机器学习 / A LEAF IN THE GARDEN
语义分割基础
机器学习 · 动手深度学习。来自 Daily Learning 的学习笔记。
简介
语义分割（Semantic segmentation），将图片中每个像素分类到对应类别
分类
- 语义分割:
- 将图片中每个像素分类到对应类别
- 不区分同一类别的不同实例
- 实例分割:
- 将图片中每个像素分类到对应实例
- 区分同一类别的不同实例
- 图像分割:
- 将图片分成不同的区域
- 不指明类别和实例，只表示相关性
实现
代码 4 · In [1]
%matplotlib inline
import os
import torch
import torchvision
import util
util.data.DATA_HUB['voc2012'] = (util.data.DATA_URL + 'VOCtrainval_11-May-2012.tar',
'4e443f8a2eca6b1dac8a6c57641b67dd40621a49')
voc_dir = util.data.download_extract('voc2012', 'VOCdevkit/VOC2012')已保存输出
/root/autodl-tmp/envs/daily_learning/lib/python3.11/site-packages/requests/__init__.py:92: RequestsDependencyWarning: Unable to find acceptable character detection dependency (chardet or charset_normalizer). warnings.warn(
已保存输出
正在从https://d2l-data.s3-accelerate.amazonaws.com/VOCtrainval_11-May-2012.tar下载./data/VOCtrainval_11-May-2012.tar...
已保存输出 · 错误记录
---------------------------------------------------------------------------KeyboardInterrupt Traceback (most recent call last)Cell In[1], line 9
5 import util
6 util.data.DATA_HUB['voc2012'] = (util.data.DATA_URL + 'VOCtrainval_11-May-2012.tar',
7 '4e443f8a2eca6b1dac8a6c57641b67dd40621a49')
8
----> 9 voc_dir = util.data.download_extract('voc2012', 'VOCdevkit/VOC2012')
File ~/autodl-tmp/sync/Daily_Learning/机器学习自学/动手深度学习/util/data.py:97, in download_extract(name, folder)
95 def download_extract(name, folder=None):
96 """下载并解压 zip/tar 文件"""
---> 97 fname = download(name)
98 base_dir = os.path.dirname(fname)
99 data_dir, ext = os.path.splitext(fname)
File ~/autodl-tmp/sync/Daily_Learning/机器学习自学/动手深度学习/util/data.py:92, in download(name, cache_dir)
90 r = requests.get(url, stream=True)
91 with open(fname, 'wb') as f:
---> 92 f.write(r.content)
93 return fname
File ~/autodl-tmp/envs/daily_learning/lib/python3.11/site-packages/requests/models.py:1042, in Response.content(self)
1040 self._content = None
1041 else:
-> 1042 self._content = b"".join(self.iter_content(CONTENT_CHUNK_SIZE)) or b""
1044 self._content_consumed = True
1045 # don't need to release the connection; that's been handled by urllib3
1046 # since we exhausted the data.
File ~/autodl-tmp/envs/daily_learning/lib/python3.11/site-packages/requests/models.py:937, in Response.iter_content.<locals>.generate()
935 if hasattr(self.raw, "stream"):
936 try:
--> 937 yield from self.raw.stream(chunk_size, decode_content=True)
938 except ProtocolError as e:
939 raise ChunkedEncodingError(e)
File ~/autodl-tmp/envs/daily_learning/lib/python3.11/site-packages/urllib3/response.py:1267, in HTTPResponse.stream(self, amt, decode_content)
1261 else:
1262 while (
1263 not is_fp_closed(self._fp)
1264 or len(self._decoded_buffer) > 0
1265 or (self._decoder and self._decoder.has_unconsumed_tail)
1266 ):
-> 1267 data = self.read(amt=amt, decode_content=decode_content)
1269 if data:
1270 yield data
File ~/autodl-tmp/envs/daily_learning/lib/python3.11/site-packages/urllib3/response.py:1110, in HTTPResponse.read(self, amt, decode_content, cache_content)
1107 if len(self._decoded_buffer) >= amt:
1108 return self._decoded_buffer.get(amt)
-> 1110 data = self._raw_read(amt)
1111 if not cache_content:
1112 self._uncached_read_occurred = True
File ~/autodl-tmp/envs/daily_learning/lib/python3.11/site-packages/urllib3/response.py:1022, in HTTPResponse._raw_read(self, amt, read1)
1019 fp_closed = getattr(self._fp, "closed", False)
1021 with self._error_catcher():
-> 1022 data = self._fp_read(amt, read1=read1) if not fp_closed else b""
1023 if amt is not None and amt != 0 and not data:
1024 # Platform-specific: Buggy versions of Python.
