本范例我们使用 torchkeras来实现对 ultralytics中的YOLOv8实例分割模型进行自定义的训练,从而对气球进行检测和分割。
尽管ultralytics提供了非常便捷且一致的训练API,再使用torchkeras实现自定义训练逻辑似乎有些多此一举。
但ultralytics的源码结构相对复杂,不便于用户做个性化的控制和修改。
并且,torchkeras在可视化上会比ultralytics的原生训练代码优雅许多。
不信的话,我们对比看看就明白啦。
此外,本文的内容对同学们熟悉ultralytics这个库的代码结构也会有帮助。
😋😋公众号算法美食屋后台回复关键词:torchkeras,获取本文notebook源代码和balloon数据集下载链接。
〇,预测过程
from ultralytics import YOLO
model = YOLO('yolov8n-seg.pt')
import numpy as np
from PIL import Image
img_path = 'park.jpg'
try:
img = Image.open(img_path)
except Exception as err:
from torchkeras.data import get_example_image
img = get_example_image(img_path)
img.save(img_path)
#可以保存预测结果以及可视化图片
result = model.predict(source=img_path, save=True,
save_txt=True, conf = 0.3)
from pathlib import Path
import ultralytics
from ultralytics.data import utils
yaml_path = str(Path(ultralytics.__file__).parent/'cfg/datasets/coco128-seg.yaml')
class_names = utils.yaml_load(yaml_path)['names']
from torchkeras import plots
boxes = result[0].boxes.data
masks = result[0].masks.data
plots.plot_instance_segmentation(img,boxes,masks,class_names)
一,准备数据
训练yolo实例分割模型需要将数据集整理成yolo数据集格式。
yolo_dataset
├── images
│ ├── train
│ │ ├── train0.jpg
│ │ └── train1.jpg
│ ├── val
│ │ ├── val0.jpg
│ │ └── val1.jpg
│ └── test
│ ├── test0.jpg
│ └── test1.jpg
└── labels
├── train
│ ├── train0.txt
│ └── train1.txt
├── val
│ ├── val0.txt
│ └── val1.txt
└── test
├── test0.txt
└── test1.txt
对于实例分割模型,标签文件(如train0.txt)格式如下:
class_id point1(x,y) point2(x,y) point3(x,y) point4(x,y),...
8 0.417781 0.771355 0.440328 0.735397 0.467375 0.658995 0.440328 0.605047 0.387719 0.524159 0.378703 0.443248 0.333625 0.436519 0.371188 0.375841 0.335125 0.364603 0.350156 0.315164 0.320094 0.299439 0.320094 0.256752 0.327609 0.198318 0.357672 0.184836 0.39825 0.155607 0.498937 0.139883 0.470375 0.0724766 0.513953 0.117407 0.553031 0.083715 0.608641 0.115164 0.67175 0.173598 0.704812 0.184836 0.710828 0.211799 0.707828 0.232033 0.718344 0.263481 0.713828 0.308435 0.707828 0.348879 0.691297 0.398318 0.676266 0.416285 0.673266 0.476963 0.641703 0.420794 0.623672 0.510678 0.604125 0.566846 0.560547 0.623037 0.547016 0.676963 0.568063 0.730888 0.607141 0.771355 0.584594 0.811799 0.506453 0.829766 0.411766 0.793832 0.420781 0.769112
注意class_id从0开始, point坐标都是相对图片长宽的相对坐标。
1,转换成yolo格式
下面将原本是json格式的balloon数据集转换成yolo格式。
import os,json
from pathlib import Path
from shutil import copyfile
from PIL import Image
from tqdm import tqdm
root_path = './datasets/balloon-seg/'
# 1,构建目录
data_root = Path(root_path)
for tp in ('images','labels'):
for part in ('train','val'):
(data_root/tp/part).mkdir(parents=True, exist_ok=True)
# 2,复制图片文件
train_images = [str(x) for x in Path('balloon/train/').rglob('*.jpg')]
val_images = [str(x) for x in Path('balloon/val/').rglob('*.jpg')]
for src_file in tqdm(train_images):
name = os.path.basename(src_file)
dst_file = root_path+'images/train/'+name
copyfile(src_file,dst_file)
for src_file in tqdm(val_images):
