Errata

While reading, I noticed a couple of minor inconsistencies that I thought might be helpful to point out for future editions or errata listings.

# Framework Reference Mismatch

In the section that reads:

"ML is a technique that enables us to do this, but to get started, we’ll need a framework—that’s where TensorFlow enters the picture. In the next section, we’ll take a look at what TensorFlow is and how to install it."

The following section actually discusses PyTorch, not TensorFlow, which may be confusing to readers expecting a TensorFlow introduction.

# Installation Instructions Reference

Similarly, in another section:

"With Python, there are many ways to install frameworks, but the default one supported by the TensorFlow team is pip. So, in your Python environment, installing PyTorch is as easy as using this:"

The transition from a TensorFlow-specific statement to PyTorch installation instructions may benefit from clarification or rephrasing to avoid ambiguity.

These are minor issues, but I thought they might be worth noting to maintain the excellent clarity and consistency throughout the book.

Thank you again for your work—it's been a valuable resource.

Note from the Author or Editor:
Please change the references to 'TensorFlow' to read 'PyTorch' instead

Text: Installing Porch in Python
Issue: Possible typo, probably intended as "Torch"
Suggested fix: Change to "Installing Torch in Python"

Note from the Author or Editor:
Doing 'PyTorch', where somehow the yT was missing :)

Text: wrapped in the torchvision.models library.
Issue: missing formatting for "torchvision.models"
Suggested fix: Format as fixed-width ("code" in Atlas) analoguous to "torch.distributed" 2 paragraphs above.

Note from the Author or Editor:
Fixed

Text: With Python, there are many ways to install frameworks, but the default one supported by the TensorFlow team is pip.
Issue: needs PT instead of TF
Suggested fix: replace "TensorFlow" with "PyTorch"

Note from the Author or Editor:
Fixed

In multiple instances the variable "y" is written in uppercase (as a capital "Y"), while in the remainder of the chapter it is consistently in lowercase.

I might be missing something here, but I suppose this has simply been overlooked in the copy editing process.

Note from the Author or Editor:
I think I got them all, changed to lowercase

Text: This will simply call the train function we specified five times […]
Issue: missing format for "train"
Suggested fix: format "train" as "literal" (ID and others) or "code" (Atlas)

Note from the Author or Editor:
I made it 'code'

Text: "dense layers"
Issue: very likely a tensor flow relict, as in PyTorch the term is "fully connected (fc)" as introduced in chapter 1
Suggested fix: Change wording "dense layer" → "fully connected layer". Maybe add a note explaining they are almost synonymous.

Note from the Author or Editor:
Avoided 'Dense' and used 'fully connected' and 'Linear' as appropriate

Text: Here, you’ll download some additional images for testing the model.
Issue: This contradicts the text on the previous page: ' In the preceding code snippet, you’ve already downloaded the training and validation datasets'.
Suggested fix: Not sure, what's best here. Maybe it's just a big misunderstanding on my side…… ¯\_(ツ)_/¯

Note from the Author or Editor:
You're right -- it's not very clearly worded. Better just to remove that sentence. Done :)

Text: missing?
Issue: Expected short explanation of transforms for validation data (2nd part of code)
Suggested fix: remove # Transforms for the validation data part or add short description of what it does.

Note from the Author or Editor:
Added blurb before code

Text: If you wanted to train a dataset to recognize […]
Issue: Possible confusion of terms
Suggested fix: replace "dataset" with, e.g., "model"

Note from the Author or Editor:
Fixed to 'model'

Text: nn.Linear(1024, 3) # Final layer for binary classification
Issue: Comment references binary classification
Suggested fix: # Final layer for ternary classification
or similar

Note from the Author or Editor:
Removed 'binary'

Text: If you explore this a little deeper, you can see that the file named scissors4.png had an output of –2.5582, –1.7362, 3.8465]
Issue: Missing opening square bracket
Suggested fix: –2.5582, –1.7362, 3.8465] → [–2.5582, –1.7362, 3.8465]

Note from the Author or Editor:
Fixed

Text: While FakeData only gives image types, you could relatively easily create your own CustomData (as we looked at earlier) […]
Issue: earlier class was called "CustomDataset"
Suggested fix: change CustomData → CustomDataset

Note from the Author or Editor:
Changed to 'your own custom data (as a CustomDataSet like we saw earlier) to provide...'

