基于百度AIStudio飞桨paddleRS-develop版道路模型开发训练
参考地址:https://aistudio.baidu.com/projectdetail/8271882
基于python35+paddle120+env环境
预测可视化结果:
(一)安装环境:
先上传本地下载的源代码PaddleRS-develop.zip
解压PaddleRS-develop.zip到目录PaddleRS
然后分别执行下面安装命令!pip install
python
!unzip -q /home/aistudio/data/data191076/PaddleRS-develop.zip && mv PaddleRS-develop PaddleRS
!pip install matplotlib==3.4 scikit-image pycocotools -t /home/aistudio/external-libraries
!pip install opencv-contrib-python -t /home/aistudio/external-libraries
!pip install -r PaddleRS/requirements.txt -t /home/aistudio/external-libraries
!pip install -e PaddleRS/ -t /home/aistudio/external-libraries
!pip install paddleslim==2.6.0 -t /home/aistudio/external-libraries
添加环境组件
python
# 因为`sys.path`可能没有及时更新,这里选择手动更新
import sys
sys.path.append('/home/aistudio/external-libraries')
sys.path.append('/home/aistudio/PaddleRS')
(二)数据预处理tran_dataPre.py
python
%run tran_dataPre.py
(三)开始模型训练
python
%run trans.py
(四) tran_dataPre.py内容如下所示:
python
#先解压数据集
#!unzip -oq -d /home/aistudio/massroad /home/aistudio/data/data56961/mass_road.zip
# 划分训练集/验证集/测试集,并生成文件名列表
import random
import os.path as osp
from os import listdir
import cv2
# 随机数生成器种子
RNG_SEED = 56961
# 调节此参数控制训练集数据的占比
TRAIN_RATIO = 0.9
# 数据集路径
DATA_DIR = '/home/aistudio/massroad'
# 分割类别
CLASSES = (
'background',
'road',
)
def write_rel_paths(phase, names, out_dir, prefix):
"""将文件相对路径存储在txt格式文件中"""
with open(osp.join(out_dir, phase+'.txt'), 'w') as f:
for name in names:
f.write(
' '.join([
osp.join(prefix, 'input', name),
osp.join(prefix, 'output', name)
])
)
f.write('\n')
random.seed(RNG_SEED)
train_prefix = osp.join('road_segmentation_ideal', 'training')
test_prefix = osp.join('road_segmentation_ideal', 'testing')
train_names = listdir(osp.join(DATA_DIR, train_prefix, 'output'))
train_names = list(filter(lambda n: n.endswith('.png'), train_names))
test_names = listdir(osp.join(DATA_DIR, test_prefix, 'output'))
test_names = list(filter(lambda n: n.endswith('.png'), test_names))
# 对文件名进行排序,以确保多次运行结果一致
train_names.sort()
test_names.sort()
random.shuffle(train_names)
len_train = int(len(train_names)*TRAIN_RATIO)
write_rel_paths('train', train_names[:len_train], DATA_DIR, train_prefix)
write_rel_paths('val', train_names[len_train:], DATA_DIR, train_prefix)
write_rel_paths('test', test_names, DATA_DIR, test_prefix)
# 写入类别信息
with open(osp.join(DATA_DIR, 'labels.txt'), 'w') as f:
for cls in CLASSES:
f.write(cls+'\n')
print("数据集划分已完成。")
# 将GT中的255改写为1,便于训练
import os.path as osp
from glob import glob
import cv2
from tqdm import tqdm
# 数据集路径
# DATA_DIR = '/home/aistudio/massroad'
train_prefix = osp.join('road_segmentation_ideal', 'training')
test_prefix = osp.join('road_segmentation_ideal', 'testing')
train_paths = glob(osp.join(DATA_DIR, train_prefix, 'output', '*.png'))
test_paths = glob(osp.join(DATA_DIR, test_prefix, 'output', '*.png'))
for path in tqdm(train_paths+test_paths):
im = cv2.imread(path, cv2.IMREAD_GRAYSCALE)
im[im>0] = 1
# 原地改写
cv2.imwrite(path, im)
(五) trans.py内容如下所示:
python
# 导入需要用到的库
import random
import os.path as osp
import cv2
import numpy as np
import paddle
import paddlers as pdrs
from paddlers import transforms as T
from matplotlib import pyplot as plt
from PIL import Image
import sys
sys.path.append('/home/aistudio/external-libraries')
sys.path.append('/home/aistudio/PaddleRS')
# 定义全局变量
# 随机种子
SEED = 56961
# 数据集存放目录
DATA_DIR = '/home/aistudio/massroad/'
# 训练集`file_list`文件路径
TRAIN_FILE_LIST_PATH = '/home/aistudio/massroad/train.txt'
