devbox
寸_铁
这个屌丝很懒，什么也没留下！
关注作者
热门标签
热门文章
当前位置:   article > 正文

自定义数据集 实例分割 MaskFormer+SOLOV2_mask2former 自定数据集

作者：寸_铁 | 2024-08-08 17:14:04

赞

踩

mask2former 自定数据集



facebookresearch/Mask2Former: Code release for "Masked-attention Mask Transformer for Universal Image Segmentation" (github.com

facebookresearch/Mask2Former: Code release for "Masked-attention Mask Transformer for Universal Image Segmentation" (github.com

WXinlong/SOLO：用于实例分割的 SOLO 和 SOLOv2，ECCV 2020 和 NeurIPS 2020。 (github.com)

1.MaskFormer

1.1 环境配置 Ubuntu20.04+cuda11.1

conda create --name mask2former python=3.8 -y
conda activate mask2former
conda install pytorch==1.9.0 torchvision==0.10.0 cudatoolkit=11.1 -c pytorch -c nvidia
pip install -U opencv-python
git clone git@github.com:facebookresearch/detectron2.git

上面这个git不好用的话就手动下载这个代码
cd detectron2
pip install -e .
pip install git+https://github.com/cocodataset/panopticapi.git
 

cd ..
git clone git@github.com:facebookresearch/Mask2Former.git

同理，可以手动下载这个源码
cd Mask2Former
pip install -r requirements.txt
cd mask2former/modeling/pixel_decoder/ops
sh make.sh

 上述内容就是完整的配置环境的过程

1.2 数据集预处理

1.2.1 划分为训练集和验证集

  1. import os
  2. import json
  3. import shutil
  4. from sklearn.model_selection import train_test_split
  5.  
  6. # 文件夹路径
  7. img_folder = '/hy-tmp/pic'
  8. mask_folder = '/hy-tmp/cv2_mask'
  9. output_dir = '/hy-tmp/output'
  10.  
  11. # 创建输出文件夹
  12. train_img_folder = os.path.join(output_dir, 'train2017')
  13. train_mask_folder = os.path.join(output_dir, 'train2017_masks')
  14. val_img_folder = os.path.join(output_dir, 'val2017')
  15. val_mask_folder = os.path.join(output_dir, 'val2017_masks')
  16.  
  17. os.makedirs(train_img_folder, exist_ok=True)
  18. os.makedirs(train_mask_folder, exist_ok=True)
  19. os.makedirs(val_img_folder, exist_ok=True)
  20. os.makedirs(val_mask_folder, exist_ok=True)
  21.  
  22. # 读取所有图像文件名
  23. img_files = [f for f in os.listdir(img_folder) if os.path.isfile(os.path.join(img_folder, f))]
  24.  
  25. # 使用 train_test_split 划分数据集
  26. train_files, val_files = train_test_split(img_files, test_size=0.2, random_state=42)
  27.  
  28. # 复制文件到相应的文件夹
  29. for img_filename in train_files:
  30.     img_name, img_ext = os.path.splitext(img_filename)
  31.     shutil.copy(os.path.join(img_folder, img_filename), os.path.join(train_img_folder, img_filename))
  32.     shutil.copy(os.path.join(mask_folder, img_name + '.png'), os.path.join(train_mask_folder, img_name + '.png'))
  33.  
  34. for img_filename in val_files:
  35.     img_name, img_ext = os.path.splitext(img_filename)
  36.     shutil.copy(os.path.join(img_folder, img_filename), os.path.join(val_img_folder, img_filename))
  37.     shutil.copy(os.path.join(mask_folder, img_name + '.png'), os.path.join(val_mask_folder, img_name + '.png'))
  38.  
  39. print("Dataset split completed.")

