OAK相机如何将 MMYOLO 内模型转换成blob格式？
编辑：OAK中国
首发：oakchina.cn
喜欢的话，请多多⭐️✍
内容可能会不定期更新，官网内容都是最新的，请查看首发地址链接。
▌前言
Hello，大家好，这里是OAK中国，我是助手君。
最近咱社群里有几个朋友在将yolox转换成blob的过程有点不清楚，所以我就写了这篇博客。（请夸我贴心！咱的原则：合理要求，有求必应！）
1.其他Yolo转换及使用教程请参考
2.检测类的yolo模型建议使用在线转换（地址），如果在线转换不成功，你再根据本教程来做本地转换。
▌.pt 转换为 .onnx
使用下列脚本 (将脚本放到 MMYOLO 根目录中) 将 pytorch 模型转换为 onnx 模型，若已安装 openvino_dev，则可进一步转换为 OpenVINO 模型：
示例用法：
# python export_onnx.py config checkpoint --img-size 320
# 例如
python export_yolo.py \
ppyoloe_plus_s_fast_8xb8-80e_coco.py \
ppyoloe_plus_s_fast_8xb8-80e_coco_20230101_154052-9fee7619.pth \
--work-dir work_dir --img-size 320
python export_onnx.py -m work_dir/ppyoloe_plus_s_fast_8xb8-80e_coco.onnx -v ppyoloe
export_yolo.py :
usage: export_yolo.py [-h] [--work-dir WORK_DIR]
[-imgsz IMG_SIZE [IMG_SIZE ...]] [-op OPSET]
config checkpoint
positional arguments:
config Config file
checkpoint Checkpoint file
options:
-h, --help show this help message and exit
--work-dir WORK_DIR Path to save export model (default: work_dir)
-imgsz IMG_SIZE [IMG_SIZE ...], --img-size IMG_SIZE [IMG_SIZE ...]
Image size of height and width (default: [640, 640])
-op OPSET, --opset OPSET
ONNX opset version (default: 12)
# coding=utf-8
import argparse
import json
import warnings
from io import BytesIO
from argparse import ArgumentDefaultsHelpFormatter
from pathlib import Path
import onnx
import torch
import torch.nn as nn
from mmdet.apis import init_detector
from mmdet.models.backbones.csp_darknet import CSPLayer, Focus
from mmengine.utils.path import mkdir_or_exist
from rich import print
from mmyolo.models import RepVGGBlock
from mmyolo.models.layers import CSPLayerWithTwoConv
warnings.filterwarnings(action="ignore", category=torch.jit.TracerWarning)
warnings.filterwarnings(action="ignore", category=torch.jit.ScriptWarning)
warnings.filterwarnings(action="ignore", category=UserWarning)
warnings.filterwarnings(action="ignore", category=FutureWarning)
warnings.filterwarnings(action="ignore", category=ResourceWarning)
def parse_args():
parser = argparse.ArgumentParser(formatter_class=ArgumentDefaultsHelpFormatter)
parser.add_argument("config", type=Path, help="Config file")
parser.add_argument("checkpoint", type=Path, help="Checkpoint file")
parser.add_argument(
"--work-dir",
default=Path("./work_dir"),
type=Path,
help="Path to save export model",
)
parser.add_argument(
"-imgsz",
"--img-size",
nargs="+",
type=int,
default=[640, 640],
help="Image size of height and width",
)
parser.add_argument("-op", "--opset", type=int, default=12, help="ONNX opset version")
args = parser.parse_args()
args.img_size *= 2 if len(args.img_size) == 1 else 1
args.work_dir = args.work_dir.resolve().absolute()
print(args)
return args
def build_model_from_cfg(config_path, checkpoint_path, device):
model = init_detector(config_path, checkpoint_path, device=device)
model.eval()
return model
class DeployFocus(nn.Module):
def __init__(self, orin_Focus: nn.Module):
super().__init__()
self.__dict__.update(orin_Focus.__dict__)
def forward(self, x):
batch_size, channel, height, width = x.shape
x = x.reshape(batch_size, channel, -1, 2, width)
x = x.reshape(batch_size, channel, x.shape[2], 2, -1, 2)
half_h = x.shape[2]
half_w = x.shape[4]
x = x.permute(0, 5, 3, 1, 2, 4)
x = x.reshape(batch_size, channel * 4, half_h, half_w)
return self.conv(x)
class DeployC2f(nn.Module):
def __init__(self, *args, **kwargs):
super().__init__()
def forward(self, x):
x_main = self.main_conv(x)
x_main = [x_main, x_main[:, self.mid_channels :, ...]]