# 验证集`file_list`文件路径
VAL_FILE_LIST_PATH = '/home/aistudio/massroad/val.txt'
# 测试集`file_list`文件路径
TEST_FILE_LIST_PATH = '/home/aistudio/massroad/test.txt'
# 数据集类别信息文件路径
LABEL_LIST_PATH = '/home/aistudio/massroad/labels.txt'
# 实验目录,保存输出的模型权重和结果
EXP_DIR = '/home/aistudio/exp/'
# 固定随机种子,尽可能使实验结果可复现
random.seed(SEED)
np.random.seed(SEED)
paddle.seed(SEED)
# 构建数据集
# 定义训练和验证时使用的数据变换(数据增强、预处理等)
train_transforms = T.Compose([
T.DecodeImg(),
# 随机裁剪
T.RandomCrop(crop_size=512),
# 以50%的概率实施随机水平翻转
T.RandomHorizontalFlip(prob=0.5),
# 以50%的概率实施随机垂直翻转
T.RandomVerticalFlip(prob=0.5),
# 将数据归一化到[-1,1]
T.Normalize(
mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5]),
T.ArrangeSegmenter('train')
])
eval_transforms = T.Compose([
T.DecodeImg(),
T.Resize(target_size=1500),
# 验证阶段与训练阶段的数据归一化方式必须相同
T.Normalize(
mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5]),
T.ArrangeSegmenter('eval')
])
# 分别构建训练和验证所用的数据集
train_dataset = pdrs.datasets.SegDataset(
data_dir=DATA_DIR,
file_list=TRAIN_FILE_LIST_PATH,
label_list=LABEL_LIST_PATH,
transforms=train_transforms,
num_workers=4,
shuffle=True
)
val_dataset = pdrs.datasets.SegDataset(
data_dir=DATA_DIR,
file_list=VAL_FILE_LIST_PATH,
label_list=LABEL_LIST_PATH,
transforms=eval_transforms,
num_workers=0,
shuffle=False
)
# 构建DeepLab V3+模型,使用ResNet-50作为backbone
model = pdrs.tasks.seg.DeepLabV3P(
in_channels=3,
num_classes=len(train_dataset.labels),
backbone='ResNet50_vd'
)
model.initialize_net(
pretrain_weights='CITYSCAPES',
save_dir=osp.join(EXP_DIR, 'pretrain'),
resume_checkpoint=None,
is_backbone_weights=False
)
# 构建优化器
optimizer = paddle.optimizer.Adam(
learning_rate=0.001,
parameters=model.net.parameters()
)
# 执行模型训练
model.train(
num_epochs=100,
train_dataset=train_dataset,
train_batch_size=8,
eval_dataset=val_dataset,
optimizer=optimizer,
save_interval_epochs=10,
# 每多少次迭代记录一次日志
log_interval_steps=30,
save_dir=EXP_DIR,
# 是否使用early stopping策略,当精度不再改善时提前终止训练
early_stop=False,
# 是否启用VisualDL日志功能
use_vdl=True,
# 指定从某个检查点继续训练
resume_checkpoint=None
)
(六)训练生成过程信息
python
Output exceeds the size limit. Open the full output data in a text editor
2024-09-05 14:16:51 [INFO] Loading pretrained model from /home/aistudio/exp/pretrain/model.pdparams
2024-09-05 14:16:53 [WARNING] [SKIP] Shape of parameters head.decoder.conv.weight do not match. (pretrained: [19, 256, 1, 1] vs actual: [2, 256, 1, 1])
2024-09-05 14:16:53 [WARNING] [SKIP] Shape of parameters head.decoder.conv.bias do not match. (pretrained: [19] vs actual: [2])
2024-09-05 14:16:53 [INFO] There are 358/360 variables loaded into DeepLabV3P.
2024-09-05 14:17:46 [INFO] [TRAIN] Epoch=1/100, Step=30/90, loss=0.133503, lr=0.001000, time_each_step=1.77s, eta=4:24:32
2024-09-05 14:18:25 [INFO] [TRAIN] Epoch=1/100, Step=60/90, loss=0.181917, lr=0.001000, time_each_step=1.31s, eta=3:14:53
2024-09-05 14:19:02 [INFO] [TRAIN] Epoch=1/100, Step=90/90, loss=0.112567, lr=0.001000, time_each_step=1.22s, eta=3:2:6
2024-09-05 14:19:03 [INFO] [TRAIN] Epoch 1 finished, loss=0.15933047160506247 .
2024-09-05 14:19:44 [INFO] [TRAIN] Epoch=2/100, Step=30/90, loss=0.141528, lr=0.001000, time_each_step=1.36s, eta=3:22:2
2024-09-05 14:20:20 [INFO] [TRAIN] Epoch=2/100, Step=60/90, loss=0.165187, lr=0.001000, time_each_step=1.22s, eta=3:0:42
2024-09-05 14:20:57 [INFO] [TRAIN] Epoch=2/100, Step=90/90, loss=0.145009, lr=0.001000, time_each_step=1.22s, eta=2:59:1
2024-09-05 14:20:58 [INFO] [TRAIN] Epoch 2 finished, loss=0.1168842613697052 .