1.2.2 转换为json格式

  1. import os
  2. import json
  3. import cv2
  4. import numpy as np
  5. from pycocotools import mask as coco_mask
  6. from PIL import Image
  7.  
  8. # 文件夹路径
  9. img_folder = '/hy-tmp/pic'
  10. mask_folder = '/hy-tmp/cv2_mask'
  11. output_dir = '/hy-tmp/1'
  12. train_img_folder = '/hy-tmp/output/train2017'
  13. train_mask_folder = '/hy-tmp/output/train2017_masks'
  14. val_img_folder = '/hy-tmp/output/val2017'
  15. val_mask_folder = '/hy-tmp/output/val2017_masks'
  16.  
  17. # 初始化 COCO 格式字典
  18. def init_coco_format():
  19.     return {
  20.         "images": [],
  21.         "annotations": [],
  22.         "categories": [{"id": 1, "name": "object", "supercategory": "object"}]
  23.     }
  24.  
  25. def create_annotations(img_folder, mask_folder, output_json_path):
  26.     coco_format = init_coco_format()
  27.     annotation_id = 1
  28.     image_id = 1
  29.  
  30.     for img_filename in os.listdir(img_folder):
  31.         img_name, img_ext = os.path.splitext(img_filename)
  32.         img_path = os.path.join(img_folder, img_filename)
  33.         mask_path = os.path.join(mask_folder, img_name + '.png')
  34.  
  35.         if not os.path.exists(mask_path):
  36.             continue
  37.  
  38.         # 读取图像
  39.         img = Image.open(img_path)
  40.         width, height = img.size
  41.  
  42.         # 生成图像信息
  43.         image_info = {
  44.             "id": image_id,
  45.             "file_name": img_filename,
  46.             "width": width,
  47.             "height": height
  48.         }
  49.         coco_format["images"].append(image_info)
  50.  
  51.         # 读取掩码
  52.         mask = Image.open(mask_path)
  53.         mask = np.array(mask)
  54.  
  55.         # 寻找掩码的边界框
  56.         contours, _ = cv2.findContours(mask, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
  57.         for contour in contours:
  58.             x, y, w, h = cv2.boundingRect(contour)
  59.             bbox = [x, y, w, h]
  60.  
  61.             # 生成掩码的二进制格式
  62.             binary_mask = np.zeros((height, width), dtype=np.uint8)
  63.             cv2.drawContours(binary_mask, [contour], -1, 1, -1)
  64.             encoded_mask = coco_mask.encode(np.asfortranarray(binary_mask))
  65.  
  66.             # 将 RLE 编码转换为可序列化的格式
  67.             rle = {
  68.                 'counts': encoded_mask['counts'].decode('utf-8'),  # 转换为字符串
  69.                 'size': encoded_mask['size']
  70.             }
  71.  
  72.             annotation = {
  73.                 "id": annotation_id,
  74.                 "image_id": image_id,
  75.                 "category_id": 1,
  76.                 "segmentation": rle,
  77.                 "area": float(coco_mask.area(encoded_mask)),
  78.                 "bbox": bbox,
  79.                 "iscrowd": 0
  80.             }
  81.  
  82.             coco_format["annotations"].append(annotation)
  83.             annotation_id += 1
  84.  
  85.         image_id += 1
  86.  
  87.     # 保存到JSON文件
  88.     with open(output_json_path, 'w') as f:
  89.         json.dump(coco_format, f)
  90.  
  91.     print(f"COCO format annotation file created at {output_json_path}.")
  92.  
  93. # 创建注释文件
  94. create_annotations(train_img_folder, train_mask_folder, os.path.join(output_dir, 'annotations/instances_train2017.json'))
  95. create_annotations(val_img_folder, val_mask_folder, os.path.join(output_dir, 'annotations/instances_val2017.json'))