x_main.extend(blocks(x_main[-1]) for blocks in self.blocks)
x_main.pop(1)
return self.final_conv(torch.cat(x_main, 1))
class HardSigmoid(nn.Module):
"""Hard Sigmoid Module"""
def __init__(self, bias=1.0, divisor=2.0, min_value=0.0, max_value=1.0):
super(HardSigmoid, self).__init__()
assert divisor != 0, "divisor is not allowed to be equal to zero"
self.bias = bias
self.divisor = divisor
self.min_value = min_value
self.max_value = max_value
def forward(self, x):
"""forward"""
x = (x + self.bias) / self.divisor
return x.clamp_(self.min_value, self.max_value)
def switch_deploy(baseModel):
for layer in baseModel.modules():
if isinstance(layer, RepVGGBlock):
layer.switch_to_deploy()
elif isinstance(layer, Focus):
baseModel.backbone.stem = DeployFocus(layer)
elif isinstance(layer, CSPLayerWithTwoConv):
setattr(layer, "__class__", DeployC2f)
elif isinstance(layer, CSPLayer):
if hasattr(layer, "attention"):
if isinstance(layer.attention.act, nn.Hardsigmoid):
layer.attention.act = HardSigmoid()
def main():
args = parse_args()
mkdir_or_exist(args.work_dir)
device = "cpu" # 'cuda:0'
output_names = None
baseModel = build_model_from_cfg(args.config.as_posix(), args.checkpoint.as_posix(), device)
switch_deploy(baseModel)
baseModel.eval()
fake_input = torch.randn(1, 3, *args.img_size).to(device)
# dry run
baseModel(fake_input)
save_onnx_path = args.work_dir.joinpath(args.config.with_suffix(".onnx").name)
# export onnx
with BytesIO() as f:
torch.onnx.export(
baseModel,
fake_input,
f,
input_names=["images"],
output_names=output_names,
opset_version=args.opset,
)
f.seek(0)
onnx_model = onnx.load(f)
onnx.checker.check_model(onnx_model)
try:
import onnxsim
onnx_model, check = onnxsim.simplify(onnx_model)
assert check, "assert check failed"
except Exception as e:
print(f"Simplify failure: {e}")
onnx.save(onnx_model, save_onnx_path)
print(f"ONNX export success, save into {save_onnx_path}")
num_classes = baseModel.bbox_head.num_classes
strides = baseModel.bbox_head.featmap_strides
labels = baseModel.dataset_meta["classes"]
anchors = (
torch.tensor(baseModel.cfg.anchors).flatten().tolist()
if hasattr(baseModel.cfg, "anchors")
else []
)
masks = (
{
f"side{int(args.img_size[0] // num)}": list(range(i * 3, i * 3 + 3))
for i, num in enumerate(strides)
}
if anchors
else {}
)
export_json = args.work_dir.joinpath("model.json")
export_json.write_text(
json.dumps(
{
"nn_config": {
"output_format": "detection",
"NN_family": "YOLO",
"input_size": f"{args.img_size[0]}x{args.img_size[1]}",
"NN_specific_metadata": {
"classes": num_classes,
"coordinates": 4,
"anchors": anchors,
"anchor_masks": masks,
"iou_threshold": 0.5,
"confidence_threshold": 0.5,
},
},
"mappings": {"labels": labels},
},
indent=4,
)
)
if __name__ == "__main__":
main()
然后使用脚本转换：
export_onnx.py:
usage: export_onnx.py [-h] -m INPUT_MODEL
[-v {yolox,yolov5,yolov6,yolov7,yolov8,ppyoloe}]
[-n NAME] [-o OUTPUT_DIR] [-b] [-s]
[-sh {1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16}]
[-t {docker,blobconverter,local}]
Tool for converting YOLO models to the blob format used by OAK
options:
-h, --help show this help message and exit
-m INPUT_MODEL, -i INPUT_MODEL, -w INPUT_MODEL, --input_model INPUT_MODEL
Path to ONNX .onnx file (default: None)
-v {yolox,yolov5,yolov6,yolov7,yolov8,ppyoloe}, --version {yolox,yolov5,yolov6,yolov7,yolov8,ppyoloe}
YOLO version (default: yolov5)
-n NAME, --name NAME The name of the model to be saved, none means using
the same name as the input model (default: None)
-o OUTPUT_DIR, --output_dir OUTPUT_DIR
Directory for saving files, none means using the same
path as the input model (default: None)
-b, --blob turn on OAK Blob export (default: False)
-s, --spatial_detection
Inference with depth information (default: False)
-sh {1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16}, --shaves {1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16}
Specifies number of SHAVE cores that converted model
