Machine Learning Course Notebooks: teaching notebooks from linear regression to CNNs

Overview

Thirteen Jupyter notebooks that walk through classic machine-learning topics: random forests and decision trees, regression on the California housing data, neural networks in TensorFlow and PyTorch, a CNN on MNIST, k-means with silhouette analysis, text classification and pandas basics. Most notebooks keep the outputs they produced, so students can read them without running anything. Six were re-run on 2026-10-13 and 2026-10-14 (four of them fixed to download their own data), including the CNN on MNIST. The other seven were re-run on 2026-10-15: two use data they already downloaded, and five whose original data files are not available now download a public substitute dataset instead, so their results differ from the stored outputs they had before (each says so in a note at the top).

These notebooks are used in the machine learning lab sessions I teach.

Demo

No hosted demo. Open any notebook in Google Colab with the badges in the Notebooks table.

Problem statement

Students learning ML need small, runnable examples on public data, each showing one idea end to end. These notebooks are that material. Before these updates most of them could not run from a clean copy: they read data from Colab or Kaggle paths that were not in the repository, and several broke on current library versions.

Approach

Each notebook is self-contained and follows the usual flow: load data, explore it, fit a model, evaluate it. All thirteen notebooks have been re-run and fetch their data in their first cells (a download from the dataset's public source, or a Keras loader), so they work in a clean checkout and in Colab. Where the original data file was lost, the notebook downloads a public substitute and says so in its first cell.

flowchart LR
  A[Download or load public dataset] --> B[Explore and prepare]
  B --> C[Fit model: forest, regression, CNN, k-means]
  C --> D[Evaluate and plot]

Loading

Data

No data is stored in this repository. Public datasets are downloaded by the notebooks into a git-ignored data/ folder; the sources and licenses are listed in THIRD_PARTY_NOTICES.md.

Dataset	Rows	Source	License / terms
Banknote Authentication (Lohweg, 2012)	1,372	UCI Machine Learning Repository	CC BY 4.0
California housing (training set)	17,000	Google Machine Learning Crash Course	Per the course page
MNIST	60,000 train / 10,000 test	Keras loader	Per the MNIST authors
Pima Indians Diabetes Database (substitute for the lost diabetes.csv)	768	OpenML data set 37 (copy of the UCI data); falls back to the identical table in plotly/datasets	OpenML licence field "Public"; MIT for the plotly/datasets repository
Titanic passenger list (substitute for the Kaggle files)	1,309 (split 891 / 418)	OpenML data set 40945 (Harrell and Cason, Vanderbilt Biostatistics)	OpenML licence field "Public"
SMS Spam Collection (substitute for the Kaggle disaster tweets)	5,574 (split 4,459 / 1,115)	UCI Machine Learning Repository	CC BY 4.0
Restaurant tips (substitute for the lost tip-amount.xlsx)	244	plotly/datasets GitHub repository	MIT for the repository; no separate licence stated for the table
WHO alcohol consumption by beverage type (substitute for the lost world_alcohol.csv)	41,521	WHO Global Health Observatory OData API, indicator SA_0000001400	CC BY 4.0 (WHO data terms)

The original files behind the last five notebooks (diabetes.csv, the Kaggle Titanic and Disaster Tweets files, tip-amount.xlsx, world_alcohol.csv) are not in this repository; the notebooks no longer use them. The source of tip-amount.xlsx and world_alcohol.csv was not recorded. Each of those five notebooks has a note at the top that names the substitute, so results differ from the earlier stored outputs.

