This document provides an overview of Apache MXNet, an open-source library for deep learning. It discusses MXNet's capabilities such as high performance scaling across GPUs, support for mobile and IoT models, and multiple language syntax. It also demonstrates MXNet through Jupyter notebooks on MNIST data and introduces Gluon, a high-level API for MXNet. Resources for learning more about MXNet, deep learning on AWS, and the presenter's blog are provided.

1. Deep Learning with
Apache MXNet
Julien Simon, AI Evangelist, EMEA
@julsimon

2. What to expect
• Apache MXNet
• Demos using Jupyter notebooks
• Resources
• Q&A

3. Apache MXNet: Open Source library for Deep Learning
Programmable Portable High Performance
Near linear scaling
across hundreds of
GPUs
Highly efficient
models for
mobile
and IoT
Simple syntax,
multiple
languages
Most Open Best On AWS
Optimized for
Deep Learning on AWS
Accepted into the
Apache Incubator

4. CPU or GPU: your choice
mod = mx.mod.Module(lenet)
mod = mx.mod.Module(lenet, context=mx.gpu(0))
mod = mx.mod.Module(lenet,
context=(mx.gpu(7), mx.gpu(8), mx.gpu(9)))

5. Ideal
Inception v3
Resnet
Alexnet
88%
Efficiency
0
64
128
192
256
1 2 4 8 16 32 64 128 256
Multi-GPU Scaling With MXNet

6. © 2017, Amazon Web Services, Inc. or its Affiliates. All rights reserved.
• Storing and accessing data in multi-dimensional arrays
àNDArray API
• Building models (layers, weights, activation functions)
à Symbol API
• Serving data during training and validation
à Iterators
• Training and using models
à Module API
The Apache MXNet API

7. Input Output
1 1 1
1 0 1
0 0 0
3
mx.sym.Convolution(data, kernel=(5,5), num_filter=20)
mx.sym.Pooling(data, pool_type="max", kernel=(2,2),
stride=(2,2)
lstm.lstm_unroll(num_lstm_layer, seq_len, len, num_hidden, num_embed)
4 2
2 0
4=Max
1
3
...
4
0.2
-0.1
...
0.7
mx.sym.FullyConnected(data, num_hidden=128)
2
mx.symbol.Embedding(data, input_dim, output_dim = k)
0.2
-0.1
...
0.7
Queen
4 2
2 0
2=Avg
Input Weights
cos(w, queen) = cos(w, king) - cos(w, man) + cos(w, woman)
mx.sym.Activation(data, act_type="xxxx")
"relu"
"tanh"
"sigmoid"
"softrelu"
Neural Art
Face Search
Image Segmentation
Image Caption
“People Riding
Bikes”
Bicycle, People,
Road, Sport
Image Labels
Image
Video
Speech
Text
“People Riding
Bikes”
Machine Translation
“Οι άνθρωποι
ιππασίας ποδήλατα”
Events
mx.model.FeedForward model.fit
mx.sym.SoftmaxOutput

8. Gluon: Deep Learning gets even easier
https://github.com/gluon-api/
• Available now in MXNet, soon in Microsoft
Cognitive Toolkit
• Developer-friendly high-level API
• Dynamic networks can be modified during training
• No compromise on performance
• Extensive model zoo
https://aws.amazon.com/blogs/aws/introducing-gluon-a-new-library-for-machine-learning-from-aws-and-microsoft/

9. Gluon Model Zoo
VGG
ResNet
AlexNet
DenseNet
SqueezeNet
Inception
MobileNet
https://mxnet.incubator.apache.org/versions/master/api/python/gluon/model_zoo.html

10. © 2017, Amazon Web Services, Inc. or its Affiliates. All rights reserved.
Deep Learning in practice
AWS Deep Learning AMI
MXNet
Torch
CTKKeras
TheanoCaffe
TensorFlow
Amazon EC2
AnacondaIntel MKL
CUDA+cuDNN Python 2+3
Caffe2
• One-click launch
• Single node or distributed
• CPU, GPU, FPGA
• NVIDIA & Intel libraries
• Anaconda Data Science Platform
• Python w/ AI/ML/DL libraries

11. Demos
https://github.com/juliensimon/dlnotebooks
1) Synthetic data set
2) Learn MNIST with a Multi-Layer Perceptron
3) Learn MNIST with the LeNet CNN
4) Predict handmade MNIST samples

12. © 2017, Amazon Web Services, Inc. or its Affiliates. All rights reserved.
The MNIST data set
• 70,000 hand-written digits
• 28x28 grayscale images

13. © 2017, Amazon Web Services, Inc. or its Affiliates. All rights reserved.
• MLP does well on MNIST, but not on real-life samples.
• MLP flattens the 28x28 image into a 784-byte vector
à the 2-D relationship between pixels is lost.
• Look at the ‘5’: the top-left angle is a unique feature. No other
digit exhibits this. Same thing for the intersection on the ‘8’.
• CNNs such as LeNet work on 2-D images and learn to detect
these unique geometric features.
• The result is a more accurate and more robust network.
Training models on MNIST

14. © 2017, Amazon Web Services, Inc. or its Affiliates. All rights reserved.
Resources
https://aws.amazon.com/machine-learning
https://aws.amazon.com/blogs/ai
https://mxnet.incubator.apache.org
https://github.com/apache/incubator-mxnet
https://github.com/gluon-api
https://medium.com/@julsimon

15. Thank you!
Julien Simon, AI Evangelist, EMEA
@julsimon