will use (default: None)
-t {docker,blobconverter,local}, --convert_tool {docker,blobconverter,local}
Which tool is used to convert, docker: should already
have docker (https://docs.docker.com/get-docker/) and
docker-py (pip install docker) installed;
blobconverter: uses an online server to convert the
model and should already have blobconverter (pip
install blobconverter); local: use openvino-dev (pip
install openvino-dev) and openvino 2022.1 ( https://do
cs.oakchina.cn/en/latest/pages/Advanced/Neural_network
s/local_convert_openvino.html#id2 ) to convert
(default: blobconverter)
# coding=utf-8
import argparse
import logging
import time
import warnings
from argparse import ArgumentDefaultsHelpFormatter
from pathlib import Path
import onnx
warnings.filterwarnings("ignore")
try:
from rich import print
from rich.logging import RichHandler
logging.basicConfig(
level="INFO",
format="%(message)s",
datefmt="[%X]",
handlers=[
RichHandler(
rich_tracebacks=False,
show_path=False,
)
],
)
except ImportError:
logging.basicConfig(
level="INFO",
format="%(asctime)s\t%(levelname)s\t%(message)s",
datefmt="[%X]",
)
def parse_args():
parser = argparse.ArgumentParser(
description="Tool for converting YOLO models to the blob format used by OAK",
formatter_class=ArgumentDefaultsHelpFormatter,
)
parser.add_argument(
"-m",
"-i",
"-w",
"--input_model",
type=Path,
required=True,
help="Path to ONNX .onnx file",
)
parser.add_argument(
"-v",
"--version",
type=str,
choices=["yolox", "yolov5", "yolov6", "yolov7", "yolov8", "ppyoloe"],
default="yolov5",
help="YOLO version",
)
parser.add_argument(
"-n",
"--name",
type=str,
help="The name of the model to be saved, "
+ "none means using the same name as the input model",
)
parser.add_argument(
"-o",
"--output_dir",
type=Path,
help="Directory for saving files, "
+ "none means using the same path as the input model",
)
parser.add_argument(
"-b",
"--blob",
action="store_true",
help="turn on OAK Blob export",
)
parser.add_argument(
"-s",
"--spatial_detection",
action="store_true",
help="Inference with depth information",
)
parser.add_argument(
"-sh",
"--shaves",
type=int,
choices=range(1, 17),
help="Specifies number of SHAVE cores that converted model will use",
)
parser.add_argument(
"-t",
"--convert_tool",
type=str,
help="Which tool is used to convert, "
+ "docker: should already have docker (https://docs.docker.com/get-docker/) "
+ "and docker-py (pip install docker) installed; "
+ "blobconverter: uses an online server to convert the model "
+ "and should already have blobconverter (pip install blobconverter); "
+ "local: use openvino-dev (pip install openvino-dev) "
+ "and openvino 2022.1 ( https://docs.oakchina.cn/en/latest/pages/"
+ "Advanced/Neural_networks/local_convert_openvino.html#id2 ) to convert",
default="blobconverter",
choices=["docker", "blobconverter", "local"],
)
args = parser.parse_args()
args.input_model = args.input_model.resolve().absolute()
if args.name is None:
args.name = args.input_model.stem
if args.output_dir is None:
args.output_dir = args.input_model.parent
if args.shaves is None:
args.shaves = 5 if args.spatial_detection else 6
return args
def modify_yolox(input_model, output_model):
t = time.time()
logging.info("Start to modify yolox with onnx %s..." % onnx.__version__)
onnx_model = onnx.load(input_model)
N, C, H, W = [
dim.dim_value for dim in onnx_model.graph.input[0].type.tensor_type.shape.dim
]
removed_outputs = [n for n in onnx_model.graph.output]
xyhw_conf_classes = int(
removed_outputs[0].type.tensor_type.shape.dim[1].dim_value + 5
)
logging.info("remove old outputs")
for n in removed_outputs:
onnx_model.graph.output.remove(n)
logging.info("get the node to be modify:")
cls_preds = []
reg_preds = []
obj_preds = []
for i, n in enumerate(onnx_model.graph.node):
if "multi_level_conv_cls" in n.name:
cls_preds.append(i)
elif "multi_level_conv_reg" in n.name:
reg_preds.append(i)
elif "multi_level_conv_obj" in n.name:
