As we have already discussed in previous chapters that, MXNet Gluon provides a clear, concise, and simple API for DL projects. It enables Apache MXNet to prototype, build, and train DL models without forfeiting the training speed.

Core Modules

Let us learn the core modules of Apache MXNet Python application programming interface (API) gluon.

gluon.nn

Gluon provides a large number of build-in NN layers in gluon.nn module. That is the reason it is called the core module.

Methods and their parameters

Following are some of the important methods and their parameters covered by mxnet.gluon.nn core module −

Implementation Examples

In the example below, we are going to use Block() which gives the base class for all neural network layers and models.

from mxnet.gluon import Block, nn
class Model(Block):
   def __init__(self, **kwargs):
      super(Model, self).__init__(**kwargs)
      # use name_scope to give child Blocks appropriate names.
      with self.name_scope():
         self.dense0 = nn.Dense(20)
         self.dense1 = nn.Dense(20)
   def forward(self, x):

      x = mx.nd.relu(self.dense0(x))
      return mx.nd.relu(self.dense1(x))

model = Model()
model.initialize(ctx=mx.cpu(0))
model(mx.nd.zeros((5, 5), ctx=mx.cpu(0)))

Output

You will see the following output −

[[0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.]
 [0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.]
 [0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.]
 [0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.]
 [0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.]]
<NDArray 5x20 @cpu(0)*gt;

In the example below, we are going to use HybridBlock() that supports forwarding with both Symbol and NDArray.

import mxnet as mx
from mxnet.gluon import HybridBlock, nn


class Model(HybridBlock):
   def __init__(self, **kwargs):
      super(Model, self).__init__(**kwargs)
      # use name_scope to give child Blocks appropriate names.
      with self.name_scope():
         self.dense0 = nn.Dense(20)
         self.dense1 = nn.Dense(20)

   def forward(self, x):
      x = nd.relu(self.dense0(x))
      return nd.relu(self.dense1(x))
model = Model()
model.initialize(ctx=mx.cpu(0))

model.hybridize()
model(mx.nd.zeros((5, 5), ctx=mx.cpu(0)))

Output

The output is mentioned below −

[[0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.]
 [0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.]
 [0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.]
 [0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.]
 [0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0. 0.]]
<NDArray 5x20 @cpu(0)>

gluon.rnn

Gluon provides a large number of build-in recurrent neural network (RNN) layers in gluon.rnn module. That is the reason, it is called the core module.

Methods and their parameters

Following are some of the important methods and their parameters covered by mxnet.gluon.nn core module:

Implementation Examples

In the example below, we are going to use GRU() which applies a multi-layer gated recurrent unit (GRU) RNN to a given input sequence.

layer = mx.gluon.rnn.GRU(100, 3)
layer.initialize()
input_seq = mx.nd.random.uniform(shape=(5, 3, 10))
out_seq = layer(input_seq)
h0 = mx.nd.random.uniform(shape=(3, 3, 100))
out_seq, hn = layer(input_seq, h0)
out_seq

Output

This produces the following output −

[[[ 1.50152072e-01 5.19012511e-01 1.02390535e-01 ... 4.35803324e-01
1.30406499e-01 3.30152437e-02]
[ 2.91542172e-01 1.02243155e-01 1.73325196e-01 ... 5.65296151e-02
1.76546033e-02 1.66693389e-01]
[ 2.22257316e-01 3.76294643e-01 2.11277917e-01 ... 2.28903517e-01
3.43954474e-01 1.52770668e-01]]


[[ 1.40634328e-01 2.93247789e-01 5.50393537e-02 ... 2.30207980e-01
6.61415309e-02 2.70989928e-02]
[ 1.11081995e-01 7.20834285e-02 1.08342394e-01 ... 2.28330195e-02
6.79589901e-03 1.25501186e-01]
[ 1.15944080e-01 2.41565228e-01 1.18612610e-01 ... 1.14908054e-01
1.61080107e-01 1.15969211e-01]]
………………………….

