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 checkpointpositional arguments:config                Config filecheckpoint            Checkpoint fileoptions:-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 OPSETONNX opset version (default: 12)

# coding=utf-8
import argparse
import json
import warnings
from io import BytesIO
from argparse import ArgumentDefaultsHelpFormatter
from pathlib import Pathimport 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 printfrom mmyolo.models import RepVGGBlock
from mmyolo.models.layers import CSPLayerWithTwoConvwarnings.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 1args.work_dir = args.work_dir.resolve().absolute()print(args)return argsdef build_model_from_cfg(config_path, checkpoint_path, device):model = init_detector(config_path, checkpoint_path, device=device)model.eval()return modelclass 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.shapex = 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 = biasself.divisor = divisorself.min_value = min_valueself.max_value = max_valuedef forward(self, x):"""forward"""x = (x + self.bias) / self.divisorreturn 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 = NonebaseModel = 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 runbaseModel(fake_input)save_onnx_path = args.work_dir.joinpath(args.config.with_suffix(".onnx").name)# export onnxwith 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 onnxsimonnx_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_classesstrides = baseModel.bbox_head.featmap_strideslabels = 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 anchorselse {})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 OAKoptions:-h, --help            show this help message and exit-m INPUT_MODEL, -i INPUT_MODEL, -w INPUT_MODEL, --input_model INPUT_MODELPath 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 usingthe same name as the input model (default: None)-o OUTPUT_DIR, --output_dir OUTPUT_DIRDirectory for saving files, none means using the samepath as the input model (default: None)-b, --blob            turn on OAK Blob export (default: False)-s, --spatial_detectionInference 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 modelwill use (default: None)-t {docker,blobconverter,local}, --convert_tool {docker,blobconverter,local}Which tool is used to convert, docker: should alreadyhave docker (https://docs.docker.com/get-docker/) anddocker-py (pip install docker) installed;blobconverter: uses an online server to convert themodel and should already have blobconverter (pipinstall blobconverter); local: use openvino-dev (pipinstall 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)

# coding=utf-8
import argparse
import logging
import time
import warnings
from argparse import ArgumentDefaultsHelpFormatter
from pathlib import Pathimport onnxwarnings.filterwarnings("ignore")try:from rich import printfrom rich.logging import RichHandlerlogging.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.stemif args.output_dir is None:args.output_dir = args.input_model.parentif args.shaves is None:args.shaves = 5 if args.spatial_detection else 6return argsdef 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, ZipFileimport blobconverterblob_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 dockerexport_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 exportlogging.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 exportlogging.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

/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：
  
  1. 选择 onnx 模型
  2. 修改 optimizer_params 为 --data_type=FP16 --scale 255 --reverse_input_channel
  3. 修改 shaves 为 6
  4. 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 npnumClasses = 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 - 5labelMap = [# "class_1","class_2","...""class_%s" % ifor 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 aboveimageQueue = device.getOutputQueue(name="image", maxSize=4, blocking=False)detectQueue = device.getOutputQueue(name="nn", maxSize=4, blocking=False)frame = Nonedetections = []# nn data, being the bounding box locations, are in <0..1> range - they need to be normalized with frame width/heightdef 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 framecv2.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.detectionsif 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↗↗