obj_preds.append(i)
logging.info(f"{cls_preds, reg_preds, obj_preds = }")
num = len(cls_preds)
for i, (cls, reg, obj) in enumerate(zip(cls_preds, reg_preds, obj_preds)):
if num == 2:
H_ = int(H / 2 ** (i + 4))
W_ = int(W / 2 ** (i + 4))
elif num == 3:
H_ = int(H / 2 ** (i + 3))
W_ = int(W / 2 ** (i + 3))
sigmoid_cls = onnx.helper.make_node(
"Sigmoid",
inputs=[onnx_model.graph.node[cls].output[0]],
outputs=[f"Sigmoid_{cls}"],
)
onnx_model.graph.node.append(sigmoid_cls)
sigmoid_obj = onnx.helper.make_node(
"Sigmoid",
inputs=[onnx_model.graph.node[obj].output[0]],
outputs=[f"Sigmoid_{obj}"],
)
onnx_model.graph.node.append(sigmoid_obj)
concat = onnx.helper.make_node(
"Concat",
inputs=[
onnx_model.graph.node[reg].output[0],
f"Sigmoid_{obj}",
f"Sigmoid_{cls}",
],
outputs=[f"output{i+1}_yolov6"],
axis=1,
)
onnx_model.graph.node.append(concat)
new_output = onnx.helper.make_tensor_value_info(
f"output{i+1}_yolov6",
onnx.TensorProto.FLOAT,
[N, xyhw_conf_classes, H_, W_],
)
onnx_model.graph.output.extend([new_output])
onnx.save(onnx_model, output_model)
logging.info("Modify complete (%.2fs).\n" % (time.time() - t))
def modify_yolov5(input_model, output_model):
t = time.time()
logging.info("Start to modify yolov5 with onnx %s..." % onnx.__version__)
onnx_model = onnx.load(input_model)
N, C, H, W = [
dim.dim_value for dim in onnx_model.graph.input[0].type.tensor_type.shape.dim
]
removed_outputs = [n for n in onnx_model.graph.output]
xyhw_conf_classes = int(
removed_outputs[0].type.tensor_type.shape.dim[1].dim_value + 5
)
logging.info("remove old outputs")
for n in removed_outputs:
onnx_model.graph.output.remove(n)
logging.info("get the node to be modify:")
convs_preds = []
for i, n in enumerate(onnx_model.graph.node):
if "convs_pred" in n.name:
convs_preds.append(i)
logging.info(f"{convs_preds = }")
num = len(convs_preds)
for i, cls in enumerate(convs_preds):
if num == 2:
H_ = int(H / 2 ** (i + 4))
W_ = int(W / 2 ** (i + 4))
elif num == 3:
H_ = int(H / 2 ** (i + 3))
W_ = int(W / 2 ** (i + 3))
sigmoid = onnx.helper.make_node(
"Sigmoid",
inputs=[onnx_model.graph.node[cls].output[0]],
outputs=[f"output{i+1}_yolov5"],
)
onnx_model.graph.node.append(sigmoid)
new_output = onnx.helper.make_tensor_value_info(
f"output{i+1}_yolov5",
onnx.TensorProto.FLOAT,
[N, xyhw_conf_classes, H_, W_],
)
onnx_model.graph.output.extend([new_output])
onnx.save(onnx_model, output_model)
logging.info("Modify complete (%.2fs).\n" % (time.time() - t))
def modify_yolov6(input_model, output_model):
t = time.time()
logging.info("Start to modify yolov6 with onnx %s..." % onnx.__version__)
onnx_model = onnx.load(input_model)
N, C, H, W = [
dim.dim_value for dim in onnx_model.graph.input[0].type.tensor_type.shape.dim
]
removed_outputs = [n for n in onnx_model.graph.output]
xyhw_conf_classes = int(
removed_outputs[0].type.tensor_type.shape.dim[1].dim_value + 5
)
logging.info("remove old outputs")
for n in removed_outputs:
onnx_model.graph.output.remove(n)
logging.info("get the node to be modify:")
cls_preds = []
for i, n in enumerate(onnx_model.graph.node):
if "cls_preds" in n.name:
cls_preds.append(i)
logging.info(f"{cls_preds = }")
num = len(cls_preds)
for i, cls in enumerate(cls_preds):
if num == 2:
H_ = int(H / 2 ** (i + 4))
W_ = int(W / 2 ** (i + 4))
elif num == 3:
H_ = int(H / 2 ** (i + 3))
W_ = int(W / 2 ** (i + 3))
sigmoid = onnx.helper.make_node(
"Sigmoid",
inputs=[onnx_model.graph.node[cls].output[0]],
outputs=[f"Sigmoid_{cls}"],
)
onnx_model.graph.node.append(sigmoid)
reduceMax = onnx.helper.make_node(
"ReduceMax",
inputs=[f"Sigmoid_{cls}"],
outputs=[f"ReduceMax_{cls}"],
keepdims=1,
axes=[1],
)
onnx_model.graph.node.append(reduceMax)
concat = onnx.helper.make_node(
"Concat",
inputs=[
onnx_model.graph.node[cls + 1].output[0],
f"ReduceMax_{cls}",
f"Sigmoid_{cls}",
],
outputs=[f"output{i+1}_yolov6r2"],
axis=1,
)
onnx_model.graph.node.append(concat)
new_output = onnx.helper.make_tensor_value_info(
