MMDetection (5)

Chapter7 Panoptic segmentation 모델 학습

Panoptic segmentation

• Panoptic segmentation = Semantic segmentation(Class별 구분, 모든 이미지에 대해서) + Instance segmentation(개별 객체별로 구분, 배경에 대해서는 예측하지 않음)

• Panoptic Segmentation의 구성

• Things: 기하학적 구조가 잘 정의되어 있고 사람, 차량, 식물 등의 이미지에서 셀 수 있는 물체

• Stuff: 정확한 기하학적 구조는 없지만 주로 질감과 재질로 특정지어지는 물체(도로, 하늘, 기타 배경)

• 실제 적용하는 곳
  -> 자율주행
  -> Medical imaging
  -> Smart cities
  -> AR/VR
  -> Surveillance and security

데이터 준비

• Dataset의 annotation파일 불러오기

with open ('파일 경로/panoptic_train2017.json', 'r') as t:
    t_data = t.read()

j = json.loads(t_data)

j.keys()

len(j['images'])

j['info']

j['annotations'][0]

j['images'][0]

• ID를 기준으로 파일 처리

image_id_list = [images_info['id'] for images_info in j['images']]

ann_id_list = [ann_info['image_id'] for ann_info in j['annotations']]

image_id_set = set(image_id_list)
ann_id_set = set(ann_id_list)

len(image_id_set.intersection(ann_id_set))

• 2천개씩 2set

import random
random.shuffle(image_id_list)

first_image_id_list = image_id_list[:2000]
second_image_id_list = image_id_list[2000:4000]

import tqdm

first_image_list = []

for image_id in tqdm.tqdm(first_image_id_list) :
    for image_data in j['images']:
        if image_data['id'] == image_id :
            first_image_list.append(image_data)
            continue

second_image_list = []

for image_id in tqdm.tqdm(second_image_id_list) :
    for image_data in j['images']:
        if image_data['id'] == image_id :
            second_image_list.append(image_data)
            continue

import tqdm

first_ann_list = []

for image_id in tqdm.tqdm(first_image_id_list) :
    for ann_data in j['annotations']:
        if ann_data['image_id'] == image_id :
            first_ann_list.append(ann_data)
            continue

second_ann_list = []

for image_id in tqdm.tqdm(second_image_id_list) :
    for ann_data in j['annotations']:
        if ann_data['image_id'] == image_id :
            second_ann_list.append(ann_data)
            continue

• 폴더 구조

# 새폴더를 만들어 줄 함수
def createDirectory(directory):
    try:
        if not os.path.exists(directory):
            os.makedirs(directory)
    except OSError:
        print("Error: Failed to create the directory.")

import os
import shutil
import json

data_root = 'e:/data/coco/'
image_path = 'train2017/'
seg_path = 'annotations/panoptic_train2017/'
ann_path = 'annotations/'
copy_path = 'e:/data/coco/firstset/'
ann_file = 'panoptic_train2017.json'

createDirectory(copy_path)
createDirectory(copy_path+image_path)
createDirectory(copy_path+seg_path)
createDirectory(copy_path+ann_path)


json_data = {'info':j['info'],
             'licenses':j['licenses'],
             'images':first_image_list,
             'annotations':first_ann_list,
             'categories':j['categories']}

with open(data_root+'firstset/'+ann_path+ann_file, 'w')as json_file:
    json.dump(json_data, json_file)

for i in tqdm.tqdm(first_ann_list):
    shutil.copyfile(
        data_root+image_path+i['file_name'].replace('png','jpg'), 
        copy_path+image_path+i['file_name'].replace('png','jpg')
        )
    
    shutil.copyfile(
        data_root+seg_path+i['file_name'], 
        copy_path+seg_path+i['file_name']
        )

