anymake / pytorchconverter

Pytorch model to caffe & 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))