开发者实战 | 英特尔锐炫™ 显卡+ oneAPI 和 OpenVINO™ 实现英特尔® 视频 AI 计算盒训推一体-上篇

wget https://learnopencv.s3.us-west-2.amazonaws.com/pothole_dataset.zipunzip -q pothole_dataset.zip

git clone https://github.com/WongKinYiu/yolov7.gitcd yolov7pip install -r requirements.txt

diff --git a/train.py b/train.py  index 86c7e48..21815b7 100644  --- a/train.py  +++ b/train.py  @@ -286,7 +286,9 @@ def train(hyp, opt, device, tb_writer=None):       model.nc       model.hyp = hyp  # attach hyperparameters to model       model.gr = 1.0  # iou loss ratio (obj_loss = 1.0 or iou)  -    model.class_weights = labels_to_class_weights(dataset.labels, nc).to(device) * nc  # attach class weights  +    #model.class_weights = labels_to_class_weights(dataset.labels, nc).to(device) * nc  # attach class weights  +    model.class_weights = labels_to_class_weights(dataset.labels, nc) * nc  # attach class weights  +    model.class_weights = model.class_weights.to(device)       model.names = names              # Start training  diff --git a/utils/autoanchor.py b/utils/autoanchor.py  index f491032..5bb3f1a 100644  --- a/utils/autoanchor.py  +++ b/utils/autoanchor.py  @@ -49,7 +49,10 @@ def check_anchors(dataset, model, thr=4.0, imgsz=640):               print(f'{prefix}ERROR: {e}')           new_bpr           if new_bpr > bpr:  # replace anchors  -            anchors = torch.tensor(anchors, device=m.anchors.device).type_as(m.anchors)  +            print("*** m.anchars.device : {}".format(m.anchors.device))  +            #anchors = torch.tensor(anchors, device=m.anchors.device).type_as(m.anchors)  +            anchors = torch.tensor(anchors).type_as(m.anchors)  +            anchors = anchors.to(m.anchors.device)               m.anchor_grid[:] = anchors.clone().view_as(m.anchor_grid)  # for inference               check_anchor_order(m)               m.anchors[:] = anchors.clone().view_as(m.anchors) / m.stride.to(m.anchors.device).view(-1, 1, 1)  # loss  diff --git a/utils/torch_utils.py b/utils/torch_utils.py  index 1e631b5..5a93bd7 100644  --- a/utils/torch_utils.py  +++ b/utils/torch_utils.py  @@ -17,6 +17,8 @@ import torch.nn as nn   import torch.nn.functional as F   import torchvision
    +import intel_extension_for_pytorch as ipex  +   try:       import thop  # for FLOPS computation   except ImportError:  @@ -64,13 +66,17 @@ def select_device(device='', batch_size=None):       # device       s = f'YOLOR 🚀 {git_describe() or date_modified()} torch {torch.__version__} '  # string       cpu = device.lower() == 'cpu'  +    xpu = device.lower() == 'xpu'       if cpu:           os.environ['CUDA_VISIBLE_DEVICES'] = '-1'  # force torch.cuda.is_available() = False  +    elif xpu:  # non-cpu device requested  +        os.environ['CUDA_VISIBLE_DEVICES'] = '-1'  # force torch.cuda.is_available() = False  +        assert torch.xpu.is_available(), f'XPU unavailable, invalid device {device} requested'  # check availability       elif device:  # non-cpu device requested           os.environ['CUDA_VISIBLE_DEVICES'] = device  # set environment variable           assert torch.cuda.is_available(), f'CUDA unavailable, invalid device {device} requested'  # check availability
    -    cuda = not cpu and torch.cuda.is_available()  +    cuda = not cpu and not xpu and torch.cuda.is_available()       if cuda:           n = torch.cuda.device_count()           if n > 1 and batch_size:  # check that batch_size is compatible with device_count  @@ -79,11 +85,19 @@ def select_device(device='', batch_size=None):           for i, d in enumerate(device.split(',') if device else range(n)):               p               s += f"{'' if i == 0 else space}CUDA:{d} ({p.name}, {p.total_memory / 1024 ** 2}MB)\n"  # bytes to MB  +            s += f"{'' if i == 0 else space}CUDA:{d} ({p.name}, {p.total_memory / 1024 ** 2}MB)\n"  # bytes to MB  +    elif xpu:  +        s += 'XPU\n'       else:           s += 'CPU\n'                  logger.info(s.encode().decode('ascii', 'ignore') if platform.system() == 'Windows' else s)  # emoji-safe    -    return torch.device('cuda:0' if cuda else 'cpu')  +    if cuda:  +        return torch.device('cuda:0')  +    elif xpu:  +        return torch.device('xpu')  +    else:  +        return torch.device('cpu')
     def time_synchronized():  @@ -371,4 +385,4 @@ class TracedModel(nn.Module):       def forward(self, x, augment=False, profile=False):           out           out = self.detect_layer(out)  -        return out  \ No newline at end of file  +        return out

patch -p1 < yolov7_xpu.patch

## content of data/pothole.yaml ##train: ../pothole_dataset/images/trainval: ../pothole_dataset/images/validtest: ../pothole_dataset/images/test   # Classesnc: 1  # number of classesnames: ['pothole']  # class names## content of data/pothole.yaml ##

## modify "nc: 80" to "nc: 1" in cfg/training/yolov7_pothole-tiny.yaml ##

cp cfg/training/yolov7-tiny.yaml cfg/training/yolov7_pothole-tiny.yaml

wget https://github.com/WongKinYiu/yolov7/releases/download/v0.1/yolov7-tiny.pt

python train.py --epochs 50 --workers 4 --device xpu --batch-size 32 --data data/pothole.yaml --img 640 640 --cfg cfg/training/yolov7_pothole-tiny.yaml --weights 'yolov7-tiny.pt' --name yolov7_tiny_pothole_fixed_res --hyp data/hyp.scratch.tiny.yaml