按照链接:https://github.com/CambriconECO/Pytorch_Yolov5_Inference/tree/eb36b68e5a4ee3fe23c45ec545bef68dced452bb 在MLU270上推理测试,画框正常了,但是保存离线模型后,使用官方pytorch docker镜像,yolov5离线推理程序推理出的结果计算mAP很低:
保存离线模型:
if opt.jit:
if opt.save:
ct.save_as_cambricon('yolov5s')
torch.set_grad_enabled(False)
ct.set_core_number(1)
trace_input = torch.randn(1, 3, 640, 640, dtype=torch.float)
trace_input = trace_input.to(ct.mlu_device())
quantized_net = torch.jit.trace(quantized_net, trace_input, check_trace = False)
拷贝离线模型并切换路径:
/torch/src/catch/examples/offline/build/yolov5
离线推理:
./yolov5_offline_multicore -offlinemodel yolov5s.cambricon -dataset_path /torch/src/catch/examples/data/COCO2017_5000/coco -images /torch/src/catch/examples/data/COCO2017_5000/coco/val2017.txt -labels ./label_map_coco.txt -simple_compile 1 -input_format 0
计算meanAP:
python3 ../../scripts/meanAP_COCO.py --file_list /torch/src/catch/examples/data/COCO2017_5000/coco/val2017.txt --result_dir ./ --ann_dir /torch/src/catch/examples/data/COCO2017_5000/coco
结果:
麻烦帮看看,有可能是什么原因导致的?