Review: Udacity’s Intro to Deep Learning with PyTorch

Introduction

Recently, I wanted to learn more about PyTorch for building neural networks. I had dabbled briefly, but not enough to have any level of competency. My existing experience is with TensorFlow and Darknet. I have heard from a number of sources that PyTorch is really impressive. Additionally, I found that they are one of the few popular networks that offers an easy to use C++ API (I use C++ for a number of personal projects). Given that, it seemed worth exploring what working with PyTorch is like.

PyTorch offers a number of tutorials on their site. However, I vaguely remember these tutorials not quite fitting what I was looking for last time I checked. To be fair, I was only dabbling and looked very briefly. They are probably very helpful and worth checking out, so don’t necessarily take my word on those.

In addition to PyTorch’s tutorials, I recently also found Udacity offered a course called Intro to Deep Learning with PyTorch. Based on the syllabus, this seemed like a promising course to cover what I wanted. It starts with an introduction to neural networks and works up through deep networks, convolutional neural networks (CNNs), recurrent neural networks (RNNs), and long short term memory (LSTM) networks. Each features implementation examples using PyTorch, so offers a great way to get up to speed on using the framework. Additionally, Udacity courses tend to have lots of exercises and quizzes throughout to make sure you understand the content. Lastly, this particular course happened to be free (at least as of the time of writing). Therefore, it seemed like a worthwhile course to explore.

After completing the course, I wrote up this post to share my experience. I quickly summarize the course content. I also include my thoughts about the course, including what it did well and not so well. Lastly, I include a brief summary of the overall usefulness of the course.

Course Content

The course contains nine sections, including a “Welcome to the course” section. Udacity estimated each section to take anywhere between 30 minutes to 5 hours. However, the 5 hour estimates also include time to complete all the coding exercises. Most sections consist of a bunch of videos interspersed with coding and quizzes.

Introduction to Neural Networks

This section presents an overview of neural network concepts. This includes logistic regression, sigmoid and softmax functions, loss functions, backpropagation, and more. It does not cover any specific network yet, but lays the concepts and mathematics that make up the foundation of neural networks.

Talking PyTorch with Soumith Chintala

This is an interview with the creator of PyTorch. It offers a history of the framework, some of its driving philosophy, and a bit about the future of PyTorch. It isn’t necessary to understand how to use it, but offers some nice context.

Introduction PyTorch

This section starts to dive into using PyTorch. It has you create a deep multi-layer perceptron network for classifying images from the FashionMNIST dataset. As part of this, it teaches you how to prepare data, define the model, train the model to learn the weights, and evaluate end accuracy. There is also a quick lesson on transfer learning.

Convolutional Neural Networks

From there, the course then walks you through the theory and application of CNNs. It first defines how convolutional layers work. It then guides you in building a CNN to classify the same FashionMNIST dataset from before. By doing this, you can compare the accuracy of both networks.

Style Transfer

This section is a quick aside into using CNNs for style transfer. Style transfer modifies one image to match the style of another. In performing this task, you learn how to manipulate a PyTorch model at a more detailed level than before. You are required to access intermediate output of several layers throughout the network, as well as freeze the weights so that backpropagation modifies the input image instead of the network.

Recurrent Neural Networks

Next, the course teaches you the theory behind RNNs and LSTMs. You demonstrate this knowledge by building a LSTM network to perform text prediction. Given an arbitrary series of characters input into the network, it tries to predict what the best next character should be.

Sentiment Prediction RNNs

The penultimate section serves as a capstone project of sorts. It walks you through an end-to-end example on building an RNN for sentiment classification of movie reviews. This includes a number of steps to pre-process the data, such as converting words to numbers and padding to a specified input length. After processing, you then build a network and train on your processed data. You then evaluate the results on a test set of data. While there is significant work involved, the instructor reviews solutions after each portion of the code.

Deploying PyTorch Models

The final section is very short. It covers how to build a network in Python and export it for use in a C++ environment. This is useful for situations where production requirements just aren’t met by Python. This section walks through the tutorial available on PyTorch’s website called Loading a TorchScript Model in C++.

Overall Thoughts

After working through all the sections, there were some aspects of the course that I liked and some that I did not care for. Overall, I definitely enjoyed the course, although I felt it is better suited for those newer to neural networks.

Pros

1. As with any Udacity course, there is a plethora of quizzes and coding examples throughout the lessons. This gave me a great opportunity to practice the material right after learning the concept.
2. Likewise, the exercises are really well put together. They follow the familiar format of template code that requires you to finish some sections. However, unlike other courses I have seen, the surrounding code is well explained. In fact, I was able to easily modify and reimplement it on my own to conduct some side exploration into automatic hyperparameter tuning.
3. The initial section on neural networks was really well done. They related the concepts very well and used a number of great analogies to drive home the ideas. Additionally, there is just the right amount of math. If someone wasn’t interested in the math, they could safely gloss over it without being too lost (at least until implementing some functions). However, for someone like me who needs to work through all the equations to feel like they have a solid grasp, there is enough to run with and understand.
4. The course has great production value. All the narrations are clear and the graphics were related and understandable. Many sections offered links to related reading that covers the certain topics in more detail. Also, the code is fairly well documented!

Cons

1. I think this course was pulled together from a number of Udacity’s other course. At times, they make references to Keras in the videos. Additionally, they repeat certain concepts in different sections, and not in the “remember from last time” sort of way. It is a free course offered on a site that usually requires payment for their courses though, so still well worth the value.
2. Not all of the exercises can be done online in their classroom workspace. They mention the ability to use the workspace at the beginning, but most of the later material requires cloning their Git repository. This isn’t a huge issue, but I don’t have a good GPU on my computer, so I had to significantly limit training on the networks and thus didn’t get nearly the accuracy as the instructors.
3. The C++ section at the end essentially just reads the online tutorial it references, so there isn’t really any value added. As this wasn’t the primary focus of the course, I can give them a pass on this.

Takeaway

Overall, I liked this course. I would consider it a practical, introductory guide to neural networks in general and PyTorch specifically. It provides a good introduction to the fundamentals of both. If you are just looking for a primer on using PyTorch and are already well-versed in neural networks, this course will be overkill for you. However, if you are just getting started and want to dive in, this is an excellent resource to use.