f"output{i+1}_yolov6r2",
onnx.TensorProto.FLOAT,
[N, xyhw_conf_classes, H_, W_],
)
onnx_model.graph.output.extend([new_output])
onnx.save(onnx_model, output_model)
logging.info("Modify complete (%.2fs).\n" % (time.time() - t))
def modify_yolov7(input_model, output_model):
t = time.time()
logging.info("Start to modify yolov7 with onnx %s..." % onnx.__version__)
onnx_model = onnx.load(input_model)
N, C, H, W = [
dim.dim_value for dim in onnx_model.graph.input[0].type.tensor_type.shape.dim
]
removed_outputs = [n for n in onnx_model.graph.output]
xyhw_conf_classes = int(
removed_outputs[0].type.tensor_type.shape.dim[1].dim_value + 5
)
logging.info("remove old outputs")
for n in removed_outputs:
onnx_model.graph.output.remove(n)
logging.info("get the node to be modify:")
convs_preds = []
for i, n in enumerate(onnx_model.graph.node):
if "convs_pred" in n.name:
convs_preds.append(i)
logging.info(f"{convs_preds = }")
num = len(convs_preds)
for i, cls in enumerate(convs_preds):
if num == 2:
H_ = int(H / 2 ** (i + 4))
W_ = int(W / 2 ** (i + 4))
elif num == 3:
H_ = int(H / 2 ** (i + 3))
W_ = int(W / 2 ** (i + 3))
sigmoid = onnx.helper.make_node(
"Sigmoid",
inputs=[onnx_model.graph.node[cls].output[0]],
outputs=[f"output{i+1}_yolov5"],
)
onnx_model.graph.node.append(sigmoid)
new_output = onnx.helper.make_tensor_value_info(
f"output{i+1}_yolov7",
onnx.TensorProto.FLOAT,
[N, xyhw_conf_classes, H_, W_],
)
onnx_model.graph.output.extend([new_output])
onnx.save(onnx_model, output_model)
logging.info("Modify complete (%.2fs).\n" % (time.time() - t))
def modify_yolov8(input_model, output_model):
t = time.time()
logging.info("Start to modify yolov8 with onnx %s..." % onnx.__version__)
onnx_model = onnx.load(input_model)
N, C, H, W = [
dim.dim_value for dim in onnx_model.graph.input[0].type.tensor_type.shape.dim
]
removed_outputs = [n for n in onnx_model.graph.output]
xyhw_conf_classes = int(
removed_outputs[0].type.tensor_type.shape.dim[1].dim_value + 5
)
logging.info("remove old outputs")
for n in removed_outputs:
onnx_model.graph.output.remove(n)
logging.info("get the node to be modify:")
cls_preds = []
reg_preds = []
for i, n in enumerate(onnx_model.graph.node):
if "cls_preds" in n.name:
if "2/Conv" in n.name:
cls_preds.append(i)
elif "reg_preds" in n.name:
if "2/Conv" in n.name:
reg_preds.append(i + 6)
logging.info(f"{cls_preds, reg_preds = }")
num = len(cls_preds)
for i, (cls, reg) in enumerate(zip(cls_preds, reg_preds)):
if num == 2:
H_ = int(H / 2 ** (i + 4))
W_ = int(W / 2 ** (i + 4))
elif num == 3:
H_ = int(H / 2 ** (i + 3))
W_ = int(W / 2 ** (i + 3))
sigmoid = onnx.helper.make_node(
"Sigmoid",
inputs=[onnx_model.graph.node[cls].output[0]],
outputs=[f"Sigmoid_{cls}"],
)
onnx_model.graph.node.append(sigmoid)
reduceMax = onnx.helper.make_node(
"ReduceMax",
inputs=[f"Sigmoid_{cls}"],
outputs=[f"ReduceMax_{cls}"],
keepdims=1,
axes=[1],
)
onnx_model.graph.node.append(reduceMax)
concat = onnx.helper.make_node(
"Concat",
inputs=[
onnx_model.graph.node[reg].output[0],
f"ReduceMax_{cls}",
f"Sigmoid_{cls}",
],
outputs=[f"output{i+1}_yolov6r2"],
axis=1,
)
onnx_model.graph.node.append(concat)
new_output = onnx.helper.make_tensor_value_info(
f"output{i+1}_yolov6r2",
onnx.TensorProto.FLOAT,
[N, xyhw_conf_classes, H_, W_],
)
onnx_model.graph.output.extend([new_output])
onnx.save(onnx_model, output_model)
logging.info("Modify complete (%.2fs).\n" % (time.time() - t))
def modify_ppyoloe(input_model, output_model):
t = time.time()
logging.info("Start to modify yolov8 with onnx %s..." % onnx.__version__)
onnx_model = onnx.load(input_model)
N, C, H, W = [
dim.dim_value for dim in onnx_model.graph.input[0].type.tensor_type.shape.dim
]
removed_outputs = [n for n in onnx_model.graph.output]
xyhw_conf_classes = int(
removed_outputs[0].type.tensor_type.shape.dim[1].dim_value + 5
)
logging.info("remove old outputs")
for n in removed_outputs:
onnx_model.graph.output.remove(n)
