上篇https://forum.cambricon.com/show-3-2292-1.html 转好模型之后就可以在MLU侧进行验证。

基于MLU验证

模型量化

代码修改（增加mlu/mlu_quant.py 文件）

## 模型量化
---
     # model = attempt_load(weights, map_location=device)  # load FP32 model
     from models.yolo import get_model
     model = get_model(opt)
     # print(model)
 
     # 配置量化参数
     import torch_mlu
     import torch_mlu.core.mlu_model as ct
     import torch_mlu.core.mlu_quantize as mlu_quantize
     qconfig={'use_avg':False, 'data_scale':1.0, 'firstconv':False, 'per_channel': False}
     # 调用量化接口
     quantized_net = mlu_quantize.quantize_dynamic_mlu(model,qconfig_spec=qconfig, dtype='int8', gen_quant=True)
     # 设置为推理模式
     quantized_net = quantized_net.eval().float()
 
     model = quantized_net
---

   # 保存量化模型

     qua_weight = opt.qua_weight
     print("SAVE quantize model:",qua_weight)
     torch.save(model.state_dict(),qua_weight)
 
---
#增加qua_weight参数
     parser.add_argument('--qua_weight', type=str,default='yolov7_intx.pth', help='model.pt path(s)')

运行命令

python mlu/mlu_quant.py --weights mlu/weight/yolov7_unzip.pt --conf 0.25 --img-size 640 --source inference/images/horses.jpg --cfg ./cfg/deploy/yolov7.yaml --qua_weight mlu/weight/yolov7_intx.pth --no-trace

基于MLU运行量化后的模型

代码修改参看（增加mlu/mlu_detect.py 文件）

 #模型加载
    from models.yolo import get_model
    model = get_model(opt)
    # print(model)

    import torch_mlu
    import torch_mlu.core.mlu_model as ct
    import torch_mlu.core.mlu_quantize as mlu_quantize

    from postprocess import MLU_PostProcessYoloV7,PostProcessPytorchYoloV7,draw_image
    from models.yolo import get_empty_model
    model = get_empty_model(opt)

    stride = 32

    # stride = int(model.stride.max())  # model stride
    imgsz = check_img_size(imgsz, s=stride)  # check img_size

    #配置 MLU core number
    ct.set_core_number(opt.core_number)
    # 设置输入图片的通道顺序，以决定首层卷积对三通道输入的补齐通道顺序。默认是 RGBA 顺序
    #ct.set_input_format(0)
    #配置MLU core类型
    ct.set_core_version(opt.mcore)
    torch.set_grad_enabled(False)

    if opt.fake_device:
        print("fake_device mode")
        ct.set_device(-1)

    device = ct.mlu_device()
    print("run on %s ..."%device)
    # 加载量化模型
    weight = weights[0]
    quantized_net = torch_mlu.core.mlu_quantize.quantize_dynamic_mlu(model)
    print('weight:',weight)
    state_dict = torch.load(weight)
    quantized_net.load_state_dict(state_dict, strict=False)
    # 设置为推理模式
    quantized_net = quantized_net.eval().float()
    quantized_net.to(device)

    model = quantized_net

    # 设置在线融合模式
    if opt.jit:
        if opt.save:
            ct.save_as_cambricon(opt.mname)

        example = torch.randn(opt.batch_size, 3, imgsz, imgsz,dtype=torch.float)
        trace_input = torch.randn(1, 3, imgsz, imgsz,dtype=torch.float)

        if opt.half_input:
            print('half_input ')
            trace_input = trace_input.type(torch.HalfTensor)
            example = example.type(torch.HalfTensor)

        print("jit trace example shape",example.shape)
        model = torch.jit.trace(model,trace_input.to(device),check_trace=False)

        #如果是生成离线模型，推理一次，直接退出，会保存离线模型
        if opt.save:
            print("save offline model mname: ",opt.mname)
            model(example.to(device))
            ct.save_as_cambricon('')
            exit(0)
    if opt.mlu_det:
        postproc = MLU_PostProcessYoloV7()
    else:
        postproc = PostProcessPytorchYoloV7(conf_thres=opt.conf_thres,iou_thres=opt.iou_thres)
    names = postproc.names
  