name = os.path.basename(src_file)
dst_file = root_path+'images/val/'+name
copyfile(src_file,dst_file)
# 3,生成标签文件
train_dir = "balloon/train/"
val_dir = "balloon/val/"
train_json_file = train_dir + "via_region_data.json"
val_json_file = val_dir + "via_region_data.json"
def get_poly(anno):
anno = anno["shape_attributes"]
px = anno["all_points_x"]
py = anno["all_points_y"]
poly = [(x + 0.5, y + 0.5) for x, y in zip(px, py)]
poly = [p for x in poly for p in x]
#box = [np.min(px), np.min(py), np.max(px), np.max(py)]
return poly
def convert_yolo(size,poly):
width,height = size
dh,dw = 1.0/height,1.0/width
poly = [dw*x if i%2==0 else dh*x for i,x in enumerate(poly)]
return poly
def write_yolo_txt(label_path,catids,yolo_polys):
lines = [f"{cls} {' '.join(f'{x:.6f}' for x in poly)}\n"
for cls,poly in zip(catids,yolo_polys)]
with open(label_path, 'w') as fl:
fl.writelines(lines)
def write_labels(data_dir,part):
with open(data_dir + "via_region_data.json") as f:
info = json.load(f)
info_values = list(info.values())
for info_value in tqdm(info_values):
img_path = data_dir + info_value['filename']
anno_list = list(info_value['regions'].values())
polys = [get_poly(anno) for anno in anno_list]
catids = np.array([0 for x in polys])
size = Image.open(img_path).size
yolo_polys = [convert_yolo(size,poly) for poly in polys]
txt_path = data_root/'labels'/part/info_value['filename'].replace('.jpg','.txt')
write_yolo_txt(txt_path,catids,yolo_polys)
write_labels(train_dir,'train')
write_labels(val_dir,'val')
100%|██████████| 63/63 [00:00<00:00, 2057.52it/s]
100%|██████████| 13/13 [00:00<00:00, 1704.63it/s]
100%|██████████| 61/61 [00:00<00:00, 2880.77it/s]
100%|██████████| 13/13 [00:00<00:00, 2820.50it/s]
2,样本可视化
from PIL import Image,ImageDraw
import os
from pathlib import Path
from shutil import copyfile
from tqdm import tqdm
import numpy as np
def get_labels_polys(img_path,gt_path):
img = Image.open(img_path)
w,h = img.size
with open(gt_path, 'r') as fl:
lines = [x.rstrip() for x in fl.readlines()]
str_data = [x.split(' ') for x in lines]
relative_polys = [[float(x) for x in arr[1:]] for arr in str_data]
labels = [int(arr[0]) for arr in str_data]
polys = [ [x*w if i%2==0 else x*h for i,x in enumerate(arr)] for arr in relative_polys]
return labels,polys
def plot_polys(image,polys):
image_result = image.copy()
draw = ImageDraw.Draw(image_result)
for poly in polys:
draw.polygon(poly, fill ="cyan", outline ="red")
return image_result
from pathlib import Path
root_path = './datasets/balloon-seg/'
data_root = Path(root_path)
val_imgs = [str(x) for x in (data_root/'images'/'val').rglob("*.jpg") if 'checkpoint' not in str(x)]
img_path = val_imgs[2]
gt_path = img_path.replace('images','labels').replace('.jpg','.txt')
labels,polys = get_labels_polys(img_path,gt_path)
plot_polys(Image.open(img_path),polys)
3,数据集配置文件
仿照 ultralytics/data/yolo/data/datasets 中已有的一些yaml数据集配置文件,构建我们自己的数据集yaml文件。
import ultralytics
print(ultralytics.__file__)
