In recent years, deep learning is a very hot technology among machine learning, and it is a field where technological progress is remarkable as new methods are proposed every year. Therefore, there must be many people who want to learn deep learning or want to build a deep learning model.
However, it is also true that building a deep learning model from scratch is a high hurdle. Therefore, Skill Up AI has opened a course where you can systematically learn how to build a deep learning model using PyTorch.
In this course, you can learn about building a deep learning model using PyTorch from data acquisition to learning, and even beginners can easily build a deep learning model.
In this article, we will introduce the PyTorch course while explaining PyTorch in the following flow.
Contents
- What is PyTorch?
- Features of PyTorch
- What you can do with PyTorch
- Building image and time series models with PyTorch
- Summary and Introduction to Pytorch Learning Course
1. What is PyTorch?
PyTorch is a deep learning library for Python developed by Facebook. Compared to other deep learning libraries, PyTorch excels in ease of use, code readability, and implementation flexibility.
There are various deep learning libraries for Python, but PyTorch is particularly popular from the above perspective. Therefore, the shortest way to tackle deep learning in Python is to implement it with PyTorch.
In practice, a library called Tensorflow is often used in addition to PyTorch , but Tensorflow has a long history and although there are many reference materials, it is inferior to late-developed libraries in terms of code readability.
On the other hand, PyTorch is superior in readability and ease of debugging (detection of coding mistakes). Another advantage of PyTorch is that it is popular among researchers and it is easy to find implementations of techniques published in recent papers. Against this background, PyTorch has grown in popularity in recent years.
As a comparison of popularity, the following table compares Tensorflow and PyTorch in terms of Google Trends, which allows comparison of the number of searches, and the number of repositories on GitHub, where many open source software are posted.
| Popularity on Google Trends* (2021/08/08 ~ 14) |
Number of related repositories on GitHub (as of 2021/08/24) |
|
|---|---|---|
| Tensorflow | 47% | 111,503 |
| PyTorch | 48% | 70,954 |
* Calculated as a relative value, with the number of searches for the keyword in the week with the highest number of searches since 2018/12/2 as 100%.
As you can see, the popularity of Google Trends is currently very thin. The number of related repositories on GitHub is slightly less for PyTorch, which is a latecomer, but this is also closing the gap year by year.
2. Features of PyTorch
PyTorch is characterized by a “Define-by-Run” design philosophy that builds a dynamic computational graph. While preparing the nodes of the neural network one by one, compute them and connect the nodes at the same time. On the other hand, Tensorflow is the exact opposite of PyTorch, building a static computation graph that writes the network from input to output and then computes.
One of the advantages of using a dynamic computation graph like PyTorch is that it is easy to debug because you code while checking the output for each node. In addition, since the inputs and outputs of networks can be described in detail, complex computational graphs can be easily created.
Therefore, it is also an advantage that it is easy to implement complicated methods such as many network branches. Due to these advantages, even complex models that change processing depending on conditions can be created intuitively, so when the author’s implementation of an international conference paper is published, there are many cases where he uses PyTorch.
Therefore, by acquiring knowledge of PyTorch, you will be able to use state-of-the-art methods quickly.
3. What you can do with PyTorch
PyTorch uses a unique data type called Tensor to handle data and parameters. Tensor types are handled more like Numpy (the standard library for data analysis in Python) variable types. Many Python users use Numpy on a regular basis, so coding in PyTorch is intuitive for most of us.
In addition, since calculations can be performed on the GPU, it is possible to build deep learning models that require a large amount of matrix calculations.
In addition, PyTorch is equipped with modules necessary for automatic differentiation, loss optimization, and other deep learning, making it easy to build models.
In this course, you will learn a series of flows such as data loading, model building, learning, and evaluation through hands-on. PyTorch may require peculiar notations, such as retrieving data in minibatches or backpropagation (a method of learning neural networks). These are one of the stumbling points for beginners.
In the course, we carefully explain the specifications of these modules through hands-on so that you can proceed with the implementation without stumbling.
One of the features of Skill Up AI’s PyTorch course is that it not only explains how to use classes and methods but also explains the principles of deep learning in an easy-to-understand manner. For example, for the gradient descent method used for learning, we will first introduce an overview of the method as shown in the figure below.
Next, I will introduce the implementation method in PyTorch so as to correspond to the outline of the method.
In this way, by introducing how to use PyTorch in a way that goes back and forth between the outline of the method and the implementation, we have devised it so that you can easily get an image of the implementation. In addition, the sample code by Jupyter Notebook is commented almost every line, which is convenient for reviewing later. In the video, not only the slides but also the sample code are carefully explained, and one of the features is that you can see and understand the whole picture from model definition to evaluation using PyTorch.
4. Build image and time series model with PyTorch
Conventional technologies that handle images and natural language, such as face recognition and machine translation, have limited model accuracy due to the large amount of data features. However, with the development of deep learning methods, it has evolved rapidly in recent years. One of the reasons why PyTorch is a positive choice for such complex tasks that require a lot of time to learn is its fast learning speed. According to the paper ” PyTorch: An Imperative Style, High-Performance Deep Learning Library ” published by Facebook and NVIDIA researchers comparing the speed of other deep learning frameworks such as PyTorch and Tensorflow, it can be processed per second. The following table compares the number of data.
Comparison of average number of data processed per second during training
| VGG-19 (image recognition model) |
ResNet-50 (image recognition model) |
GNMT (machine translation model) |
|
|---|---|---|---|
| Tensorflow | 1422 | 66 | 9631 |
| PyTorch | 1547 | 119 | 15512 |
(Partially quoted from ” PyTorch: An Imperative Style, High-Performance Deep Learning Library “)
In any case, PyTorch can process more data per second than Tensorflow, and it is expected that training can be performed faster.
In the PyTorch course, we will build networks such as CNN for image feature extraction and RNN for processing time series and natural language data, and actually implement image processing and time series data processing. These methods are essential items in the current deep learning field. By learning how to build a model and the actual flow through the course, it becomes possible to implement various applied methods.
5. Summary and Introduction to Pytorch Learning Course
In this blog, I explained the recent popularity and characteristics of PyTorch, the reason why it is being applied to images and time-series data, and touched on the features of the skill-up AI PyTorch course.
The Pytorch course is especially recommended for those who want to work on deep learning but do not know where to start, or for those who have a vague understanding of deep learning but find it difficult to connect it to actual coding.
It is also recommended for those who have passed the Deep Learning Test E qualification and want to implement a cutting-edge deep learning model in the future. At the moment, the E qualification is only for his NumPy implementation, so many people may think that even if they have already passed the E qualification, it is not enough. If you are like that, let’s improve his level further with the PyTorch course!
This course is an interactive course where you can experience a series of deep learning modeling processes while moving the code hands-on. If you are interested, please take a look at the course page.
