Errata

(p71) In footnote,

sklearn.set_config(pandas_in_out=True) --> sklearn.set_config(transform_output="pandas")

Thank you! :)

>>> y_train_pred_forest = y_probas_forest[:, 1] >= 0.5 # positive proba ≥ 50%
>>> f1_score(y_train_5, y_pred_forest)

Instead of:

>>> y_train_pred_forest = y_probas_forest[:, 1] >= 0.5 # positive proba ≥ 50%
>>> f1_score(y_train_5, y_train_pred_forest)

I agree with Valerio's Jan 09, 2024 submission, there is a missing open parenthesis in this equation, and I'm assuming (as Valerio did) that it should've been before the h.

The formula for the number of polynomial features should be (n+d)!/(d!n!)-1 instead of (n+d)!/(d!n!). For example, for a polynomial of degree d=2 with n=1 feature x, we get 2 features (x and x^2), not 3. The reason is that we subtract the bias term 1, so the original formula is changed here.

Figure 6-5 was changed in the 3rd edition of the book, now the figure does not fit the text above and Figure 6-4 any more (the x-values do not match).

Figure contents out of sync with Github version

Under the correlation matrix, the paragraph says "...The new bedrooms_ratio is much more correlated with median house value than the total number of rooms or bedrooms..."

This is clearly a slight mistake as the bedroom_ratio is the least correlated to the median house value. What was probably meant instead is the rooms_per_house attribute.

The typo appears in Chapter 3: Classification, specifically on page 104. The error is as follows: In the last paragraph of page 104, there is an issue with the wording. The paragraph currently reads as follows: "One pixel's intensity is from 0 to 255 that's right, but 0 indicates white and 255 indicates black in grayscale map." However, the correct phrasing should be: "The pixel intensity ranges from 0 to 255, where 0 indicates black and 255 indicates white in a grayscale map." This correction clarifies the intended meaning and ensures accurate understanding for readers.

The boundaries showed in the Figure 6-5 are different from what you can find in DT nodes of Figure 6-4. I ran the Python notebook to see if the two images have the same boundaries and in that case everything is correct!

>>>voting_clf.score(X_test, y_test)
fails because the probability of the SVC estimator reverts to False after execution of the fit() method on the VotingClassifier. To make the code work, I had to instantiate the VotingClassifier (code of page 214) with ('svc', SVC(probability=True, random_state=42)).

I tested this several times and indeed voting_clf.named_estimators_["svc"].probability is reset to False after execution of voting_clf.fit(X_train, y_train).

There appears to be a mistake in Equation 7-1 as it is missing a denominator.
Thanks!

The code looks for a 2% threshold, but the paragraph above (twice) says 4.

The text in the Jupyter notebook on GitHub consistently uses 2.

The code `housing = fetch_california_housing()`, and subsequent code `fashion_mnist = tf.keras.datasets.fashion_mnist.load_data()` (on p.318, ~3rd paragraph), will not work in all circumstances. It fails on macos 13.6.4 running Python 3.11.8, with the following error:

URL fetch failure ...certificate verify failed: self-signed certificate in certificate chain (_ssl.c:1006)

This can be fixed by running the following code prior to `housing = fetch_california_housing()`:

import os, ssl
if (not os.environ.get('PYTHONHTTPSVERIFY', '') and
getattr(ssl, '_create_unverified_context', None)):
ssl._create_default_https_context = ssl._create_unverified_context

The code does not work because "decay is deprecated in the new Keras optimizer, pleasecheck the docstring for valid arguments, or use the legacy optimizer, e.g., tf.keras.optimizers.legacy.SGD.".

I understand that Power scheduling is replaced by the PolynomialDecay scheduler, though both are only equivalent as long as t/s << 1.

Probably the rest of the chapter on schedulers needs to be adapted as well to the Learning rate schedules API.

Finally it might be useful to specify that a "(training) step" consists of one iteration on one batch. The terms "step" and "iteration" could benefit from a definition somewhere in the book.

In the build function, the parent's build method is not called (contrary to what the text says). On Github it is called:
super().build(batch_input_shape) # must be at the end

In addition to the error reported earlier (stride = 1 and not 2), the accolade representing fw=3 is one unit too large.

To my understanding, the lambda function S should actually be defined as:
>>> lambda S: (S[:, 0], S[:, 1:, target_col])
instead of
>>> lambda S: (S[:, 0], S[:, 1:, 1])
Otherwise, the `to_seq2seq_dataset` function parameter `target_col` is not even used.