2024-09-05 14:21:39 [INFO] [TRAIN] Epoch=3/100, Step=30/90, loss=0.126603, lr=0.001000, time_each_step=1.38s, eta=3:22:13
2024-09-05 14:22:16 [INFO] [TRAIN] Epoch=3/100, Step=60/90, loss=0.117296, lr=0.001000, time_each_step=1.22s, eta=2:58:14
2024-09-05 14:22:53 [INFO] [TRAIN] Epoch=3/100, Step=90/90, loss=0.072859, lr=0.001000, time_each_step=1.23s, eta=2:58:46
2024-09-05 14:22:53 [INFO] [TRAIN] Epoch 3 finished, loss=0.10787189056475957 .
2024-09-05 14:23:34 [INFO] [TRAIN] Epoch=4/100, Step=30/90, loss=0.081685, lr=0.001000, time_each_step=1.37s, eta=3:18:39
2024-09-05 14:24:11 [INFO] [TRAIN] Epoch=4/100, Step=60/90, loss=0.087735, lr=0.001000, time_each_step=1.23s, eta=2:57:28
2024-09-05 14:24:48 [INFO] [TRAIN] Epoch=4/100, Step=90/90, loss=0.084795, lr=0.001000, time_each_step=1.22s, eta=2:55:44
2024-09-05 14:24:49 [INFO] [TRAIN] Epoch 4 finished, loss=0.10476481277081702 .
2024-09-05 14:25:30 [INFO] [TRAIN] Epoch=5/100, Step=30/90, loss=0.098625, lr=0.001000, time_each_step=1.37s, eta=3:16:59
2024-09-05 14:26:07 [INFO] [TRAIN] Epoch=5/100, Step=60/90, loss=0.078188, lr=0.001000, time_each_step=1.24s, eta=2:57:12
2024-09-05 14:26:43 [INFO] [TRAIN] Epoch=5/100, Step=90/90, loss=0.098015, lr=0.001000, time_each_step=1.21s, eta=2:52:11
2024-09-05 14:26:44 [INFO] [TRAIN] Epoch 5 finished, loss=0.10311256903741095 .
2024-09-05 14:27:25 [INFO] [TRAIN] Epoch=6/100, Step=30/90, loss=0.109136, lr=0.001000, time_each_step=1.38s, eta=3:16:8
...
2024-09-05 15:39:38 [INFO] Start to evaluate (total_samples=81, total_steps=81)...
2024-09-05 15:40:14 [INFO] [EVAL] Finished, Epoch=40, miou=0.716638, category_iou=[0.96831487 0.46496069], oacc=0.969164, category_acc=[0.97447995 0.81316509], kappa=0.619485, category_F1-score=[0.98390241 0.63477565] .
2024-09-05 15:40:14 [INFO] Current evaluated best model on eval_dataset is epoch_10, miou=0.7255623401044613
2024-09-05 15:40:18 [INFO] Model saved in /home/aistudio/exp/epoch_40.
(七) 测试集预测结果:
python
# 构建测试集
test_dataset = pdrs.datasets.SegDataset(
data_dir=DATA_DIR,
file_list=TEST_FILE_LIST_PATH,
label_list=LABEL_LIST_PATH,
transforms=eval_transforms,
num_workers=0,
shuffle=False
)
# 为模型加载历史最佳权重
state_dict = paddle.load(osp.join(EXP_DIR, 'best_model/model.pdparams'))
model.net.set_state_dict(state_dict)
# 执行测试
test_result = model.evaluate(test_dataset)
print(
"测试集上指标:IoU为{:.2f},Acc为{:.2f},Kappa系数为{:.2f}, F1为{:.2f}".format(
test_result['category_iou'][1],
test_result['category_acc'][1],
test_result['kappa'],
test_result['category_F1-score'][1]
)
)
python
2024-09-05 20:07:40 [INFO] 13 samples in file /home/aistudio/massroad/test.txt
2024-09-05 20:07:41 [INFO] Start to evaluate (total_samples=13, total_steps=13)...