  ---
  #推理及后处理部分
  ---
        # Inference
        t1 = time_synchronized()
        with torch.no_grad():   # Calculating gradients would cause a GPU memory leak
            detect_out = model(img)
            # pred = model(img, augment=opt.augment)[0]
        t2 = time_synchronized()
        if len(detect_out) == 1:
            pred = detect_out.cpu().type(torch.FloatTensor) if opt.half_input else detect_out.cpu()
        else:
            pred = [out.cpu().type(torch.FloatTensor) for out in detect_out]
        # print("mlu pred:{} {}".format(type(pred),pred))

        # from mlu.tools.dump_npy import save_npy
        # save_npy(pred,"mlu")
        # # from postprocess import draw_image
        if opt.mlu_det:
            pred = postproc.get_boxes(pred)
            p, s, im0,   = path, '', im0s, getattr(dataset, ' ', 0)
            p = Path(p)  # to Path
            save_path = str(save_dir / p.name)  # img.jpg
            print(save_path)
            draw_image(pred, img, im0s, path, save_path, names)
            exit(0)
            
        pred = postproc.yolo_det(pred)[0]

备注：后处理参看postprocess.py，mlu推理代码参看mlu_detect.py

运行命令（MLU+CPU NMS）

#逐层运行+CPU NMS
python mlu/mlu_detect.py --weights mlu/weight/yolov7_intx.pth --conf 0.25 --img-size 640 --source inference/images/horses.jpg --cfg ./cfg/deploy/yolov7.yaml
# 融合模型运行+CPU NMS
python mlu/mlu_detect.py --weights mlu/weight/yolov7_intx.pth --conf 0.25 --img-size 640 --source inference/images/horses.jpg --cfg ./cfg/deploy/yolov7.yaml --jit
#生成离线模型
python mlu/mlu_detect.py --weights mlu/weight/yolov7_intx.pth --conf 0.25 --img-size 640 --source inference/images/horses.jpg --cfg ./cfg/deploy/yolov7.yaml --jit --mcore MLU270 --mname yolov7_4b4c --core 4 --batch 4 --save 

运行命令（MLU+mlu Yolo det）

#逐层运行 + MLU yolo detection output
python mlu/mlu_detect.py --weights mlu/weight/yolov7_intx.pth --conf 0.25 --img-size 640 --source inference/images/horses.jpg --cfg ./cfg/deploy/yolov7.yaml --mlu_det
#融合模型运行 + MLU yolo detection output
python mlu/mlu_detect.py --weights mlu/weight/yolov7_intx.pth --conf 0.25 --img-size 640 --source inference/images/horses.jpg --cfg ./cfg/deploy/yolov7.yaml --jit --mlu_det
#生成MLU270离线模型 4B4C
python mlu/mlu_detect.py --weights mlu/weight/yolov7_intx.pth --conf 0.25 --img-size 640 --source inference/images/horses.jpg --cfg ./cfg/deploy/yolov7.yaml --jit --mlu_det --mcore MLU270 --mname yolov7_4b4c --core 4 --batch 4 --save 

#生成MLU270离线模型 1B4C
python mlu/mlu_detect.py --weights mlu/weight/yolov7_intx.pth --conf 0.25 --img-size 640 --source inference/images/horses.jpg --cfg ./cfg/deploy/yolov7.yaml --jit --mlu_det --mcore MLU270 --mname yolov7_1b4c --core 4 --batch 1 --save

#生成MLU220离线模型 4B4C
python mlu/mlu_detect.py --weights mlu/weight/yolov7_intx.pth --conf 0.25 --img-size 640 --source inference/images/horses.jpg --cfg ./cfg/deploy/yolov7.yaml --jit --mlu_det --mcore MLU220 --mname mlu220_yolov7_1b4c --core 4 --batch 1 --save