1.2.3 实例分割数据集

  1. import os
  2. import json
  3. import cv2
  4. import numpy as np
  5. from pycocotools import mask as coco_mask
  6. from PIL import Image
  7. from sklearn.model_selection import train_test_split
  8. import shutil
  9.  
  10. # 文件夹路径
  11. img_folder = '/hy-tmp/pic'
  12. mask_folder = '/hy-tmp/cv2_mask'
  13. output_dir = '/hy-tmp/json'
  14.  
  15. # 创建输出文件夹
  16. train_img_folder = os.path.join(output_dir, 'train2017')
  17. train_mask_folder = os.path.join(output_dir, 'train2017_masks')
  18. val_img_folder = os.path.join(output_dir, 'val2017')
  19. val_mask_folder = os.path.join(output_dir, 'val2017_masks')
  20.  
  21. os.makedirs(train_img_folder, exist_ok=True)
  22. os.makedirs(train_mask_folder, exist_ok=True)
  23. os.makedirs(val_img_folder, exist_ok=True)
  24. os.makedirs(val_mask_folder, exist_ok=True)
  25.  
  26. # 读取所有图像文件名
  27. img_files = [f for f in os.listdir(img_folder) if os.path.isfile(os.path.join(img_folder, f))]
  28.  
  29. # 使用 train_test_split 划分数据集
  30. train_files, val_files = train_test_split(img_files, test_size=0.2, random_state=42)
  31.  
  32. # 复制文件到相应的文件夹
  33. for img_filename in train_files:
  34.     img_name, img_ext = os.path.splitext(img_filename)
  35.     shutil.copy(os.path.join(img_folder, img_filename), os.path.join(train_img_folder, img_filename))
  36.     shutil.copy(os.path.join(mask_folder, img_name + '.png'), os.path.join(train_mask_folder, img_name + '.png'))
  37.  
  38. for img_filename in val_files:
  39.     img_name, img_ext = os.path.splitext(img_filename)
  40.     shutil.copy(os.path.join(img_folder, img_filename), os.path.join(val_img_folder, img_filename))
  41.     shutil.copy(os.path.join(mask_folder, img_name + '.png'), os.path.join(val_mask_folder, img_name + '.png'))
  42.  
  43. print("Dataset split completed.")
  44.  
  45. # 初始化 COCO 格式字典
  46. def init_coco_format():
  47.     return {
  48.         "images": [],
  49.         "annotations": [],
  50.         "categories": [{"id": 1, "name": "object", "supercategory": "object"}]
  51.     }
  52.  
  53. def create_instance_annotations(img_folder, mask_folder, output_json_path):
  54.     coco_format = init_coco_format()
  55.     annotation_id = 1
  56.     image_id = 1
  57.  
  58.     for img_filename in os.listdir(img_folder):
  59.         img_name, img_ext = os.path.splitext(img_filename)
  60.         img_path = os.path.join(img_folder, img_filename)
  61.         mask_path = os.path.join(mask_folder, img_name + '.png')
  62.  
  63.         if not os.path.exists(mask_path):
  64.             continue
  65.  
  66.         # 读取图像
  67.         img = Image.open(img_path)
  68.         width, height = img.size
  69.  
  70.         # 生成图像信息
  71.         image_info = {
  72.             "id": image_id,
  73.             "file_name": img_filename,
  74.             "width": width,
  75.             "height": height
  76.         }
  77.         coco_format["images"].append(image_info)
  78.  
  79.         # 读取掩码
  80.         mask = Image.open(mask_path)
  81.         mask = np.array(mask)
  82.  
  83.         # 寻找掩码的边界框
  84.         contours, _ = cv2.findContours(mask, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
  85.         for contour in contours:
  86.             x, y, w, h = cv2.boundingRect(contour)
  87.             bbox = [x, y, w, h]
  88.  
  89.             # 生成掩码的二进制格式
  90.             binary_mask = np.zeros((height, width), dtype=np.uint8)
  91.             cv2.drawContours(binary_mask, [contour], -1, 1, -1)
  92.             encoded_mask = coco_mask.encode(np.asfortranarray(binary_mask))
  93.  
  94.             # 将 RLE 编码转换为可序列化的格式
  95.             rle = {
  96.                 'counts': encoded_mask['counts'].decode('utf-8'),  # 转换为字符串
  97.                 'size': encoded_mask['size']
  98.             }
  99.  
  100.             annotation = {
  101.                 "id": annotation_id,
  102.                 "image_id": image_id,
  103.                 "category_id": 1,
  104.                 "segmentation": rle,
  105.                 "area": float(coco_mask.area(encoded_mask)),
  106.                 "bbox": bbox,
  107.                 "iscrowd": 0
  108.             }
  109.  
  110.             coco_format["annotations"].append(annotation)
  111.             annotation_id += 1
  112.  
  113.         image_id += 1
  114.  
  115.     # 保存到JSON文件
  116.     with open(output_json_path, 'w') as f:
  117.         json.dump(coco_format, f)
  118.  
  119.     print(f"COCO format instance annotation file created at {output_json_path}.")
  120.  
  121. # 创建实例分割注释文件
  122. create_instance_annotations(train_img_folder, train_mask_folder, os.path.join(output_dir, 'annotations/instances_train2017.json'))
  123. create_instance_annotations(val_img_folder, val_mask_folder, os.path.join(output_dir, 'annotations/instances_val2017.json'))