Example

hn

Output

This produces the following output −

[[[-6.08105101e-02 3.86217088e-02   6.64453954e-03 8.18805695e-02
3.85607071e-02 -1.36945639e-02 7.45836645e-03 -5.46515081e-03
9.49622393e-02 6.39371723e-02 -6.37890724e-03 3.82240303e-02
9.11015049e-02 -2.01375950e-02 -7.29381144e-02 6.93765879e-02
2.71829776e-02 -6.64435029e-02 -8.45306814e-02 -1.03075653e-01
6.72040805e-02 -7.06537142e-02 -3.93818803e-02 5.16211614e-03
-4.79770005e-02 1.10734522e-01 1.56721435e-02 -6.93409378e-03
1.16915874e-01 -7.95962065e-02 -3.06530762e-02 8.42394680e-02
7.60370195e-02 2.17055440e-01 9.85361822e-03 1.16660878e-01
4.08297703e-02 1.24978097e-02 8.25245082e-02 2.28673983e-02
-7.88266212e-02 -8.04114193e-02 9.28791538e-02 -5.70827350e-03
-4.46166918e-02 -6.41122833e-02 1.80885363e-02 -2.37745279e-03
4.37298454e-02 1.28888980e-01 -3.07202265e-02 2.50503756e-02
4.00907174e-02 3.37077095e-03 -1.78839862e-02 8.90695080e-02
6.30150884e-02 1.11416787e-01 2.12221760e-02 -1.13236710e-01
5.39616570e-02 7.80710578e-02 -2.28817668e-02 1.92073174e-02
………………………….

In the example below we are going to use LSTM() which applies a long-short term memory (LSTM) RNN to a given input sequence.

layer = mx.gluon.rnn.LSTM(100, 3)
layer.initialize()

input_seq = mx.nd.random.uniform(shape=(5, 3, 10))
out_seq = layer(input_seq)
h0 = mx.nd.random.uniform(shape=(3, 3, 100))
c0 = mx.nd.random.uniform(shape=(3, 3, 100))
out_seq, hn = layer(input_seq,[h0,c0])
out_seq

Output

The output is mentioned below −

[[[ 9.00025964e-02 3.96071747e-02 1.83841765e-01 ... 3.95872220e-02
1.25569820e-01 2.15555862e-01]
[ 1.55962542e-01 -3.10300849e-02 1.76772922e-01 ... 1.92474753e-01
2.30574399e-01 2.81707942e-02]
[ 7.83204585e-02 6.53361529e-03 1.27262697e-01 ... 9.97719541e-02
1.28254429e-01 7.55299702e-02]]
[[ 4.41036932e-02 1.35250352e-02 9.87644792e-02 ... 5.89378644e-03
5.23949116e-02 1.00922674e-01]
[ 8.59075040e-02 -1.67027581e-02 9.69351009e-02 ... 1.17763653e-01
9.71239135e-02 2.25218050e-02]
[ 4.34580036e-02 7.62207608e-04 6.37005866e-02 ... 6.14888743e-02
5.96345589e-02 4.72368896e-02]]
……………

Example

hn

Output

When you run the code, you will see the following output −

[
[[[ 2.21408084e-02 1.42750628e-02 9.53067932e-03 -1.22849066e-02
1.78788435e-02 5.99269159e-02 5.65306023e-02 6.42553642e-02
6.56616641e-03 9.80876666e-03 -1.15729487e-02 5.98640442e-02
-7.21173314e-03 -2.78371759e-02 -1.90690923e-02 2.21447181e-02
8.38765781e-03 -1.38521893e-02 -9.06938594e-03 1.21346042e-02
6.06449470e-02 -3.77471633e-02 5.65885007e-02 6.63008019e-02
-7.34188128e-03 6.46054149e-02 3.19911093e-02 4.11194898e-02
4.43960279e-02 4.92892228e-02 1.74766723e-02 3.40303481e-02
-5.23341820e-03 2.68163737e-02 -9.43402853e-03 -4.11836170e-02
1.55221792e-02 -5.05655073e-02 4.24557598e-03 -3.40388380e-02
……………………

Training Modules

The training modules in Gluon are as follows −

gluon.loss

In mxnet.gluon.loss module, Gluon provides pre-defined loss function. Basically, it has the losses for training neural network. That is the reason it is called the training module.