%%writefile balloon-seg.yaml
# Train/val/test sets as 1) dir: path/to/imgs, 2) file: path/to/imgs.txt, or 3) list: [path/to/imgs1, path/to/imgs2, ..]
path: balloon-seg # dataset root dir
train: images/train # train images (relative to 'path') 128 images
val: images/val # val images (relative to 'path') 128 images
test: # test images (optional)
# Classes
names:
0: ballon
# Download script/URL (optional)
# download: https://ultralytics.com/assets/coco128.zip
Overwriting balloon-seg.yaml
4,数据管道
import torch
from torch.utils.data import DataLoader
from ultralytics.cfg import get_cfg
from ultralytics.utils import DEFAULT_CFG,yaml_load
from ultralytics.data.utils import check_cls_dataset, check_det_dataset
from ultralytics.data import build_yolo_dataset,build_dataloader
overrides = {'task':'segment',
'data':'balloon-seg.yaml',
'imgsz':640,
'workers':4
}
cfg = get_cfg(cfg = DEFAULT_CFG,overrides=overrides)
data_info = check_det_dataset(cfg.data)
ds_train = build_yolo_dataset(cfg,img_path=data_info['train'],batch=cfg.batch,
data = data_info,mode='train',rect=False,stride=32)
ds_val = build_yolo_dataset(cfg,img_path=data_info['val'],batch=cfg.batch,data = data_info,
mode='val',rect=False,stride=32)
dl_train = DataLoader(ds_train,batch_size = cfg.batch, num_workers = cfg.workers,
collate_fn = ds_train.collate_fn)
dl_val = DataLoader(ds_val,batch_size = cfg.batch, num_workers = cfg.workers,
collate_fn = ds_val.collate_fn)
二,定义模型
可以选择 yolov8n-seg,yolov8s-seg,yolov8l-seg,等官方定义好的模型结构,
也可以通过修改yaml模型配置文件来实现用户自定义的模型结构。
from ultralytics.nn.tasks import SegmentationModel
model = SegmentationModel(cfg = 'yolov8n-seg.yaml', ch=3, nc=1)
#weights = torch.hub.load_state_dict_from_url('https://github.com/ultralytics/assets/releases/download/v0.0.0/yolov8n.pt')
weights = torch.load('yolov8n-seg.pt')
model.load(weights['model'])
model.args = cfg
model.nc = data_info['nc'] # attach number of classes to model
model.names = data_info['names']
三,训练模型
我们分别演示使用ultralytics原生接口以及使用torchkeras的KerasModel两种接口训练模型的方法。
1, 使用ultralytics原生接口
from ultralytics import YOLO
model = YOLO('yolov8n-seg.pt')
results = model.train(data = 'balloon-seg.yaml', epochs = 50, workers=2)
#自动调参接口
#yolo_model = YOLO('yolov8n-seg.pt')
#yolo_model.tune(data='balloon-seg.yaml', epochs=10,
# iterations=300, optimizer='AdamW',
# plots=True, save=False, val=True)
2, torchkeras梦中情炉接口
#测试loss计算过程
for batch in dl_train:
break
for key,value in batch.items():
if isinstance(value,torch.Tensor):
batch[key] = batch[key].cuda()
model = model.cuda()
model.train();
batch['img'] = batch['img'].float()/255.0
loss,_ = model.forward(batch)
loss
tensor(89.2061)
from torchkeras import KerasModel
#我们需要修改StepRunner以适应Yolov8的数据集格式
class StepRunner:
def __init__(self, net, loss_fn, accelerator, stage = "train", metrics_dict = None,
optimizer = None, lr_scheduler = None
):
self.net,self.loss_fn,self.metrics_dict,self.stage = net,loss_fn,metrics_dict,stage
self.optimizer,self.lr_scheduler = optimizer,lr_scheduler
self.accelerator = accelerator
self.net.train()
def __call__(self, batch):
batch['img'] = batch['img'].float()/255
#loss
loss,_ = model.forward(batch)
#backward()
if self.optimizer is not None and self.stage=="train":
self.accelerator.backward(loss)
self.optimizer.step()
if self.lr_scheduler is not None:
self.lr_scheduler.step()
self.optimizer.zero_grad()
all_loss = self.accelerator.gather(loss).sum()
#losses
step_losses = {self.stage+"_loss":all_loss.item()}
#metrics
step_metrics = {}
if self.stage=="train":
if self.optimizer is not None:
step_metrics['lr'] = self.optimizer.state_dict()['param_groups'][0]['lr']
else:
step_metrics['lr'] = 0.0
return step_losses,step_metrics
KerasModel.StepRunner = StepRunner
optimizer = torch.optim.AdamW(model.parameters(), lr=1e-4)
keras_model = KerasModel(net = model,
loss_fn = None,
optimizer = optimizer)
keras_model.fit(train_data=dl_train,
val_data=dl_val,
epochs = 200,
ckpt_path='checkpoint',
patience=20,
monitor='val_loss',
