#-*- coding:utf-8 -*- import torch #量化仅可用cpu model = ResNet().cpu() model = torch.load_state_dict(torch.load(weights)) #Specify quantization configuration #在这一步声明了对称量化或非对称量化，及量化bit数 #如下代码中采用了默认的非对称量化，及8bit model.qconfig = torch.quantization.default_qconfig model = torch.quantization.prepare(model) #Convert to quantized model model = torch.quantization.convert(model) #Save model, 保存后模型的size显著减小，但性能损失相对较大 #故，建议考虑量化感知训练 torch.save(model.state_dict(), "path.pt")