logging.info("get the node to be modify:")
cls_preds = []
reg_preds = []
for i, n in enumerate(onnx_model.graph.node):
if "cls_preds" in n.name:
cls_preds.append(i)
elif "reg_preds" in n.name:
reg_preds.append(i + 4)
logging.info(f"{cls_preds, reg_preds = }")
num = len(cls_preds)
for i, (cls, reg) in enumerate(zip(cls_preds, reg_preds)):
if num == 2:
H_ = int(H / 2 ** (i + 4))
W_ = int(W / 2 ** (i + 4))
elif num == 3:
H_ = int(H / 2 ** (i + 3))
W_ = int(W / 2 ** (i + 3))
sigmoid = onnx.helper.make_node(
"Sigmoid",
inputs=[onnx_model.graph.node[cls].output[0]],
outputs=[f"Sigmoid_{cls}"],
)
onnx_model.graph.node.append(sigmoid)
reduceMax = onnx.helper.make_node(
"ReduceMax",
inputs=[f"Sigmoid_{cls}"],
outputs=[f"ReduceMax_{cls}"],
keepdims=1,
axes=[1],
)
onnx_model.graph.node.append(reduceMax)
reshape_shape = onnx.helper.make_tensor(
f"reshape_shape_{reg}", onnx.TensorProto.INT64, dims=[3], vals=[1, -1, 4]
)
onnx_model.graph.initializer.append(reshape_shape)
reshape = onnx.helper.make_node(
"Reshape",
inputs=[onnx_model.graph.node[reg].output[0], f"reshape_shape_{reg}"],
outputs=[f"Reshape_{reg}"],
allowzero=0,
)
onnx_model.graph.node.append(reshape)
transpose = onnx.helper.make_node(
"Transpose",
inputs=[f"Reshape_{reg}"],
outputs=[f"Transpose_{reg}"],
perm=[0, 2, 1],
)
onnx_model.graph.node.append(transpose)
reshape_shape2 = onnx.helper.make_tensor(
f"reshape_shape2_{reg}",
onnx.TensorProto.INT64,
dims=[4],
vals=[1, 4, H_, W_],
)
onnx_model.graph.initializer.append(reshape_shape2)
reshape2 = onnx.helper.make_node(
"Reshape",
inputs=[f"Transpose_{reg}", f"reshape_shape2_{reg}"],
outputs=[f"Reshape_{reg}_2"],
allowzero=0,
)
onnx_model.graph.node.append(reshape2)
concat = onnx.helper.make_node(
"Concat",
inputs=[
f"Reshape_{reg}_2",
f"ReduceMax_{cls}",
f"Sigmoid_{cls}",
],
outputs=[f"output{i+1}_yolov6r2"],
axis=1,
)
onnx_model.graph.node.append(concat)
new_output = onnx.helper.make_tensor_value_info(
f"output{i+1}_yolov6r2",
onnx.TensorProto.FLOAT,
[N, xyhw_conf_classes, H_, W_],
)
onnx_model.graph.output.extend([new_output])
onnx.save(onnx_model, output_model)
logging.info("Modify complete (%.2fs).\n" % (time.time() - t))
def convert(convert_tool, output_model, shaves, output_dir, name, **kwargs):
t = time.time()
export_dir: Path = output_dir.joinpath(name + "_openvino")
export_dir.mkdir(parents=True, exist_ok=True)
export_xml = export_dir.joinpath(name + ".xml")
export_blob = export_dir.joinpath(name + ".blob")
if convert_tool == "blobconverter":
from zipfile import ZIP_LZMA, ZipFile
import blobconverter
blob_path = blobconverter.from_onnx(
model=str(output_model),
data_type="FP16",
shaves=shaves,
use_cache=False,
version="2022.1",
output_dir=export_dir,
optimizer_params=[
"--scale=255",
"--reverse_input_channel",
"--use_new_frontend",
],
download_ir=True,
)
with ZipFile(blob_path, "r", ZIP_LZMA) as zip_obj:
for name in zip_obj.namelist():
zip_obj.extract(
name,
output_dir,
)
blob_path.unlink()
elif convert_tool == "docker":
import docker
export_dir = Path("/io").joinpath(export_dir.name)
export_xml = export_dir.joinpath(name + ".xml")
export_blob = export_dir.joinpath(name + ".blob")
client = docker.from_env()
image = client.images.pull("openvino/ubuntu20_dev", tag="2022.1.0")
docker_output = client.containers.run(
image=image.tags[0],
command='bash -c "mo -m '
+ f"{name}.onnx -n {name} -o {export_dir} "
+ "--static_shape --reverse_input_channels --scale=255 --use_new_frontend "
+ "&& echo 'MYRIAD_ENABLE_MX_BOOT NO' | tee /tmp/myriad.conf >> /dev/null "
+ "&& /opt/intel/openvino/tools/compile_tool/compile_tool -m "
+ f"{export_xml} -o {export_blob} -ip U8 "
+ f"-VPU_NUMBER_OF_SHAVES {shaves} -VPU_NUMBER_OF_CMX_SLICES {shaves} "
+ '-d MYRIAD -c /tmp/myriad.conf"',
remove=True,
volumes=[
f"{output_dir}:/io",
],
working_dir="/io",
)
logging.info(docker_output.decode("utf8"))
else:
import subprocess as sp
# OpenVINO export
logging.info("Starting to export OpenVINO...")
