Errata

"Siddha Gangju" should be "Siddha Ganju"

Note from the Author or Editor:
Oh, really sorry about that! This is now fixed.

The text says that thethas value is 1.5, but if we check the plots in the previous page, the max is around 1.65

Note from the Author or Editor:
Great catch, thanks! Indeed, the max is around 1.65 (about 1.66, in fact). I fixed the text.

Actually, make_column_selector is a class, not a function.
So I suggest that

"make_column_selector() function" --> "make_column_selector class"
Two "selector function" --> "selector"

Thank you!

Note from the Author or Editor:
Good point, thanks. Here's the updated paragraph:

Since listing all the column names is not very convenient, Scikit-Learn provides a `make_column_selector` class that you can use to automatically select all the features of a given type, such as numerical or categorical. You can pass a selector to the `ColumnTransformer` instead of column names or indices. Moreover, if you don't care about naming the transformers, you can use `make_column_transformer()`, which chooses the names for you, just like `make_pipeline()` does. For example, the following code creates the same `ColumnTransformer` as earlier, except the transformers are automatically named `"pipeline-1"` and `"pipeline-2"` instead of `"num"` and `"cat"`:

Hi Aurelien,

(p91, p92) In the middle of page,

I suggest to change `RandomForest` --> `RandomForestRegressor`

Thank you!

Note from the Author or Editor:
Good catch, thanks. This is now fixed.

Hi Aurelien,

I think `memory` of `Pipeline` class is just parameter, not hyperparameter.

Thanks!

Note from the Author or Editor:
Good point, thanks. I replaced "hyperparameter" with "parameter".

While the encoder uses LSTM with 512 unit, in the following section it mentions GRU with 10 units.

This is confusing, I think.

Note from the Author or Editor:
Thanks for your feedback.
Indeed, this was a really confusing mistake, my apologies. Here's the corrected note:

The `Bidirectional` layer will create a clone of the `LSTM` layer (but in the reverse direction), and it will run both and concatenate their outputs. So although the `LSTM` layer has 256 units, the `Bidirectional` layer will output 512 values per time step.

The "Punctuation" section mentions "Punctutation" (note the extra t).

I checked Amazon "look inside", which indicates the same error in the print version.

Note from the Author or Editor:
Good catch, thank you! That's a typo, it should indeed be "punctuation".

Dear Mr. Geron,

I hope this message finds you well. I would like to report a few issues I encountered in Figure B2 and the explanation of autodifferentiation in the accompanying text. Please see the details below:

1. Missing Function Definition in Figure B2: The function f is not explicitly mentioned in Figure B2. The correct function should be f(x, y) = y \cdot x^2 + y + 2 .

2. Missing Multiplication in Figure B2: In Figure B2, one of the blue blocks lacks the indication of the mathematical operation of multiplication, which seems to be missing from the diagram.

3. Unclear Explanation of Autodiff: Just below Equation B3, the explanation of autodiff with a general function h is unclear. While the implementation of autodiff with basic operations (addition, subtraction, multiplication) is understandable, it is not clear how autodiff would work (operationally on a computer) when the function h cannot be written in closed form. Could you clarify how autodiff is implemented in such cases when h does not have a closed-form expression?

In short, Appendix B is badly written.

Thank you for considering these points, and I look forward to your response.

Best regards,
Sashwat Tanay

Note from the Author or Editor:
Thanks for your feedback!

1. The function f is defined at the start of the appendix, but you're right, it should also be visible on the figure. I just added it.

2. Thanks, indeed the multiplication sign was missing. Fixed! :)

3. What kind of non-closed-form example do you have in mind? Loops and conditionals? If so, there are two options: eager mode or graph mode. In eager mode, the computation graph is built on the fly during the forward pass, so you can use loops and conditionals without any problem (e.g., if you run a long loop, the graph will just become large, and the backward pass will work normally on that graph). In graph mode, TensorFlow will build a new computation graph that will contain nodes for the loops and conditionals, and it will run the appropriate operations during forward and backward pass.
I could add this to appendix B, but I think it might confuse the reader, and I see it as a technical detail. The main idea of reverse-mode autodiff is that you go forward through the computation graph, you store all the intermediate results, and you flow the gradients backward by applying the chain rule at each step.

"Luckily, it [RandomForestClassifier] has a predict_proba() method that returns class probabilities for each instance, and we can just use the probability of the positive class as a score, so it will work fine. [Footnote 4] ...

Footnote 4: Scikit-Learn classifiers always have either a decision_function() method or a predict_proba() method, or sometimes both.

WARNING
These are estimated probabilities, not actual probabilities. ...
The sklearn.calibration package contains tools to calibrate the estimated probabilities and make them much closer to actual probabilities.
See the extra material section in this chapter’s notebook for more details."

Corrections:
a) Given this WARNING, "so it will work fine" seems misleading and incorrect.

b) The chapter 3 notebook at
github dot com/ageron/handson-ml3/blob/main/03_classification.ipynb
is missing code that uses the sklearn.calibration package or that addresses the topic of "calibration".

c) I think this WARNING should be integrated into the core text of the textbook, and taught prior to showing how to obtain so-called "probabilities" using the predict_proba() method, because, some people might skip the warnings.