1.2.4 全景分割数据集

  1. import os
  2. import json
  3. import cv2
  4. import numpy as np
  5. from pycocotools import mask as coco_mask
  6. from PIL import Image
  7.  
  8. # 初始化 COCO 格式字典
  9. def init_coco_format():
  10.     return {
  11.         "images": [],
  12.         "annotations": [],
  13.         "categories": [{"id": 1, "name": "object", "supercategory": "object"}],
  14.         "licenses": [],
  15.         "info": {
  16.             "year": 2023,
  17.             "version": "1.0",
  18.             "description": "COCO Panoptic Dataset",
  19.             "contributor": "",
  20.             "url": "",
  21.             "date_created": "2023-01-01"
  22.         }
  23.     }
  24.  
  25. def create_panoptic_annotations(img_folder, mask_folder, output_json_path):
  26.     coco_format = init_coco_format()
  27.     panoptic_annotations = []
  28.     annotation_id = 1
  29.     image_id = 1
  30.  
  31.     for img_filename in os.listdir(img_folder):
  32.         img_name, img_ext = os.path.splitext(img_filename)
  33.         img_path = os.path.join(img_folder, img_filename)
  34.         mask_path = os.path.join(mask_folder, img_name + '.png')
  35.  
  36.         if not os.path.exists(mask_path):
  37.             continue
  38.  
  39.         # 读取图像
  40.         img = Image.open(img_path)
  41.         width, height = img.size
  42.  
  43.         # 生成图像信息
  44.         image_info = {
  45.             "id": image_id,
  46.             "file_name": img_filename,
  47.             "width": width,
  48.             "height": height
  49.         }
  50.         coco_format["images"].append(image_info)
  51.  
  52.         # 读取掩码
  53.         mask = Image.open(mask_path)
  54.         mask = np.array(mask)
  55.  
  56.         # 生成全景注释信息
  57.         panoptic_annotation = {
  58.             "image_id": image_id,
  59.             "file_name": img_filename.replace('.jpg', '.png'),
  60.             "segments_info": []
  61.         }
  62.  
  63.         # 寻找掩码的边界框
  64.         contours, _ = cv2.findContours(mask, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
  65.         for contour in contours:
  66.             x, y, w, h = cv2.boundingRect(contour)
  67.             bbox = [x, y, w, h]
  68.  
  69.             # 生成掩码的二进制格式
  70.             binary_mask = np.zeros((height, width), dtype=np.uint8)
  71.             cv2.drawContours(binary_mask, [contour], -1, 1, -1)
  72.             encoded_mask = coco_mask.encode(np.asfortranarray(binary_mask))
  73.  
  74.             # 将 RLE 编码转换为可序列化的格式
  75.             rle = {
  76.                 'counts': encoded_mask['counts'].decode('utf-8'),  # 转换为字符串
  77.                 'size': encoded_mask['size']
  78.             }
  79.  
  80.             segment_info = {
  81.                 "id": annotation_id,
  82.                 "category_id": 1,
  83.                 "iscrowd": 0,
  84.                 "bbox": bbox,
  85.                 "area": float(coco_mask.area(encoded_mask)),
  86.                 "segmentation": rle
  87.             }
  88.  
  89.             panoptic_annotation["segments_info"].append(segment_info)
  90.             annotation_id += 1
  91.  
  92.         panoptic_annotations.append(panoptic_annotation)
  93.         image_id += 1
  94.  
  95.     # 保存到JSON文件
  96.     os.makedirs(os.path.dirname(output_json_path), exist_ok=True)
  97.     with open(output_json_path, 'w') as f:
  98.         json.dump({
  99.             "images": coco_format["images"],
  100.             "annotations": panoptic_annotations,
  101.             "categories": coco_format["categories"],
  102.             "licenses": coco_format["licenses"],
  103.             "info": coco_format["info"]
  104.         }, f)
  105.  
  106.     print(f"COCO format panoptic annotation file created at {output_json_path}.")
  107.  
  108. # 创建全景注释文件
  109. create_panoptic_annotations(train_img_folder, train_mask_folder, os.path.join(output_dir, 'annotations/panoptic_train2017.json'))
  110. create_panoptic_annotations(val_img_folder, val_mask_folder, os.path.join(output_dir, 'annotations/panoptic_val2017.json'))

