Mxnet:以全连接层为例子自定义新的操作(层)

在使用深度学习平台时，光会使用其中已定义好的操作有时候是满足不了实际使用的，一般需要我们自己定义新的操作。但是，绝大多数深度平台都是编译好的，很难再次编写。本文以Mxnet为例，官方给出四种定义新操作的方法，

分别调用：

１、mx.operator.CustomOp

２、mx.operator.NDArrayOp

３、mx.operator.NumpyOp

４、使用 C++　定义底层

并且给出了重新定义softmax层的例子。但是sofetmax操作只有前向操作，也没有参数，与我们通常需要需要使用的情况不符，官方文档也没有一个有参数的中间层例子。在此博主给出了一个重新定义全连接操作的例子，希望能够给大家带来帮助。

# pylint: skip-file
import os
from data import mnist_iterator
import mxnet as mx
import numpy as np
import logging
from numpy import *

class Dense(mx.operator.CustomOp):

    def __init__(self, num_hidden):
        self.num_hidden = num_hidden

    def forward(self, is_train, req, in_data, out_data, aux):
        x = in_data[0]
        w = in_data[1]
        b = in_data[2]
        y = out_data[0]
        y[:] = mx.nd.add(mx.nd.dot(x, w.T), b)
        self.assign(out_data[0], req[0], mx.nd.array(yy))

    def backward(self, req, out_grad, in_data, out_data, in_grad, aux):
        dx = in_grad[0]
        dw = in_grad[1]
        db = in_grad[2]
        dy = out_grad[0]
        x = in_data[0]
        w = in_data[1]
        dw[:] = mx.nd.dot(dy.T, x)
        dx[:] = mx.nd.dot(dy, w)
        db[:] = mx.nd.sum(dy, axis=0)
        self.assign(in_grad[0], req[0], dx)
        self.assign(in_grad[1], req[0], dw)
        self.assign(in_grad[2], req[0], db)


@mx.operator.register("dense")
class DenseProp(mx.operator.CustomOpProp):
    def __init__(self, num_hidden):
        super(DenseProp, self).__init__(True)
        # we use constant bias here to illustrate how to pass arguments
        # to operators. All arguments are in string format so you need
        # to convert them back to the type you want.
        self.num_hidden = long(num_hidden)

    def list_arguments(self):
        return ['data', 'weight', 'bias']

    def list_outputs(self):
        #  this can be omitted if you only have 1 output.
        return ['output']

    def infer_shape(self, in_shapes):
        data_shape = in_shapes[0]
        weight_shape = (self.num_hidden, in_shapes[0][1])
        bias_shape = (self.num_hidden,)
        output_shape = (data_shape[0], self.num_hidden)
        return [data_shape, weight_shape, bias_shape], [output_shape], []

    def infer_type(self, in_type):
        dtype = in_type[0]
        return [dtype, dtype, dtype], [dtype], []

    def create_operator(self, ctx, in_shapes, in_dtypes):
        #  create and return the CustomOp class.
        return Dense(self.num_hidden)


# define mlp
data = mx.symbol.Variable('data')
##This is the new defined layer
fc1 = mx.symbol.Custom(data, name='fc1', op_type='dense', num_hidden=128)
act1 = mx.symbol.Activation(data=fc1, name='relu1', act_type="relu")
fc2 = mx.symbol.FullyConnected(data=act1, name = 'fc2', num_hidden = 64)
act2 = mx.symbol.Activation(data = fc2, name='relu2', act_type="relu")
fc3 = mx.symbol.FullyConnected(data = act2, name='fc3', num_hidden=10)
mlp = mx.symbol.Softmax(data = fc3, name = 'softmax')
train, val = mnist_iterator(batch_size=100, input_shape = (784,))
logging.basicConfig(level=logging.DEBUG)
model = mx.model.FeedForward(
    ctx = mx.gpu(1), symbol = mlp, num_epoch = 20,
    learning_rate = 0.1, momentum = 0.9, wd = 0.00001)
model.fit(X=train, eval_data=val,
          batch_end_callback=mx.callback.Speedometer(100,100))