mode='min',
mixed_precision='no',
plot= True
)
#关闭mosaic增强再训一次
ds_train.close_mosaic(cfg)
keras_model.from_scratch = False
keras_model.fit(train_data=dl_train,
val_data=dl_val,
epochs = 200,
ckpt_path='checkpoint',
patience=20,
monitor='val_loss',
mode='min',
mixed_precision='no',
plot= True
)
四,评估模型
为了便于评估 map等指标,我们将权重再次保存后,用ultralytics的原生YOLO接口进行加载后评估。
from ultralytics import YOLO
keras_model.load_ckpt('checkpoint')
save_dic = dict(model = keras_model.net, train_args =dict(cfg))
torch.save(save_dic, 'best_yolo.pt')
from ultralytics import YOLO
best_model = YOLO(model = 'best_yolo.pt')
metrics = best_model.val(data = cfg.data )
import pandas as pd
df = pd.DataFrame()
df['metric'] = metrics.keys
for i,c in best_model.names.items():
df[c] = metrics.class_result(i)
df
五,使用模型
from pathlib import Path
root_path = './datasets/balloon-seg/'
data_root = Path(root_path)
val_imgs = [str(x) for x in (data_root/'images'/'train').rglob("*.jpg") if 'checkpoint' not in str(x)]
img_path = val_imgs[10]
Image.open(img_path)
result = best_model(img_path,conf=0.1)
from torchkeras import plots
masks = result[0].masks.data
boxes = result[0].boxes.data
plots.plot_instance_segmentation(Image.open(img_path),boxes,masks,
class_names = ['balloon'],min_score=0.0)
六, 导出onnx格式
success = best_model.export(format='onnx',dynamic=True)
model = YOLO('best_yolo.onnx',task='segment')
img_path = val_imgs[5]
result = model.predict(img_path,save_txt=True);
from torchkeras import plots
masks = result[0].masks.data
boxes = result[0].boxes.data
plots.plot_instance_segmentation(Image.open(img_path),boxes,masks,
class_names = ['balloon'],min_score=0.0)
import torch
from torch import nn
import onnxruntime
from PIL import Image
from ultralytics.models.yolo.segment.predict import SegmentationPredictor
from ultralytics.yolo.utils.torch_utils import select_device
class OnnxModel(nn.Module):
def __init__(self,weights,
device=torch.device('cpu'),
dnn=False,
data=None,
fp16=False,
fuse=True,
verbose=True):
super().__init__()
w = weights
nn_module = False
onnx = True
pt, jit, xml, engine, coreml, saved_model, pb, tflite, edgetpu, tfjs, paddle, triton = [False]*12
nhwc = False
stride = 32
model, metadata = None, None
cuda = torch.cuda.is_available() and device.type != 'cpu'
names = ['circle']
self.__dict__.update(locals()) # assign all variables to self
providers = ['CUDAExecutionProvider', 'CPUExecutionProvider'] if cuda else [
'CPUExecutionProvider']
self.session = onnxruntime.InferenceSession(w, providers=providers)
self.output_names = [x.name for x in self.session.get_outputs()]
def forward(self, im, augment=False, visualize=False):
im = im.cpu().numpy()
y = self.session.run(self.output_names, {self.session.get_inputs()[0].name: im})
if isinstance(y, (list, tuple)):
return self.from_numpy(y[0]) if len(y) == 1 else [self.from_numpy(x) for x in y]
else:
return self.from_numpy(y)
def from_numpy(self, x):
return torch.tensor(x).to(self.device) if isinstance(x, np.ndarray) else x
def warmup(self,imgsz=(1, 3, 640, 640)):
im = torch.empty(*imgsz, dtype=torch.half if self.fp16 else torch.float,
device=self.device) # input
for _ in range(2 if self.jit else 1): #
self.forward(im) # warmup
class Predictor(SegmentationPredictor):
def setup_model(self, model, verbose=True):
device = select_device(self.args.device, verbose=verbose)
model = model or self.args.model
self.args.half &= device.type != 'cpu'
self.model = OnnxModel(model,
device=device,
dnn=self.args.dnn,
data=self.args.data,
fp16=self.args.half,
verbose=verbose)
self.device = device
self.model.eval()
args = dict(model='best_yolo.onnx')
predictor = Predictor(overrides=args)
result = predictor(source = Image.open(img_path) )
from torchkeras import plots
masks = result[0].masks.data
boxes = result[0].boxes.data
plots.plot_instance_segmentation(Image.open(img_path),boxes,masks,
class_names = ['balloon'],min_score=0.0)
公众号算法美食屋后台回复关键词:torchkeras,获取本文notebook代码和balloon数据集,以及更多有趣范例~