OpenVINO_cmd = (
f"mo --input_model {output_model} --output_dir {export_dir} "
+ "--data_type FP16 --scale 255 --reverse_input_channel"
)
try:
sp.check_output(OpenVINO_cmd, shell=True)
logging.info("OpenVINO export success, saved as %s" % export_dir)
except sp.CalledProcessError:
logging.exception("")
logging.warning("OpenVINO export failure!")
logging.warning(
"By the way, you can try to export OpenVINO use:\n\t%s" % OpenVINO_cmd
)
# OAK Blob export
logging.info("Then you can try to export blob use:")
blob_cmd = (
"echo 'MYRIAD_ENABLE_MX_BOOT ON' | tee /tmp/myriad.conf\n"
+ f"compile_tool -m {export_xml} -o {export_blob} "
+ "-ip U8 -d MYRIAD "
+ f"-VPU_NUMBER_OF_SHAVES {shaves} -VPU_NUMBER_OF_CMX_SLICES {shaves} "
+ "-c /tmp/myriad.conf"
)
logging.info("%s" % blob_cmd)
logging.info(
"compile_tool maybe in the path: "
+ "/opt/intel/openvino/tools/compile_tool/compile_tool, "
+ "if you install openvino 2022.1 with apt"
)
logging.info("Convert complete (%.2fs).\n" % (time.time() - t))
if __name__ == "__main__":
modify_onnx = {
"yolox": modify_yolox,
"yolov5": modify_yolov5,
"yolov6": modify_yolov6,
"yolov7": modify_yolov7,
"yolov8": modify_yolov8,
"ppyoloe": modify_ppyoloe,
}
args = parse_args()
logging.info(args)
print()
output_model = args.output_dir / (args.name + ".onnx")
modify_onnx[args.version](input_model=args.input_model, output_model=output_model)
if args.blob:
convert(output_model=output_model, **vars(args))
可以使用 Netron 查看模型结构
▌转换
openvino 本地转换
onnx -> openvino
mo 是 openvino_dev 2022.1 中脚本，
安装命令为pip install openvino-dev
mo --input_model ppyoloe_plus_s_fast_8xb8-80e_coco.onnx --scale 255 --reverse_input_channel
openvino -> blob
<path>/compile_tool -m ppyoloe_plus_s_fast_8xb8-80e_coco.xml \
-ip U8 -d MYRIAD \
-VPU_NUMBER_OF_SHAVES 6 \
-VPU_NUMBER_OF_CMX_SLICES 6
在线转换
blobconvert 网页
- 进入网页，按下图指示操作：
- 修改参数，转换模型：
- 选择 onnx 模型
- 修改
optimizer_params为--data_type=FP16 --scale 255 --reverse_input_channel - 修改
shaves为6 - 转换
blobconverter python 代码
blobconverter.from_onnx(
"ppyoloe_plus_s_fast_8xb8-80e_coco.onnx",
optimizer_params=[
" --scale 255",
"--reverse_input_channel",
],
shaves=6,
)
blobconvert cli
blobconverter --onnx ppyoloe_plus_s_fast_8xb8-80e_coco.onnx -sh 6 -o . --optimizer-params "scale=255 --reverse_input_channel"
▌DepthAI 示例
正确解码需要可配置的网络相关参数：
使用 export_yolo.py 转换模型时会将相关参数写入 json 文件中，可根据 json 文件中数据添加下列参数
- setNumClasses – YOLO 检测类别的数量
- setIouThreshold – iou 阈值
- setConfidenceThreshold – 置信度阈值，低于该阈值的对象将被过滤掉
- setAnchors – yolo 锚点
- setAnchorMasks – 锚掩码
相关示例可参考
- yoloX请查看
- yolov5请查看
- yolov6请查看
- yolov7请查看
- yolov8请查看
- yolov5lite请查看
# coding=utf-8
import cv2
import depthai as dai
import numpy as np
numClasses = 80
model = dai.OpenVINO.Blob("ppyoloe_plus_s_fast_8xb8-80e_coco.blob")
dim = next(iter(model.networkInputs.values())).dims
W, H = dim[:2]
output_name, output_tenser = next(iter(model.networkOutputs.items()))
if "yolov6" in output_name:
numClasses = output_tenser.dims[2] - 5
else:
