百科.dev
全部条目在线工具AI 编程趋势榜开源项目技术资讯提交条目
登录
< 返回工具列表
P

PytorchConverter

> 编程语言
开源

从 Pytorch 模型转换为 Caffe 和 ncnn

393 stars0 点赞1 次浏览
访问官网GitHub

工具介绍

从 Pytorch 模型转换为 Caffe 和 ncnn

Pytorch Converter

Pytorch model to Caffe & ncnn

Model Examples

  • SqueezeNet from torchvision
  • DenseNet from torchvision
  • ResNet50 (with ceiling_mode=True)
  • MobileNet
  • AnimeGAN pretrained model from author (https://github.com/jayleicn/animeGAN)
  • SSD-like object detection net(for ncnn)
  • UNet (no pretrained model yet, just default initialization)

Attentions

  • Mind the difference on ceil_mode of pooling layer among Pytorch and Caffe, ncnn

    • You can convert Pytorch models with all pooling layer's ceil_mode=True.
    • Or compile a custom version of Caffe/ncnn with floor() replaced by ceil() in pooling layer inference.
  • Python Errors: Use Pytorch 0.2.0 Only to Convert Your Model

    • Higher version of pytorch 0.3.0, 0.3.1, 0.4.0 seemingly have blocked third party model conversion.
    • Please note that you can still TRAIN your model on pytorch 0.3.0~0.4.0. The converter running on 0.2.0 could still load higher version models correctly.
  • Other Python packages requirements:

    • to Caffe: numpy, protobuf (to gen caffe proto)
    • to ncnn: numpy
    • for testing Caffe result: pycaffe, cv2
  • Model Loading Error

    • Use compatible model saving & loading method, e.g.

      # Saving, notice the difference on DataParallel
      net_for_saving = net.module if use_nn_DataParallel else net
      torch.save(net_for_saving.state_dict(), path)
      
      # Loading
      net.load_state_dict(torch.load(path, map_location=lambda storge, loc: storage))

Issues· 0 开放

查看全部 Issues在 GitHub 打开

暂无开放 Issues,或尚未同步最近议题。

> 标签

Python

暂无评论,来聊聊你的看法吧

> 工具信息

发布日期2026年8月1日
最后更新2026年9月17日
分类编程语言
定价开源

> 相关工具

T
TypeScript
JavaScript 的超集,为前端与全栈提供静态类型
P
Python
通用编程语言,广泛用于 Web、数据与 AI
G
Go
Google 推出的简洁高效系统语言