Tutorial

This tutorial will show you how to use cxxnet train a feed forward neural network, then show you how to modify the feed forward network to a simple convolution neural network.

Feed Forward Example

=== Reference configuration example [MNIST.conf] (../blob/master/example/MNIST/MNIST.conf)

Before you start

• Run the command ./run.sh MNIST.conf in the folder example/MNIST/ to get data.

Setup data iterator configuration

cxxnet use iterator to provide data batch to the network trainer. First we need to set the data type (eval or data). Then we need to specify iterator type (mnist, cifar, image, imgbin, etc). Then set some attribute to the iterator including shuffle, file path and so on. Here is an example for MNIST

Setup training iterator

This part is about Data Iterator setting

• Change path_image to the path of training image file, change path_label to the path of training label file you download just now

data = train
iter = mnist
    path_img = "./data/train-images-idx3-ubyte.gz"
    path_label = "./data/train-labels-idx1-ubyte.gz"
    shuffle = 1
iter = end

Setup test iterator

• Change path_image to the path of test image file, change path_label to the path of test label file you download just now

eval = test
iter = mnist
    path_img = "./data/t10k-images-idx3-ubyte.gz"
    path_label = "./data/t10k-labels-idx1-ubyte.gz"
iter = end

Setup network structure

This part is about Layer Setting

Network structure start with declaration "netconfig=start" and end with "netconfig=end". They layer is declared in the format "layer[ from_num -> to_num ] = layer_type:nick" Then comes the parameters of the layer. Here is an example for MNIST

netconfig=start
layer[0->1] = fullc:fc1
  nhidden = 100
  init_sigma = 0.01
layer[1->2] = sigmoid:se1
layer[2->3] = fullc:fc1
  nhidden = 10
  init_sigma = 0.01
layer[3->3] = softmax
netconfig=end

Notice some special layer like softmax and dropout use self-loop, which means the input node equals output node.

Setup input size and batch size

In this section, we need to set the input shape and batch size. The input shape should be 3 numbers, split by ','; The 3 numbers is channel, height and width for 3D input or 1, 1, dim for 1D vector input. In this example it is

input_shape = 1,1,784
batch_size = 100

Setup global parameters

This part is about Global Setting

Global parameters are used for setting the trainer behavior. In this example, we use the following configuration.

dev = cpu
save_model = 15
max_round = 15
num_round = 15
train_eval = 1
random_type = gaussian

First set working device dev ( cpu or gpu ); frequent to save mode save_model ; training round num_round and max_round and whether to print training set evaluation train_eval. The random_type defines weight initialization method. We provides ( gaussian method and xavier method)

Setup learning parameters

This part is about Updater Setting

learning parameter change the updater behavior. eta is known as learning rate, and wd is known as weight decay. And momentum will help train faster.

eta = 0.1
momentum = 0.9
wd  = 0.0

Alternatively, we can set parameters specifically for bias and weight connections, using following configs.

wmat:eta = 0.1
bias:eta = 0.2
momentum = 0.9
wd  = 0.0

Here we set 0.1 learning rate for weight connections, and 0.2 learning rate for bias. Adding prefix wmat: to any parameter settings and they will be only applied to weight connections. While bias: stands for update parameters of bias

Metric method

For classification, we use error to metric performance. So set

metric = error

Running experiment

Make a folder named "models" for saving the training model in the same folder you calling cxxnet_learner
Then just run

../../bin/cxxnet ./MNIST.conf

or

./run.sh ./MNIST.conf

Then you will get a nearly 98% correct result in just several seconds.

Do prediction

Add these info in the configure file MNIST.conf

# Data iterator setting
pred = pred.txt
iter = mnist
    path_img = "./data/t10k-images-idx3-ubyte.gz"
    path_label = "./data/t10k-labels-idx1-ubyte.gz"
iter = end
# Global Setting
task = pred
model_in = ./models/0014.model

run

../../bin/cxxnet MNIST.conf

The prediction result will be stored in the pred.txt

Continue Training

===

Now we get a model file models/0014.model. Add these configuration into the original configuration file:

max_round = 3
model_in = ./models/0014.model

Then run ../../bin/cxxnet MNIST.conf, it will continue train extra 3 round based on previous model. If use continue=1 instead of model_in, cxxnet will search the model folder and use the model file with largest number.

Convolution Example

=== Reference configuration example MNIST_CONV.conf

Use convolution layer is easy in cxxnet. Based on previous configuration, make the following changes

• Data Iterator, set input_flat = 0 to make input in 3D: (1,28,28)

# Data iterator setting
data = train
iter = mnist
    path_img = "./data/train-images-idx3-ubyte.gz"
    path_label = "./data/train-labels-idx1-ubyte.gz"
    input_flat = 0
    shuffle = 1
iter = end

eval = test
iter = mnist
    input_flat = 0
    path_img = "./data/t10k-images-idx3-ubyte.gz"
    path_label = "./data/t10k-labels-idx1-ubyte.gz"
iter = end

• Update a convolution network structure

# Network structure setting
netconfig=start
layer[0->1] = conv:cv1
  kernel_size = 3
  pad = 1
  stride = 2
  nchannel = 32
  random_type = xavier
layer[1->2] = max_pooling
  kernel_size = 3
  stride = 2
layer[2->3] = flatten
layer[3->3] = dropout
  threshold = 0.5
layer[3->4] = fullc:fc1
  nhidden = 100
  init_sigma = 0.01
layer[4->5] = sigmoid:se1
layer[5->6] = fullc:fc1
  nhidden = 10
  init_sigma = 0.01
layer[6->6] = softmax
netconfig=end

• Set trainer input shape

input_shape = 1,28,28

• Try to use GPU train the conv net, set dev=gpu

dev = gpu

• Then run

../../bin/cxxnet ./MNIST_CONV.conf

or

./run.sh ./MNIST_CONV.conf

You will get a result near 99% in a few seconds.