py build install

model training

In this Demo, we use the Mask2Former image segmentation algorithm to train the ResNet50 backbone network and download the sample configuration file

To adapt to small data, we set the configuration file mask2Former_ r50_ 8xb2_ e50_ The following fields of instance.py are modified to reduce the number of training epochs and increase the frequency of printing logs.

Note: If COCO complete data training is used, it is recommended to use a single 8-card for training to ensure the effect;

To ensure the model effect, we finetune the pre training model and execute the following command to start the training.