工程代码:https://github.com/eriklindernoren/PyTorch-YOLOv3
第一步:下载Darknet 版的yolov3.weights
第二部:使用以上工程代码加载模型:model.load_darknet_weights() 并保存为pytorch模型格式yolov3.pth:
torch.save(model.state_dict(), pytorch_model)
第三步:用 torch.jit.trace convert 模型为 .pt 格式。 这里注意:参数check_trace=False,不然会报错
traced_script_module = torch.jit.trace(model, example,check_trace=False)
参考:https://github.com/pytorch/pytorch/issues/23993
注意:torch.jit.trace()不支持控制流语句,需将models.py 中 if targets is None 以下语句注释掉,增加
return output,0 因为作为推断,所以注释掉的语句不影响。
import numpy as np import torch # import torchsnooper from models import Darknet device = torch.device("cuda" if torch.cuda.is_available() else "cpu") # with torch.jit.optimized_execution(True): def save_pytorch_model(darknet_model,pytorch_model,config_file): model = Darknet(config_file).to(device) model.load_darknet_weights(darknet_model) torch.save(model.state_dict(), pytorch_model) print('pytorch model saved!') def convert_pytorch_model_to_libtorch(config_file,pytorch_model,libtorch_model): model = Darknet(config_file).to(device) model.load_state_dict(torch.load(pytorch_model)) model.eval() example = torch.rand(1,3,416,416).cuda() with torch.jit.optimized_execution(True): print('pp') # example 报错 Expected type 'tuple', got 'Tensor' instead ,可能是input没有放cuda上 traced_script_module = torch.jit.trace(model, example,check_trace=False)# save the converted model print('oo') traced_script_module.save(libtorch_model) output = traced_script_module(torch.rand(1,3,416,416).cuda()) print(output) print('trace model saved!')