1025 # Close the connection when no data is returned
(...) 1030 # not properly close the connection in all cases. There is
1031 # no harm in redundantly calling close.
1032 self._fp.close()
File ~/autodl-tmp/envs/daily_learning/lib/python3.11/site-packages/urllib3/response.py:1005, in HTTPResponse._fp_read(self, amt, read1)
1002 return self._fp.read1(amt) if amt is not None else self._fp.read1()
1003 else:
1004 # StringIO doesn't like amt=None
-> 1005 return self._fp.read(amt) if amt is not None else self._fp.read()
File ~/autodl-tmp/envs/daily_learning/lib/python3.11/http/client.py:478, in HTTPResponse.read(self, amt)
475 if self.length is not None and amt > self.length:
476 # clip the read to the "end of response"
477 amt = self.length
--> 478 s = self.fp.read(amt)
479 if not s and amt:
480 # Ideally, we would raise IncompleteRead if the content-length
481 # wasn't satisfied, but it might break compatibility.
482 self._close_conn()
File ~/autodl-tmp/envs/daily_learning/lib/python3.11/socket.py:718, in SocketIO.readinto(self, b)
716 while True:
717 try:
--> 718 return self._sock.recv_into(b)
719 except timeout:
720 self._timeout_occurred = True
File ~/autodl-tmp/envs/daily_learning/lib/python3.11/ssl.py:1314, in SSLSocket.recv_into(self, buffer, nbytes, flags)
1310 if flags != 0:
1311 raise ValueError(
1312 "non-zero flags not allowed in calls to recv_into() on %s" %
1313 self.__class__)
-> 1314 return self.read(nbytes, buffer)
1315 else:
1316 return super().recv_into(buffer, nbytes, flags)
File ~/autodl-tmp/envs/daily_learning/lib/python3.11/ssl.py:1166, in SSLSocket.read(self, len, buffer)
1164 try:
1165 if buffer is not None:
-> 1166 return self._sslobj.read(len, buffer)
1167 else:
1168 return self._sslobj.read(len)
KeyboardInterrupt: 读入内存
代码 6
def read_voc_images(voc_dir, is_train=True):
"""读取VOC数据集中的图像和标签"""
txt_fname = os.path.join(voc_dir, 'ImageSets', 'Segmentation',
'train.txt' if is_train else 'val.txt')
mode = torchvision.io.image.ImageReadMode.RGB
with open(txt_fname, 'r') as f:
images = f.read().split()
features, labels = [], []
for i, fname in enumerate(images):
features.append(torchvision.io.read_image(os.path.join(
voc_dir, 'JPEGImages', f'{fname}.jpg')))
labels.append(torchvision.io.read_image(os.path.join( # label 也是图片，每个像素的值是类别索引
voc_dir, 'SegmentationClass' ,f'{fname}.png'), mode))
return features, labels
train_features, train_labels = read_voc_images(voc_dir, True)检查图片格式
代码 8
n = 5
imgs = train_features[:n] + train_labels[:n]
imgs = [img.permute(1,2,0) for img in imgs]
util.show_images(imgs, 2, n)标号映射
代码 10
VOC_COLORMAP = [[0, 0, 0], [128, 0, 0], [0, 128, 0], [128, 128, 0],
[0, 0, 128], [128, 0, 128], [0, 128, 128], [128, 128, 128],
[64, 0, 0], [192, 0, 0], [64, 128, 0], [192, 128, 0],
[64, 0, 128], [192, 0, 128], [64, 128, 128], [192, 128, 128],
[0, 64, 0], [128, 64, 0], [0, 192, 0], [128, 192, 0],
[0, 64, 128]]
VOC_CLASSES = ['background', 'aeroplane', 'bicycle', 'bird', 'boat',
'bottle', 'bus', 'car', 'cat', 'chair', 'cow',
'diningtable', 'dog', 'horse', 'motorbike', 'person',
'potted plant', 'sheep', 'sofa', 'train', 'tv/monitor']代码 11
def voc_colormap2label():
"""建立从RGB到类别索引的映射"""
def hashing(colormap):