• 두번째 set 저장

import os
import shutil

data_root = 'e:/data/coco/'
image_path = 'train2017/'
seg_path = 'annotations/panoptic_train2017/'
ann_path = 'annotations/'
copy_path = 'e:/data/coco/secondset/'
ann_file = 'panoptic_train2017.json'


createDirectory(copy_path)
createDirectory(copy_path+image_path)
createDirectory(copy_path+seg_path)
createDirectory(copy_path+ann_path)

json_data = {'info':j['info'],
             'licenses':j['licenses'],
             'images':second_image_list,
             'annotations':second_ann_list,
             'categories':j['categories']}

with open(data_root+'secondset/'+ann_path+ann_file, 'w')as json_file:
    json.dump(json_data, json_file)

for i in tqdm.tqdm(second_ann_list):
    shutil.copyfile(
        data_root+image_path+i['file_name'].replace('png','jpg'), 
        copy_path+image_path+i['file_name'].replace('png','jpg')
        )
    shutil.copyfile(data_root+seg_path+i['file_name'], 
                    copy_path+seg_path+i['file_name']
                    )

• val 세트 저장

with open ('e:/data/coco/annotations/panoptic_val2017.json', 'r') as t:
    t_data = t.read()

j = json.loads(t_data)

image_id_list = [images_info['id'] for images_info in j['images']]
ann_id_list = [ann_info['image_id'] for ann_info in j['annotations']]

image_id_set = set(image_id_list)
ann_id_set = set(ann_id_list)

len(image_id_set.intersection(ann_id_set))

import random
random.shuffle(image_id_list)

first_image_id_list = image_id_list[:500]
second_image_id_list = image_id_list[500:1000]

import tqdm

first_image_list = []

for image_id in tqdm.tqdm(first_image_id_list) :
    for image_data in j['images']:
        if image_data['id'] == image_id :
            first_image_list.append(image_data)
            continue

second_image_list = []

for image_id in tqdm.tqdm(second_image_id_list) :
    for image_data in j['images']:
        if image_data['id'] == image_id :
            second_image_list.append(image_data)
            continue

import tqdm

first_ann_list = []

for image_id in tqdm.tqdm(first_image_id_list) :
    for ann_data in j['annotations']:
        if ann_data['image_id'] == image_id :
            first_ann_list.append(ann_data)
            continue

second_ann_list = []

for image_id in tqdm.tqdm(second_image_id_list) :
    for ann_data in j['annotations']:
        if ann_data['image_id'] == image_id :
            second_ann_list.append(ann_data)
            continue

import os
import shutil
import json

data_root = 'e:/data/coco/'
image_path = 'val2017/'
seg_path = 'annotations/panoptic_val2017/'
ann_path = 'annotations/'
copy_path = 'e:/data/coco/firstset/'
ann_file = 'panoptic_val2017.json'

createDirectory(copy_path)
createDirectory(copy_path+image_path)
createDirectory(copy_path+seg_path)
createDirectory(copy_path+ann_path)


json_data = {'info':j['info'],
             'licenses':j['licenses'],
             'images':first_image_list,
             'annotations':first_ann_list,
             'categories':j['categories']}

with open(data_root+'firstset/'+ann_path+ann_file, 'w')as json_file:
    json.dump(json_data, json_file)

for i in tqdm.tqdm(first_ann_list):
    shutil.copyfile(
        data_root+image_path+i['file_name'].replace('png','jpg'), 
        copy_path+image_path+i['file_name'].replace('png','jpg')
        )
    shutil.copyfile(
        data_root+seg_path+i['file_name'], 
        copy_path+seg_path+i['file_name']
        )

import os
import shutil

data_root = 'e:/data/coco/'
image_path = 'val2017/'
seg_path = 'annotations/panoptic_val2017/'
ann_path = 'annotations/'
copy_path = 'e:/data/coco/secondset/'
ann_file = 'panoptic_val2017.json'


createDirectory(copy_path)
createDirectory(copy_path+image_path)
createDirectory(copy_path+seg_path)
createDirectory(copy_path+ann_path)