1.2.5 全景分割数据集找到对应的图片（.png格式）

  1. import json
  2. import os
  3. import shutil
  4.  
  5. # 定义文件路径
  6. json_path = 'path/to/panoptic_train2017.json'
  7. img_dir = 'path/to/img'
  8. output_dir = 'path/to/output_folder'
  9.  
  10. # 如果输出文件夹不存在，则创建
  11. if not os.path.exists(output_dir):
  12.     os.makedirs(output_dir)
  13.  
  14. # 加载 JSON 文件
  15. with open(json_path, 'r', encoding='utf-8') as f:
  16.     data = json.load(f)
  17.  
  18. # 假设 JSON 文件中有一个键 "images"，其中包含每个图像的相关信息
  19. image_info_list = data.get('images', [])
  20.  
  21. # 遍历每个图像信息并复制对应的图片到输出文件夹
  22. for image_info in image_info_list:
  23.     # 假设每个图像信息包含 "file_name" 键
  24.     file_name = image_info.get('file_name')
  25.     
  26.     if file_name:
  27.         # 构建原始图片路径
  28.         img_path = os.path.join(img_dir, file_name)
  29.         
  30.         # 构建目标图片路径
  31.         output_path = os.path.join(output_dir, file_name)
  32.         
  33.         # 检查文件是否存在并复制
  34.         if os.path.isfile(img_path):
  35.             shutil.copy(img_path, output_path)
  36.             print(f"图片 {file_name} 已复制到 {output_path}")
  37.         else:
  38.             print(f"图片 {file_name} 不存在于路径 {img_path}")
  39.     else:
  40.         print("图像信息中缺少 'file_name' 键")
  41.  
  42.

1.3训练过程

单卡训练太慢了，需要好几天，选择多卡训练

单卡训练的代码

  1. python train_net.py \
  2.   --config-file configs/coco/panoptic-segmentation/maskformer2_R50_bs16_50ep.yaml \
  3.   --num-gpus 1 SOLVER.IMS_PER_BATCH 2 SOLVER.BASE_LR 0.00025

下图为正式运行的过程

多卡训练的代码

遇到的小插曲

报错

TypeError: init() got an unexpected keyword argument 'dtype'

修改events.py文件以及

from packaging.version import parse as LooseVersion

1.4测试过程

2. SOLOV2

2.1 环境配置

2.1.1选择pytorch版本

torch                     1.5.0+cu101              pypi_0    pypi

2.1.2 编译

git clone https://github.com/WXinlong/SOLO.git
cd SOLO

 pip install -r requirements/build.txt
pip install "git+https://github.com/cocodataset/cocoapi.git#subdirectory=PythonAPI"
pip install -v -e .  # or "python setup.py develop"