测试集上指标:IoU为0.47,Acc为0.82,Kappa系数为0.62, F1为0.64
(八)预测结果可视化情况:
python
# 预测结果可视化
# 重复运行本单元可以查看不同结果
def read_image(path):
im = cv2.imread(path)
return im[...,::-1]
def show_images_in_row(ims, fig, title='', quantize=False):
n = len(ims)
fig.suptitle(title)
axs = fig.subplots(nrows=1, ncols=n)
for idx, (im, ax) in enumerate(zip(ims, axs)):
# 去掉刻度线和边框
ax.spines['top'].set_visible(False)
ax.spines['right'].set_visible(False)
ax.spines['bottom'].set_visible(False)
ax.spines['left'].set_visible(False)
ax.get_xaxis().set_ticks([])
ax.get_yaxis().set_ticks([])
if isinstance(im, str):
im = read_image(im)
if quantize:
im = (im*255).astype('uint8')
if im.ndim == 2:
im = np.tile(im[...,np.newaxis], [1,1,3])
ax.imshow(im)
# 需要展示的样本个数
num_imgs_to_show = 4
# 随机抽取样本
chosen_indices = random.choices(range(len(test_dataset)), k=num_imgs_to_show)
# 参考 https://stackoverflow.com/a/68209152
fig = plt.figure(constrained_layout=True)
fig.suptitle("Test Results")
subfigs = fig.subfigures(nrows=3, ncols=1)
# 读取输入影像并显示
im_paths = [test_dataset.file_list[idx]['image'] for idx in chosen_indices]
show_images_in_row(im_paths, subfigs[0], title='Image')
# 获取模型预测输出
with paddle.no_grad():
model.net.eval()
preds = []
for idx in chosen_indices:
input, mask = test_dataset[idx]
input = paddle.to_tensor(input["image"]).unsqueeze(0)
logits, *_ = model.net(input)
pred = paddle.argmax(logits[0], axis=0)
preds.append(pred.numpy())
show_images_in_row(preds, subfigs[1], title='Pred', quantize=True)
# 读取真值标签并显示
im_paths = [test_dataset.file_list[idx]['mask'] for idx in chosen_indices]
show_images_in_row(im_paths, subfigs[2], title='GT', quantize=True)
# 渲染结果
fig.canvas.draw()
Image.frombytes('RGB', fig.canvas.get_width_height(), fig.canvas.tostring_rgb())
(九) 导出静态模型
训练后保存的模型为动态模型,布署发布模型为静态模型,因此需要导出操作
python
import matplotlib.pyplot as plt
import random
import cv2
import numpy as np
import paddle
import paddlers as pdrs
from PIL import Image
import os
from paddlers.tasks import load_model
model_path = './exp/best_model'
img_14="i:/cwgis_ai/cup/mass_road/road_segmentation_ideal/testing/input/img-14.png"
img_10="i:/cwgis_ai/cup/mass_road/road_segmentation_ideal/testing/input/img-10.png"
#save_dir="./models/road_infer_model_100"
save_dir="./models/road_infer_model_100_custom"
# export model OK
# Set environment variables
os.environ['PADDLEX_EXPORT_STAGE'] = 'True'
os.environ['PADDLESEG_EXPORT_STAGE'] = 'True'
# Load model from directory
model = load_model(model_path)
#fixed_input_shape = None
#fixed_input_shape = [1500,1500]
fixed_input_shape = [17761,25006] #[w,h]
# Do dynamic-to-static cast 动态到静态的转换
# XXX: Invoke a protected (single underscore) method outside of subclasses.
model.export_inference_model(save_dir, fixed_input_shape)
(十) 预测单张图片代码
python
import matplotlib.pyplot as plt
import random
import cv2
import numpy as np
import paddle
import paddlers as pdrs
from PIL import Image
import os
from paddlers.tasks import load_model
# 因为`sys.path`可能没有及时更新,这里选择手动更新
import sys
sys.path.append('/home/aistudio/external-libraries')
sys.path.append('/home/aistudio/PaddleRS')
img_14="./massroad/road_segmentation_ideal/testing/input/img-14.png"
img_10="./massroad/road_segmentation_ideal/testing/input/img-10.png"
img_5="./massroad/road_segmentation_ideal/testing/input/img-5.png"
customImg="./customImage/DeepLearning_Image.png" #file tif to png
#model_dir="./models/road_infer_model_100"
#model_dir="./models/road_infer_model_100_None"
model_dir="./models/road_infer_model_100_custom"
#model = pdrs.deploy.Predictor(model_dir)
model = pdrs.deploy.Predictor(model_dir,use_gpu=True)
# 读取输入影像并显示
im_paths = [customImg]
im_lis = []
for name in im_paths:
print(name)
img = cv2.imread(name)
print(img.shape)
#img = paddle.to_tensor(img) #.unsqueeze(0) #标量输入
im_lis.append(img)
# 获取模型预测输出img_file=img_10
preds = []
results = model.predict(im_lis)
#print(results)
label_map=results[0]["label_map"]
#print(label_map)
label_map[label_map>0] = 255
cv2.imwrite('./outImage/label_map_custom.png', label_map)
score_map=results[0]["score_map"]
#cv2.imwrite('./outImage/score_map.png', score_map[0])
print(score_map)
print("预测完成")
本blog地址:https://blog.csdn.net/hsg77