dts3/LDL

Learning Deep Learning (Nvidia)

LDL

This code repository contains code examples associated with the book Learning Deep Learning (LDL) by Magnus Ekman (ISBN: 9780137470358).

Related web sites:

• Publisher (Pearson) site with supplemental material and links to buy LDL: http://informit.com/title/9780137470358
• Author's site: http://ldlbook.com
• NVIDIA Deep Learning Institute (DLI) book site: http://www.nvidia.com/dli-books

LDL can also be purchased at Amazon.com: https://www.amazon.com/Learning-Deep-Practice-Transformers-TensorFlow/dp/0137470355

Some of the code examples rely on datasets that are not in the repository itself. This file provides all information needed to obtain these datasets and to run the programming examples. Further, the code examples are well documented in the Jupyter notebook versions. However, the purpose of the examples are to illustrate concepts taught in the LDL book. The examples should be considered in that context, and are best consumed in conjunction with reading the book.

Docker Files

The repository contains two Docker files to simplify running the code examples under Docker:

• Dockerfile_tf - for the TensorFlow versions of the code examples
• Dockerfile_pt - for the PyTorch versions of the code examples

See the author's site for blog posts describing how to set up and run Docker on Linux and Windows: http://ldlbook.com

Code Examples

The code examples can be divided into three categories in the following three directories:

• stand_alone - stand-alone examples not relying on a Deep Learning (DL) framework
• tf_framework - examples based on the TensorFlow DL framework
• pt_framework - examples based on the PyTorch DL framework

There is a one-to-one mapping between the code examples in the tf_framework and pt_framework directory. Pick a framework of your choice or learn both!

The initial versions of these programming examples were tested with versions 2.4 and 2.5 of TensorFlow and versions 1.8.0 and 1.9.0 of PyTorch. The most recent versions have been tested with version 2.9.1 of TensorFlow and version 1.12.1 of PyTorch. TensorFlow is sometimes rather verbose when using GPU acceleration. To make it less verbose, set the environment variable TF_CPP_MIN_LOG_LEVEL to the value 2. If you are using bash, this can be done with export TF_CPP_MIN_LOG_LEVEL=2.

The naming of each code example follows the pattern cXeY_DESCRIPTION.py where X represents the chapter number, Y the example number in that chapter, and DESCRIPTION is a brief description of what the example is doing. The examples named aFeY_DESCRIPTION.py are not from a regular chapter but from Appendix F.

Apart from the three directories containing code examples, there is a single directory named data that is supposed to contain datasets needed by some of the code examples. The repository contains some of these assets but the user needs to download additional datasets to fully populate it. Instructions to that is found in the section Datasets below.

Each code example is expected to be run from within the directory where the code example is located, as it uses a relative path to access the dataset. That is, you first need to change to the stand_alone directory before running code examples located in that directory.

Because of the stochastic nature of DL algorithms, the results may vary from run to run. That is, it is expected that your results will not exactly reproduce the results stated in the book.

Alternative Versions

Some of the code examples have alternative versions to work around issues observed on some platforms. This applies to the following code examples:

• tf_framework/c11_e1_autocomplete_no_rdo - this version does not use recurrent dropout, which causes hangs on some platforms
• tf_framework/c12_e1_autocomplete_embeddin_no_rdo - this version does not use recurrent dropout, which causes hangs on some platforms
• tf_framework/c17_e4_nas_random_hill_multiprocess - this version spawns multiple processes, which works around a memory leak problem observed on some platforms
• tf_framework/c17_e5_nas_evolution_multiprocess - this version spawns multiple processes, which works around a memory leak problem observed on some platforms

Supporting Spreadsheet

Apart from the code examples, this repository also contains a spreadsheet named network_example.xlsx. The spreadsheet provides additional insight into the basic workings of neurons and the learning process. It is unlikely that this spreadsheet is useful without first reading the corresponding description in LDL.

The spreadsheet consists of three tabs, each corresponding to a specific section of the initial chapters:

• perceptron_learning corresponds to the section The Perceptron Learning Algorithm in Chapter 1, The Rosenblatt Perceptron.
• backprop_learning corresponds to the section Using Backpropagation to Compute the Gradient in Chapter 3, Sigmoid Neurons and Backpropagation.
• xor_example corresponds to the section Programming Example: Learning the XOR Function in Chapter 3.

Datasets

Some of the programming examples rely on datasets accessible through the DL framework but others need to be downloaded and placed in the appropriate location. This section describes how to obtain the ones that need to be downloaded. All program examples assume that the downloaded datasets are placed in the directory named data in the root of the code example directory tree.

MNIST

The MNIST Database of handwritten digits can be obtained from http://yann.lecun.com/exdb/mnist.

Download the following files:

• train-images-idx3-ubyte.gz
• train-labels-idx1-ubyte.gz
• t10k-images-idx3-ubyt.gz
• t10k-labels-idx1-ubyte.gz

Once downloaded, gunzip them to the data/mnist/ directory. You need the Python package idx2numpy to use this version of the MNIST dataset.

BOOKSTORE SALES DATA FROM US CENSUS BUREAU

Sales data from the United States Census Bureau can be obtained from https://www.census.gov/econ/currentdata.

Select Monthly Retail Trade and Food Services and click the Submit button. That should take you to a page where you need to specify five different steps. Select:

• Monthly Retail Trade and Food Services
• Start: 1992 End: 2020
• 451211: Book Stores
• Sales - Monthly
• U.S. Total

Make sure that the checkbox Not Seasonally Adjusted is checked. Then click the GET DATA button. That should result in a table with data values. Download it to a comma-separated values (CSV) file by clicking the link TXT. Remove the first few lines in the downloaded CSV file so the file starts with a single line containing headings saying "Period,Value" followed by one line for each month. Further, remove any lines with non-numerical values, such as "NA", at the end of the file. Name the file book_store_sales.csv and copy to the data directory.

FRANKENSTEIN FROM PROJECT GUTENBERG

The text for Mary Shelley's Frankenstein can be downloaded from https://www.gutenberg.org/files/84/84-0.txt. Rename the file to frankenstein.txt and copy to the data directory.

GloVe WORD EMBEDDINGS

The GloVe word embeddings file, which is close to 1 GB in size, can be downloaded from http://nlp.stanford.edu/data/glove.6B.zip. Unzip it after downloading and copy the file glove.6B.100d.txt to the data directory.

ANKI BILINGUAL SENTENCE PAIRS

The Anki bilingual sentence pairs can be downloaded from http://www.manythings.org/anki/fra-eng.zip. Unzip it after download and copy the file fra.txt to the data directory.

COCO

Create a directory named coco inside of the data directory. Download the following file: http://images.cocodataset.org/annotations/annotations_trainval2014.zip. Unzip it and copy the file captions_train2014.json to the directory coco. Download the following 13 GB file: http://images.cocodataset.org/zips/train2014.zip. Unzip it into the data/coco/ directory so the path to the unzipped directory is data/coco/train2014/.