 2.2 数据集预处理

将数据集转换为coco数据集格式

2.2.1 把img以及对应的mask转换为train_2017.json以及val_2017.json格式

代码如下

  1. import os
  2. import json
  3. import cv2
  4. import numpy as np
  5. from pycocotools import mask as coco_mask
  6. from PIL import Image
  7. from sklearn.model_selection import train_test_split
  8.  
  9. # 文件夹路径
  10. img_folder = '/hy-tmp/pic'
  11. mask_folder = '/hy-tmp/cv2_mask'
  12. output_train_json_path = '/hy-tmp/coco/annotations/instances_train2017.json'
  13. output_val_json_path = '/hy-tmp/coco/annotations/instances_val2017.json'
  14.  
  15. # 初始化 COCO 格式字典
  16. def init_coco_format():
  17.     return {
  18.         "images": [],
  19.         "annotations": [],
  20.         "categories": [{"id": 1, "name": "object", "supercategory": "object"}]
  21.     }
  22.  
  23. # 读取文件名
  24. img_files = [f for f in os.listdir(img_folder) if os.path.isfile(os.path.join(img_folder, f))]
  25. train_files, val_files = train_test_split(img_files, test_size=0.2, random_state=42)
  26.  
  27. def create_annotations(files, coco_format, start_image_id=1, start_annotation_id=1):
  28.     image_id = start_image_id
  29.     annotation_id = start_annotation_id
  30.  
  31.     for img_filename in files:
  32.         img_name, img_ext = os.path.splitext(img_filename)
  33.         img_path = os.path.join(img_folder, img_filename)
  34.         mask_path = os.path.join(mask_folder, img_name + '.png')  # 假设掩码是 .png 格式
  35.  
  36.         if not os.path.exists(mask_path):
  37.             continue
  38.  
  39.         # 读取图像
  40.         img = Image.open(img_path)
  41.         width, height = img.size
  42.  
  43.         # 生成图像信息
  44.         image_info = {
  45.             "id": image_id,
  46.             "file_name": img_filename,
  47.             "width": width,
  48.             "height": height
  49.         }
  50.         coco_format["images"].append(image_info)
  51.  
  52.         # 读取掩码
  53.         mask = Image.open(mask_path)
  54.         mask = np.array(mask)
  55.  
  56.         # 寻找掩码的边界框
  57.         contours, _ = cv2.findContours(mask, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
  58.         for contour in contours:
  59.             x, y, w, h = cv2.boundingRect(contour)
  60.             bbox = [x, y, w, h]
  61.  
  62.             # 生成掩码的二进制格式
  63.             binary_mask = np.zeros((height, width), dtype=np.uint8)
  64.             cv2.drawContours(binary_mask, [contour], -1, 1, -1)
  65.             encoded_mask = coco_mask.encode(np.asfortranarray(binary_mask))
  66.  
  67.             # 将 RLE 编码转换为可序列化的格式
  68.             rle = {
  69.                 'counts': encoded_mask['counts'].decode('utf-8'),  # 转换为字符串
  70.                 'size': encoded_mask['size']
  71.             }
  72.  
  73.             annotation = {
  74.                 "id": annotation_id,
  75.                 "image_id": image_id,
  76.                 "category_id": 1,
  77.                 "segmentation": rle,
  78.                 "area": float(coco_mask.area(encoded_mask)),
  79.                 "bbox": bbox,
  80.                 "iscrowd": 0
  81.             }
  82.  
  83.             coco_format["annotations"].append(annotation)
  84.             annotation_id += 1
  85.  
  86.         image_id += 1
  87.  
  88.     return coco_format, image_id, annotation_id
  89.  
  90. # 初始化 COCO 格式字典
  91. coco_train_format = init_coco_format()
  92. coco_val_format = init_coco_format()
  93.  
  94. # 创建训练集注释
  95. coco_train_format, train_image_id, train_annotation_id = create_annotations(train_files, coco_train_format)
  96.  
  97. # 创建验证集注释
  98. coco_val_format, _, _ = create_annotations(val_files, coco_val_format, start_image_id=train_image_id, start_annotation_id=train_annotation_id)
  99.  
  100. # 保存训练集注释到JSON文件
  101. with open(output_train_json_path, 'w') as f:
  102.     json.dump(coco_train_format, f)
  103.  
  104. # 保存验证集注释到JSON文件
  105. with open(output_val_json_path, 'w') as f:
  106.     json.dump(coco_val_format, f)
  107.  
  108. print("COCO format annotation files created.")