return (colormap[0] * 256 + colormap[1]) * 256 + colormap[2]
colormap2label = torch.zeros(256 ** 3, dtype=torch.uint8) # 存储每个颜色对应的类别索引
for i, colormap in enumerate(VOC_COLORMAP):
colormap2label[hashing(colormap)] = i # 将每个颜色映射到对应的类别索引
return colormap2label
def voc_label_indices(colormap, colormap2label):
"""将彩色标签映射到类别索引"""
def hashing(colormap):
return (colormap[:, :, 0] * 256 + colormap[:, :, 1]) * 256 + colormap[:, :, 2]
colormap = colormap.permute(1, 2, 0).numpy().astype('int32') # 将通道维度移到最后
idx = hashing(colormap) # 将RGB映射到整数
return colormap2label[idx] # 返回类别索引例如
代码 13
y = voc_label_indices(train_labels[0], voc_colormap2label())
y[105:115, 130:140], VOC_CLASSES[1] # 显示某个区域类别索引(1)和类别名称(aeroplane)随机剪裁的图像增广，还需要同样剪裁标签
代码 15
def voc_rand_crop(feature, label, height, width):
"""随机裁剪图像和标签"""
# get_params 返回裁剪窗口的左上角坐标和高宽
rect = torchvision.transforms.RandomCrop.get_params(feature, (height, width))
# 通用的参数裁剪图像和标签
feature = torchvision.transforms.functional.crop(feature, *rect)
label = torchvision.transforms.functional.crop(label, *rect)
return feature, label代码 16
imgs = []
for _ in range(n):
imgs += voc_rand_crop(train_features[0], train_labels[0], 200, 300)
imgs = [img.permute(1,2,0) for img in imgs]
util.show_images(imgs[::2]+imgs[1::2], 2, n) # 显示裁剪后的图像和标签自定义数据集类
代码 18
class VOCSegDataset(torch.utils.data.Dataset):
"""VOC语义分割数据集"""
def __init__(self, is_train: bool, crop_size: tuple[int, int], voc_dir: str):
self.transform = torchvision.transforms.Normalize(mean=[0.485, 0.456, 0.406],
std=[0.229, 0.224, 0.225]) # ImageNet 标准
self.crop_size = crop_size
features, labels = read_voc_images(voc_dir, is_train)
self.features = [self.normalize_image(feat) for feat in self.filter(features)]
self.labels = self.filter(labels)
self.colormap2label = voc_colormap2label()
print(f'read {len(self.features)} examples')
def normalize_image(self, img):
"""标准化图像"""
return self.transform(img.float())
def filter(self, imgs):
"""过滤掉小于裁剪尺寸的图像"""
return [img for img in imgs if (img.shape[1] >= self.crop_size[0] and
img.shape[2] >= self.crop_size[1])]
def __getitem__(self, idx):
"""返回裁剪后的图像和标签"""
feature, label = voc_rand_crop(self.features[idx], self.labels[idx],
*self.crop_size)
return feature.float(), voc_label_indices(label, self.colormap2label)
def __len__(self):
"""返回数据集大小"""
return len(self.features)读取数据集
代码 20
crop_size = (320, 480)
voc_train = VOCSegDataset(is_train=True, crop_size=crop_size, voc_dir=voc_dir)
voc_test = VOCSegDataset(is_train=False, crop_size=crop_size, voc_dir=voc_dir)代码 21
train_iter = torch.utils.data.DataLoader(voc_train, batch_size=64, shuffle=True, drop_last=True, num_workers=0)
for X, Y in train_iter:
print(X.shape)
print(Y.shape)
break整合所有组件
代码 23
def load_data_voc(batch_size, crop_size):
"""加载VOC数据集"""
voc_dir = util.data.download_extract('voc2012', 'VOCdevkit/VOC2012')
train_iter = torch.utils.data.DataLoader(
VOCSegDataset(True, crop_size, voc_dir),
batch_size=batch_size,
shuffle=True,
drop_last=True,
num_workers=util.data.get_dataloader_workers())
test_iter = torch.utils.data.DataLoader(
VOCSegDataset(False, crop_size, voc_dir),
batch_size=batch_size,
drop_last=True,
num_workers=0)
return train_iter, test_iter