I think understanding the difference between non-calibrated and calibrated probabilities is an important point too many textbooks omit.

Note from the Author or Editor:
Thanks for your message. I agree that it's important to make it clear that classifiers output *estimated* probabilities, not actual probabilities.

a) You're right, the phrase "so it will work fine" was unfortunate, I was referring to the fact that you wouldn't get an error, not that there's nothing to worry about. I rephrased this for the next reprints and the next edition.

b) I initially intended to add a code example showing how to use sklearn.calibration, but decided that the documentation was clear enough, but I forgot to remove the reference in the text. I just removed it.

c) The scorpion notes are intended to stand out and represent important warnings. I asked O'Reilly and several friends, and they all agree that important messages should be put in such boxes. However, I replaced "class probabilities" with "estimated class probabilities" the first times these terms are used, which is just before the warning message.

Thanks again for your feedback!

"can be simply be copied" -> "can simply be copied"

Note from the Author or Editor:
Good catch, thanks!

Because combination attributes is not added yet, there are 9 numerical attributes.
So, "there are now 11 numerical attributes, you would get 11^2 = 121 plots" should be "there are now 9 numerical attributes, you would get 9^2 = 81 plots".
Thanks

Note from the Author or Editor:
Good catch, thanks, this is now fixed.

In the second paragraph on page 93, right below the code it states that the best model is obtained by setting max_features to 8, but the output of the code states it should be set to 6.

Note from the Author or Editor:
Good catch, thanks! This is fixed now.

Hi,

After `>>> y_some_digits_pred = (y_scores > threshold)`, `>>> y_some_digits_pred` should be added.

Thank you!

Note from the Author or Editor:
Great catch, thanks! This is now fixed.

Hi,

"the function adds a last precision of 0 and a last recall of 1" should be "the function adds a last precision of 1 and a last recall of 0".

Thank you!

Note from the Author or Editor:
Great catch, thanks! This is now fixed.

Hi,

In the last paragraph on p126 and the first paragraph on p127, `ChainClassifier` should be `ClassifierChain`.

Thank you!

Note from the Author or Editor:
Good catch, thanks. This is now fixed.

I believe that Figure 6-4 is from the 2nd edition, and it is different from the one in the 3rd edition notebook.
Additionally, the explanation below Figure 6-2 needs to be updated to reflect this change.
Thanks!

Note from the Author or Editor:
Good catch, thanks! This is now fixed.

The page break between pages 232 and 233 garbles up the text.

It currently reads:
", and use these (page break) can be used as the input features..."
instead of:
", and use these (page break) as the input features ..."

Note from the Author or Editor:
Good catch, thanks. Fixed!

The homl.info link to extra-anns is broken.

Note from the Author or Editor:
Thanks for your feedback. This is now fixed.

The footnote states: "Due to TensorFlow issue #46858, the call to super().build() may fail in this case, unless the issue was fixed by the time you read this. If not, you need to replace this line with self.built = True."

However, on the code snippet, neither "super().build()" nor "self.built = True" appear. They do appear in the Google colab and and GitHub.

The footnote (as well as the code in Google colab and GitHub) should be removed, as "super().build()" is deprecated in the latest TensorFlow versions (> v2.5.0), Keras will call it for us.

Note from the Author or Editor:
Good catch, thanks. Indeed, calling super().build() or setting self.built = True is no longer needed, and I removed it from the code in the book, but I forgot to remove the footnote. I'll also update the Colab notebooks now.

In Fig 14-3, the bottom blue 3x3 square should be sifted by one square, since the stride in that figure is 1.

Note from the Author or Editor:
Great catch, thanks a lot! This figure was correct in the previous editions, but it was not very pretty, so we tried to beautify it, but in the process we made a copy/paste error with Figure 14-4. My apologies, I should have caught that. This is now fixed.

The last 2 paragraphs of page 489 in the printed edition state (sorry for the subscripts not showing up correctly):

"With padding="valid", if the width of the input is ih, then the output width is equal to (ih – fh + sh) / sh, rounded down. Recall that fh is the kernel width, and sh is the horizontal stride. Any remainder in the division corresponds to ignored columns on the right side of the input image. The same logic can be used to compute the output height, and any ignored rows at the bottom of the image.

With padding="same", the output width is equal to ih / sh, rounded up. To make this possible, the appropriate number of zero columns are padded to the left and right of the input image (an equal number if possible, or just one more on the right side). Assuming the output width is ow, then the number of padded zero columns is (ow – 1) × sh + fh – ih. Again, the same logic can be used to compute the output height and the number of padded rows."

In all other locations, for example, fh refers to the height of the kernel/receptive field, not the width. It seems like these last 2 paragraphs are mixing up height and width in a couple of locations (i.e. the subscripts should be all w's or all h's).

Note from the Author or Editor:
Good catch, thanks. You're right, all the h subscripts in these two paragraphs should we w subscripts, my apologies for the confusion.

Incomplete reference in index: sklearn.model_selection.RandomizedSearchCV is already introduced on pages 93/94 (in addition to p. 248)