Caffe has been a well-known and widely-used deep learning framework. Now MXNet has supported calling most caffe operators(layers) and loss functions directly in its symbolic graph! Using one's own customized caffe layer is also effortless.

Besides Caffe, MXNet has already embedded Torch modules and its tensor mathematical functions. (link)

This blog demonstrates two steps to use Caffe op in MXNet:

  • How to install MXNet with Caffe support.

  • How to embed Caffe op into MXNet's symbolic graph.

Install Caffe With MXNet interface

Compile with Caffe

  • In mxnet folder, open (if you haven't already, copy make/ (Linux) or make/ (Mac) into MXNet root folder as and uncomment the lines CAFFE_PATH = $(HOME)/caffe and MXNET_PLUGINS += plugin/caffe/ Modify CAFFE_PATH to your caffe installation if necessary.
  • Run make clean && make to build with caffe support.

Caffe Operator (Layer)

Caffe's neural network operator and loss functions are supported by MXNet through mxnet.symbol.CaffeOp and mxnet.symbol.CaffeLoss respectively. For example, the following code shows multi-layer perception network for classifying MNIST digits (full code):


data = mx.symbol.Variable('data')
fc1  = mx.symbol.CaffeOp(data_0=data, num_weight=2, name='fc1', prototxt="layer{type:\"InnerProduct\" inner_product_param{num_output: 128} }")
act1 = mx.symbol.CaffeOp(data_0=fc1, prototxt="layer{type:\"TanH\"}")
fc2  = mx.symbol.CaffeOp(data_0=act1, num_weight=2, name='fc2', prototxt="layer{type:\"InnerProduct\" inner_product_param{num_output: 64} }")
act2 = mx.symbol.CaffeOp(data_0=fc2, prototxt="layer{type:\"TanH\"}")
fc3 = mx.symbol.CaffeOp(data_0=act2, num_weight=2, name='fc3', prototxt="layer{type:\"InnerProduct\" inner_product_param{num_output: 10}}")
mlp = mx.symbol.SoftmaxOutput(data=fc3, name='softmax')

Let's break it down. First data = mx.symbol.Variable('data') defines a variable as placeholder for input. Then it's fed through Caffe operators with fc1 = mx.symbol.CaffeOp(data_0=data, num_weight=2, name='fc1', prototxt="layer{type:\"InnerProduct\" inner_product_param{num_output: 128} }").

The inputs to caffe op are named as data_i for i=0 ... num_data-1 as num_data is the number of inputs. You may skip the argument, as the example does, if its value is 1. While num_weight is number of blobs_(weights). Its default value is 0, as many ops maintain no weight. prototxt is the configuration string.

We could also replace the last line by:

label = mx.symbol.Variable('softmax_label')
mlp = mx.symbol.CaffeLoss(data=fc3, label=label, grad_scale=1, name='softmax', prototxt="layer{type:\"SoftmaxWithLoss\"}")

to use loss function in caffe.

Use customized caffe operators

Running customized operator from mxnet is no difference than using regular ones. There's no need to add any code in mxnet, as mxnet directly calls caffe layer registry.


Caffe-plugin is contributed by Haoran Wang.

Haoran is an incoming master student of MCDS program at Carnegie Mellon University. He received his Bachelor degree in Computer Science from ACM Class at Shanghai Jiao Tong University.

He has many thanks to Minjie Wang, Tianqi Chen, Junyuan Xie and Prof. Zheng Zhang for their helpful advices on implementation and documentation of caffe-plugin.