2.2.2 把数据集划分为train和val

  1. import os
  2. import json
  3. import shutil
  4.  
  5. # 文件夹路径
  6. img_folder = '/hy-tmp/pic'
  7. val_img_folder = '/hy-tmp/coco/train2017'
  8. val_json_path = '/hy-tmp/coco/annotations/instances_train2017.json'
  9.  
  10. # 创建 val2017 文件夹，如果不存在
  11. os.makedirs(val_img_folder, exist_ok=True)
  12.  
  13. # 读取 val2017.json 文件
  14. with open(val_json_path, 'r') as f:
  15.     val_data = json.load(f)
  16.  
  17. # 获取 val2017.json 中所有图片文件名
  18. val_img_files = [img_info['file_name'] for img_info in val_data['images']]
  19.  
  20. # 复制图片到 val2017 文件夹
  21. for img_filename in val_img_files:
  22.     src_path = os.path.join(img_folder, img_filename)
  23.     dst_path = os.path.join(val_img_folder, img_filename)
  24.     if os.path.exists(src_path):
  25.         shutil.copy2(src_path, dst_path)
  26.         print(f"Copied: {src_path} to {dst_path}")
  27.     else:
  28.         print(f"File not found: {src_path}")
  29.  
  30. print("All val2017 images have been copied.")

2.2.3 如果img和mask不是成对数据

首先运行下面的代码，然后在运行2.2.1以及2.2.2

  1. import os
  2.  
  3. # 文件夹路径
  4. img_folder = '/hy-tmp/pic'
  5. mask_folder = '/hy-tmp/cv2_mask'
  6.  
  7. # 获取文件夹中所有文件名（不包括扩展名）
  8. img_files = {os.path.splitext(f)[0] for f in os.listdir(img_folder) if os.path.isfile(os.path.join(img_folder, f))}
  9. mask_files = {os.path.splitext(f)[0] for f in os.listdir(mask_folder) if os.path.isfile(os.path.join(mask_folder, f))}
  10.  
  11. # 找出在img文件夹中但不在mask文件夹中的文件
  12. img_files_to_delete = img_files - mask_files
  13. # 找出在mask文件夹中但不在img文件夹中的文件
  14. mask_files_to_delete = mask_files - img_files
  15.  
  16. # 删除img文件夹中的不对应文件
  17. for file in img_files_to_delete:
  18.     file_path = os.path.join(img_folder, file + '.jpg')  # 假设文件扩展名是.jpg，根据实际情况修改
  19.     if os.path.exists(file_path):
  20.         os.remove(file_path)
  21.         print(f"Deleted from img: {file_path}")
  22.  
  23. # 删除mask文件夹中的不对应文件
  24. for file in mask_files_to_delete:
  25.     file_path = os.path.join(mask_folder, file + '.png')  # 假设文件扩展名是.png，根据实际情况修改
  26.     if os.path.exists(file_path):
  27.         os.remove(file_path)
  28.         print(f"Deleted from mask: {file_path}")
  29.  
  30. print("Deletion completed.")

2.3 使用单卡进行训练

python tools/train.py configs/solov2/solov2_r50_fpn_8gpu_1x.py

成功运行界面如下图所示,目前官方给的epoch是12个，未修改

2.4 使用单卡进行测试

python tools/test_ins.py configs/solov2/solov2_r50_fpn_8gpu_1x.py /hy-tmp/SOLO/work_dirs/solov2_release_r50_fpn_8gpu_1x/latest.pth --show --out results_solo.pkl --eval segm

结果如下图

2.5 使用单卡进行可视化

python tools/test_ins_vis.py configs/solov2/solov2_r50_fpn_8gpu_1x.py  /hy-tmp/SOLO/work_dirs/solov2_release_r50_fpn_8gpu_1x/latest.pth --show --save_dir  work_dirs/vis_solo

结果在work_dirs/vis_solo文件夹下面

声明：本文内容由网友自发贡献，不代表【wpsshop博客】立场，版权归原作者所有，本站不承担相应法律责任。如您发现有侵权的内容，请联系我们。转载请注明出处：https://www.wpsshop.cn/w/寸_铁/article/detail/949200
推荐阅读
相关标签

Copyright © 2003-2013 www.wpsshop.cn 版权所有，并保留所有权利。

  

闽ICP备14008679号