Text: One more thing to consider when using custom splits is that the name random doesn’t mean […]
Issue: name is 'random_split'
Suggested fix: replace 'random' → 'random_split'

Note from the Author or Editor:
Fixed

Text: For example, batching, image augmentation, mapping to feature columns, and other such logic […]
Issue: only mention of term 'feature columns' in the book
Suggested fix: rephrase for consistency with existing material

Note from the Author or Editor:
Changed to 'features' to be more generic

Text: Whenever you’re dealing with training or inference and you want the data or model to be on the accelerator, you’ll see something like .to(“cuda”) […]
Issue: use of typographical/smart quotes in code.
Suggested fix: replace .to(“cuda”) → .to("cuda")

Note from the Author or Editor:
Fixed

Text: This chapter covered the data ecosystem in PyTorch and introduced you to the dataset and DataLoader classes.
Issue: dataset class name should be uppercase
Suggested fix: dataset → Dataset

Note from the Author or Editor:
Fixed

Text: Then, you’ll be given the sequences representing the three sentences.
Issue: only two sentences
Suggested fix: replace three → two

Note from the Author or Editor:
Fixed

Text: […] so add 24 to get to 24, 24.
Issue: This looks like an array or a vector with missing brackets ([24, 24]).
Suggested fix: replace 24, 24 → 2,424 as shown in sample output above.

Note from the Author or Editor:
Fixed

Text: It was a very short sentence, so it’s padded up to 85 characters with a lot of zeros!
Issue: The array contains tokens.
Suggested fix: replace "characters" → "tokens"

Note from the Author or Editor:
Fixed

Text: You can then set the loss function and classifier to this. (Note that the LR is 0.001, or 1e–3.):
Issue: Text refers to optimizer, not classifier
Suggested fix: replace "classifier" → "optimizer"

Note from the Author or Editor:
Fixed

Text: […] while the loss for the test set diverged after 15
epochs, the validation loss turned to increase […]
Issue: Text mentions test set while chart is labelled as "Training and Validation Loss".
Suggested fix: replace "test set" → "training set"

Note from the Author or Editor:
Fixed

Text:
Figure 7-9. Accuracy for LSTM over 30 epochs
Figure 7-10. Loss with LSTM over 30 epochs
Issue: training ran für 300 epochs
Suggested fix: replace "30" → "300"

Note from the Author or Editor:
Fixed

Text: This model shows a total of 406.817 parameters of which only 6,817 are trainable, so training will be fast!
Issue: period instead of decimal comma
Suggested fix: replace '406.817' → '406,817'

Note from the Author or Editor:
Fixed

Text: You’ll want to create a single string with all the text and set that to be your data. Use \n for the line breaks. Then, this corpus can be easily loaded and tokenized. First, the tokenize function will split the text into individual words, and then the create_word_dictionary will create a dictionary with an index for each individual word in the text:
Issue: Missing formatting for "\n" and "tokenize".
Suggested fix: <code>\n</code>, <code>tokenize</code>.

Note from the Author or Editor:
Thanks, fixed!

Text: […] # features and targets […]
Issue: These were called "features and labels" before
Suggested fix: Unify to use "features and labels".

Note from the Author or Editor:
A little more nuanced than that -- when it's a sequence, I like to think of the label as the value that you are targeting, hence the interchangeable language. I edited the prose at the top to use 'target label' to clarify this, so when in code you only see 'target' it makes sense.

Text: As discussed in earlier chapters, with dropout, neighboring neurons are randomly dropped out (ignored) during training to avoid a familiarity bias.
Issue: Bias was introduced as "proximity bias" in ch. 3
Suggested fix: familiarity bias → proximity bias

Note from the Author or Editor:
Fixed

Text: python3 -m venv chapter12env
Issue: Example is from ch 13
Suggested fix: change chapter12env → chapter13env

Text: Before running it, make sure you have a model-store (or similar) directory that you will store the archived model in.
Issue: At the beginning of the chapter this directory was created under the name "model_store".
Suggested Change model-store → model_store

Text: On macOS, the shared RAM with the M-Series chips works well, while running on an M1 Mac with 16 Gb, the Gemma 2B is fast and smooth with Ollama.
Issue: "Gb" (Gigabit) vs "GB" (Gigabyte)
Suggested fix: replace Gb → GB

Text: Next, we can wrap all this code in a function called analyze, with this signature:
Issue: According to following code examples, function is called 'analyze_file'
Suggested fix: replace: function called analyze → function called analyze_file

Note from the Author or Editor:
Fixed

Text:If the process completes successfully, the JobId is updated […]
Issue: variable name in code is jobId (lowercase 'j')
Suggested fix: replace JobId → jobId in prose

Note from the Author or Editor:
Fixed