numClasses = output_tenser.dims[2] // 3 - 5
labelMap = [
# "class_1","class_2","..."
"class_%s" % i
for i in range(numClasses)
]
# Create pipeline
pipeline = dai.Pipeline()
# Define sources and outputs
camRgb = pipeline.create(dai.node.ColorCamera)
detectionNetwork = pipeline.create(dai.node.YoloDetectionNetwork)
xoutRgb = pipeline.create(dai.node.XLinkOut)
xoutNN = pipeline.create(dai.node.XLinkOut)
xoutRgb.setStreamName("image")
xoutNN.setStreamName("nn")
# Properties
camRgb.setPreviewSize(W, H)
camRgb.setResolution(dai.ColorCameraProperties.SensorResolution.THE_1080_P)
camRgb.setInterleaved(False)
camRgb.setColorOrder(dai.ColorCameraProperties.ColorOrder.BGR)
# Network specific settings
detectionNetwork.setBlob(model)
detectionNetwork.setConfidenceThreshold(0.5)
# Yolo specific parameters
detectionNetwork.setNumClasses(numClasses)
detectionNetwork.setCoordinateSize(4)
detectionNetwork.setAnchors([])
detectionNetwork.setAnchorMasks({})
detectionNetwork.setIouThreshold(0.5)
# Linking
camRgb.preview.link(detectionNetwork.input)
camRgb.preview.link(xoutRgb.input)
detectionNetwork.out.link(xoutNN.input)
# Connect to device and start pipeline
with dai.Device(pipeline) as device:
# Output queues will be used to get the rgb frames and nn data from the outputs defined above
imageQueue = device.getOutputQueue(name="image", maxSize=4, blocking=False)
detectQueue = device.getOutputQueue(name="nn", maxSize=4, blocking=False)
frame = None
detections = []
# nn data, being the bounding box locations, are in <0..1> range - they need to be normalized with frame width/height
def frameNorm(frame, bbox):
normVals = np.full(len(bbox), frame.shape[0])
normVals[::2] = frame.shape[1]
return (np.clip(np.array(bbox), 0, 1) * normVals).astype(int)
def drawText(frame, text, org, color=(255, 255, 255), thickness=1):
cv2.putText(
frame, text, org, cv2.FONT_HERSHEY_SIMPLEX, 0.5, (0, 0, 0), thickness + 3, cv2.LINE_AA
)
cv2.putText(
frame, text, org, cv2.FONT_HERSHEY_SIMPLEX, 0.5, color, thickness, cv2.LINE_AA
)
def drawRect(frame, topLeft, bottomRight, color=(255, 255, 255), thickness=1):
cv2.rectangle(frame, topLeft, bottomRight, (0, 0, 0), thickness + 3)
cv2.rectangle(frame, topLeft, bottomRight, color, thickness)
def displayFrame(name, frame):
color = (128, 128, 128)
for detection in detections:
bbox = frameNorm(
frame, (detection.xmin, detection.ymin, detection.xmax, detection.ymax)
)
drawText(
frame=frame,
text=labelMap[detection.label],
org=(bbox[0] + 10, bbox[1] + 20),
)
drawText(
frame=frame,
text=f"{detection.confidence:.2%}",
org=(bbox[0] + 10, bbox[1] + 35),
)
drawRect(
frame=frame,
topLeft=(bbox[0], bbox[1]),
bottomRight=(bbox[2], bbox[3]),
color=color,
)
# Show the frame
cv2.imshow(name, frame)
while True:
imageQueueData = imageQueue.tryGet()
detectQueueData = detectQueue.tryGet()
if imageQueueData is not None:
frame = imageQueueData.getCvFrame()
if detectQueueData is not None:
detections = detectQueueData.detections
if frame is not None:
displayFrame("rgb", frame)
if cv2.waitKey(1) == ord("q"):
break
▌参考资料
https://docs.oakchina.cn/en/latest/
https://www.oakchina.cn/selection-guide/
OAK中国
| OpenCV AI Kit在中国区的官方代理商和技术服务商
| 追踪AI技术和产品新动态
戳「+关注」获取最新资讯↗↗