Errata

My version is for Kindle.

Figure 1-6 renders a map of the world. There are two points labelled B, one in red, coherent with the text, and a black B, which I believe is a mistake.

The second convolution calculation result should be 0.3 instead of the -0.1 result at the right of the frame.

The map erroneously shows two points labeled as B. In order to be consistent with the text:
- Point B located in Alaska should remain in the picture
- Duplicate Point B located in the middle of Asia should be removed

(1st Edition, 1st Release)

Combination 1 of 16 page appears once in table 1-1, not twice.

I suggest to change ShortHairShortFlat to LongHairSTraight in topType of which face_id is 0.

(1st Edition, 1st Release)

"where \hat{\theta_{kl}} is the number of times...."

should be

"where n_{kl} is the number of times...."

You need to cd into the GDL_code directory before entering the git pull command.

It is claimed that neural networks work best when the input is in the range -1 to 1. However, the provided source code keeps the input in the range 0 to 1. It is confusing why the provided source code does not result in the inputs to be in the range that claims to work best with neural networks.

In the code section of Example 2.1 building the network using Sequential model, Flatten() should come first before Dense(200)?
This would match with Example 2.2 that uses Functional API.

At the very end of the code listing, there's an ">" which should not be there.

(1st Edition, 1st Release)

In categorical corss-entropy eq.,

-\sum^n_{i=1} y_i log(p_i)

should be

-\sum^n_{i=1} \sum^C_{k=1} y^k_i log(p^k_i)

(1st Edition, 1st Release)

"...a rate of approximately 160 microseconds"

should be

"...a rate of approximately 8 seconds"

It seems that the part "8s" in the output shown in Figure 2-7 was neglected. To be more precise, the 7th epoch took more than 10 seconds, and it took about 8 seconds 362 microseconds on average.

the point [-3.5, -0.5]

seems to be

point [-3.5, 0.5]

in the Figure3-1

The example gives y = x_train as an argument to the model's fit function instead of y = y_test.

The text describes the choice of the loss function as the Root Mean Square (RMSE) error. However, the formula implemented in example 3-5 (and also in the models/AE.py code) is an implementation of the Mean Square Error (i.e., it is missing a square root)

This will train and do something, but it is a different function, not the one stated.

(1st Edition, 1st Release)

In KL loss eq.,

= 1/2 \sum ...

should be

= - 1/2 \sum ...

Unbalanced parentheses on the left-hand side of the second equation: inside the brackets, it's written as

N(μ, σ || N(0,1)

but it must be:

N(μ, σ) || N(0,1)

It should be clarified that a separate download of the feature file list_attr_celeba.csv is required. Furthermore, the image id in the feature file has to be changed from File_name to image_id otherwise the code in 03_06_vae_faces_analysis throws a KeyError.

> We pass this through four Conv2D layers, the first two preceded by Upsampling2D layers

I think that Each Conv2D is preceded by UpSampling2D layer in the code.

It says "the model has three outputs: 1..., -1,... and dummy 0 vector.

But the code in example 4.10 shows as the third option validity interpolated, which is not a 0 vector.

(1st Edition, 1st Release)

8 x 8 single-channel

should be

16 x 16 single-channel

2**4 should be 2**3.

> The include_top = False parameter
specifies that we do not need to load the weights for the final dense layers of the
networks that result in the classification of the image.

"include_top" specifies whether to include the 3 fully-connected layers at the top of the network, not to load the weights.

See also https://keras.io/applications/#vgg19

(1st Edition, 1st Release)

Above picture in 249 page, "./vae/vae.json" should be "./vae/weights.h5".

Above picture in 257 page, "./vae/rnn.json" should be "./rnn/weights.h5".

(1st Edition, 1st Release)

Two variable names is used interchangable. For example, full_model and vae_full.
encoder, decoder, encoder_mu_log_var has same situation.

Please use only a name not to confuse readers.

(1st Edition, 1st Release)

Standard deviation is intialized at 0.1 in the repo., not 0.5.

(1st Edition, 1st Release)

"key-value input" should be "key input".