Methods and their parameters

Following are some of the important methods and their parameters covered by mxnet.gluon.loss training module:

Example

As we know that mxnet.gluon.loss.loss will calculate the MSE(Mean Squared Error) between label and prediction (pred). It is done with the help of following formula:

gluon.parameter

mxnet.gluon.parameter is a container that holds the parameters i.e. weights of the Blocks.

Methods and their parameters

Following are some of the important methods and their parameters covered by mxnet.gluon.parameter training module −

Implementation Example

In the example below, we will initialize parameters and the gradients arrays by using initialize() method as follows −

weight = mx.gluon.Parameter(''weight'', shape=(2, 2))
weight.initialize(ctx=mx.cpu(0))
weight.data()

Output

The output is mentioned below −

[[-0.0256899 0.06511251]
[-0.00243821 -0.00123186]]
<NDArray 2x2 @cpu(0)>

Example

weight.grad()

Output

The output is given below −

[[0. 0.]
[0. 0.]]
<NDArray 2x2 @cpu(0)>

Example

weight.initialize(ctx=[mx.gpu(0), mx.gpu(1)])
weight.data(mx.gpu(0))

Output

You will see the following output −

[[-0.00873779 -0.02834515]
 [ 0.05484822 -0.06206018]]
<NDArray 2x2 @gpu(0)>

Example

weight.data(mx.gpu(1))

Output

When you execute the above code, you should see the following output −

[[-0.00873779 -0.02834515]
 [ 0.05484822 -0.06206018]]
<NDArray 2x2 @gpu(1)>

gluon.trainer

mxnet.gluon.trainer applies an Optimizer on a set of parameters. It should be used together with autograd.

Methods and their parameters

Following are some of the important methods and their parameters covered by mxnet.gluon.trainer training module −

Data Modules

The data modules of Gluon are explained below −

gluon.data

Gluon provides a large number of build-in dataset utilities in gluon.data module. That is the reason it is called the data module.

Classes and their parameters

Following are some of the important methods and their parameters covered by mxnet.gluon.data core module. These methods are typically related to Datasets, Sampling, and DataLoader.

Implementation Examples

In the example below, we are going to use gluon.data.BatchSampler() API, which wraps over another sampler. It returns the mini batches of samples.

import mxnet as mx
from mxnet.gluon import data
sampler = mx.gluon.data.SequentialSampler(15)
batch_sampler = mx.gluon.data.BatchSampler(sampler, 4, ''keep'')
list(batch_sampler)

Output

The output is mentioned below −

[[0, 1, 2, 3], [4, 5, 6, 7], [8, 9, 10, 11], [12, 13, 14]]

gluon.data.vision.datasets

Gluon provides a large number of pre-defined vision dataset functions in gluon.data.vision.datasets module.

Classes and their parameters

MXNet provides us useful and important datasets, whose classes and parameters are given below −

Example

In the example below, we are going to show the use of ImageListDataset(), which is used for loading image files that are specified by a list of entries −

# written to text file *.lst

0 0 root/cat/0001.jpg
1 0 root/cat/xxxa.jpg
2 0 root/cat/yyyb.jpg
3 1 root/dog/123.jpg
4 1 root/dog/023.jpg
5 1 root/dog/wwww.jpg

# A pure list, each item is a list [imagelabel: float or list of float, imgpath]

[[0, root/cat/0001.jpg]
[0, root/cat/xxxa.jpg]
[0, root/cat/yyyb.jpg]
[1, root/dog/123.jpg]
[1, root/dog/023.jpg]
[1, root/dog/wwww.jpg]]

Utility Modules

The utility modules in Gluon are as follows −

gluon.utils

Gluon provides a large number of build-in parallelisation utility optimiser in gluon.utils module. It provides variety of utilities for training. That is the reason it is called the utility module.

Functions and their parameters

Following are the functions and their parameters consisting in this utility module named gluon.utils −