json_data = {'info':j['info'],
             'licenses':j['licenses'],
             'images':second_image_list,
             'annotations':second_ann_list,
             'categories':j['categories']}

with open(data_root+'secondset/'+ann_path+ann_file, 'w')as json_file:
    json.dump(json_data, json_file)

for i in tqdm.tqdm(second_ann_list):
    shutil.copyfile(
        data_root+image_path+i['file_name'].replace('png','jpg'), 
        copy_path+image_path+i['file_name'].replace('png','jpg')
        )
    shutil.copyfile(
        data_root+seg_path+i['file_name'], 
        copy_path+seg_path+i['file_name']
        )

학습하기

• Config 파일 설정

!mim download mmdet --config panoptic_fpn_r101_fpn_1x_coco --dest ./

from mmengine import Config

cfg = Config.fromfile('./panoptic_fpn_r101_fpn_1x_coco.py')

cfg.data_root = 'e:/data/coco/'

cfg.train_dataloader.dataset.ann_file = 'firstset/annotations/panoptic_train2017.json'
cfg.train_dataloader.dataset.data_root = cfg.data_root

cfg.val_dataloader.dataset.ann_file = 'firstset/annotations/panoptic_val2017.json'
cfg.val_dataloader.dataset.data_root = cfg.data_root

cfg.test_dataloader.dataset = cfg.val_dataloader.dataset

cfg.val_evaluator.ann_file = 'e:/data/coco/firstset/annotations/panoptic_val2017.json'
cfg.val_evaluator.seg_prefix = 'e:/data/coco/firstset/annotations/panoptic_val2017'
cfg.test_evaluator = cfg.val_evaluator

cfg.optim_wrapper.optimizer.lr = 0.02 / 8

cfg.visualizer.vis_backends.append({"type":'TensorboardVisBackend'})

cfg.train_cfg.max_epochs = 4

with open('./panoptic_firstset_config.py', 'w')as f:
    f.write(cfg.pretty_text)

• 학습 진행

!python tools/train.py ./panoptic_firstset_config.py

!python tools/test.py panoptic_fpn_r101_fpn_1x_coco.py ./work_dirs/panoptic_fpn_r101_fpn_1x_coco/epoch_4.pth 

• 이어서 학습

from mmengine import Config

cfg = Config.fromfile('./panoptic_fpn_r101_fpn_1x_coco.py')

• 이전 학습 결과 불러옴

cfg.data_root = 'e:/data/coco/'
cfg.load_from = './work_dirs/panoptic_firstset_config/epoch_12.pth' # 이전 학습 결과

cfg.train_dataloader.dataset.ann_file = 'secondset/annotations/panoptic_train2017.json'
cfg.train_dataloader.dataset.data_root = cfg.data_root

cfg.val_dataloader.dataset.ann_file = 'secondset/annotations/panoptic_val2017.json'
cfg.val_dataloader.dataset.data_root = cfg.data_root

cfg.test_dataloader.dataset = cfg.val_dataloader.dataset

cfg.val_evaluator.ann_file = 'e:/data/coco/secondset/annotations/panoptic_val2017.json'
cfg.val_evaluator.seg_prefix = 'e:/data/coco/secondset/annotations/panoptic_val2017/'
cfg.test_evaluator = cfg.val_evaluator

cfg.optim_wrapper.optimizer.lr = 0.02/8

cfg.visualizer.vis_backends.append({"type":'TensorboardVisBackend'})

with open('./panoptic_secondset_config.py','w')as f:
    f.write(cfg.pretty_text)

!python tools/train.py ./panoptic_firstset_config.py

!python tools/test.py panoptic_fpn_r101_fpn_1x_coco.py ./work_dirs/panoptic_fpn_r101_fpn_1x_coco/epoch_4.pth 

이 글은 제로베이스 데이터 취업 스쿨의 강의 자료 일부를 발췌하여 작성되었습니다