kuangliu / mocha

Convert torch model to/from caffe model easily

Mocha: Convert torch model to/from caffe model easily

Caffe: load with Python.
Torch: load with Lua.

Q. How to connect Python & Lua code?
A. We don't need directly interact Python code with Lua code, which is complex. We use intermediate file exchange, the model parameters are saved to disk layer by layer as .npy file, and then loaded to caffe/torch separately.

TODO:

• Caffe split layer -> Torch ConcatTable
• PReLU
• More tests

mocha's Issues

Wrong bottom_name

This project has very few dependencies, I would like to thank the author. However, there are some problems about the relu layer.

Since the .prototxt file is builded based on the net.json file, and the 'bottom_name' is based on the 'name' in the json file. It will cause error if the relu layer is used since 'top' and 'bottom' name are both not the 'name' of the layer.

Main error:
The 'name' of net_config is not equal to the 'bottom_name' of caffe model
https://github.com/kuangliu/mocha/blob/master/torch/export.lua

https://github.com/kuangliu/mocha/blob/master/caffe/import.py#L115

Example:

layer {
  name: "relu2"
  type: "ReLU"
  bottom: "conv1"
  top: "conv1"
}
layer {
  name: "pool3"
  type: "Pooling"
  bottom: "relu2" ---------------------> wrong, it should be "conv1"
  top: "pool3"
  pooling_param {
    pool: MAX
    kernel_h: 2
    kernel_w: 2
    stride_h: 2
    stride_w: 2
    pad_h: 0
    pad_w: 0
  }
}

How to use this tool?

Hi,
Could you kindly explain how to use this tool? Since there is not much information in READEME.
For example, I have a pre-trained Torch model (nn4.small2.v1.t7) which is from OpenFace,
how to use this tool to convert the Torch model(nn4.small2.v1.t7) to Caffe by this tool?

Thanks,
Victor