Notebooks

Notebook	Topic	Dataset	Outputs (cells with output / code cells)	Open in Colab	Status
ANN_using_pytorch.ipynb	Feed-forward neural network in PyTorch (banknote forgery detection)	Banknote Authentication (UCI), downloaded by the notebook	yes (14/23), no errors		Verified (re-run 2026-10-15, UCI banknote data): 17 s on CPU with PyTorch 2.5.1, which is now pinned in requirements.txt
ANN_using_tensorflow.ipynb	Neural network in TensorFlow/Keras, with a network diagram (drawn with matplotlib)	Banknote Authentication (UCI), downloaded by the notebook	yes (61/78), no errors		Verified (re-run 2026-10-15, UCI banknote data): 26 s. Fixed an empty Dense(), a missing Input import and undefined y_pred2 / y_class2; ann_visualizer (needs Graphviz) replaced by a matplotlib function
CNN_Identify_the_Digits_(MNIST).ipynb	CNN digit classifier	MNIST (Keras loader); the optional personal-image cells (2.jpeg) are skipped when the file is missing	yes (35/41), no errors		Verified (re-run 2026-10-14, full settings: batch size 10, 5 epochs, CPU, 12 min): test accuracy 0.9905. Three cells adjusted so it runs outside Colab
CNN_Identify_the_Digits_(MNIST)_validation_split.ipynb	Companion of the CNN above with a held-out validation split and fixed seeds	MNIST (Keras loader)	yes (10/10), no errors		Verified (re-run 2026-10-14, seed 0; seeds 0, 1, 2 in the log): see the section below
Natural_Language_Processing_with_Disaster_Tweets.ipynb	Uses the SMS Spam Collection; file name kept for link stability. Text preprocessing, word2vec and Naive Bayes on short texts (now SMS spam, formerly tweets)	SMS Spam Collection (UCI, CC BY 4.0), downloaded by the notebook; substitute for the Kaggle "Disaster Tweets" data, which is not available here	yes (39/52), no errors		Verified (re-run 2026-10-15, SMS Spam Collection substitute data): 11 s on gensim 4. Results differ from the earlier tweet outputs
SigmoidNeuron.ipynb	Sigmoid neuron from scratch	Synthetic, generated in the notebook	yes (13/26)		Verified (re-run 2026-10-14, 56 s): runs clean after replacing tqdm_notebook with tqdm
banknote_randomforest.ipynb	Random forest vs decision tree on banknote data	Banknote Authentication (UCI), downloaded by the notebook	yes (3/11)		Verified (re-run 2026-10-13, 35 s): runs clean. Data is downloaded in code
california_regression.ipynb	Linear regression in TensorFlow, then a random forest	California housing (Google MLCC), downloaded by the notebook	yes (17/27)		Verified (re-run 2026-10-13, 5 min): runs clean after fixing a plot error on current NumPy
diabetes.ipynb	Decision tree on diabetes data, tree plot	Pima Indians Diabetes (OpenML data set 37), downloaded by the notebook; substitute for the lost diabetes.csv	yes (10/19), no errors		Verified (re-run 2026-10-15, Pima Indians Diabetes substitute data): 10 s. sklearn.externals.six removed and the pydotplus/graphviz plot replaced by plot_tree. Results differ from the earlier outputs
mnist_opencv.ipynb	Small CNN on MNIST	MNIST (Keras loader)	yes (14/20)		Verified (re-run 2026-10-13, 3 min): runs clean after fixing a removed Keras method
plot_kmeans_silhouette_analysis.ipynb	K-means and silhouette analysis	Synthetic, make_blobs	yes (1/2)		Executed 2026-10-13 (11 s). It had no outputs before
tip_amount_tensorflow.ipynb	Linear regression in TensorFlow	Restaurant tips (plotly/datasets), downloaded by the notebook; substitute for the lost tip-amount.xlsx	yes (9/13), no errors		Verified (re-run 2026-10-15, restaurant tips substitute data): 15 s. The stored KeyError is fixed. Results differ from the earlier outputs
titanic-tutorial.ipynb	Titanic survival: feature engineering and models	Titanic passenger list (OpenML data set 40945), downloaded by the notebook; substitute for the Kaggle files, which need a login	yes (33/45), no errors		Verified (re-run 2026-10-15, OpenML Titanic substitute data): 36 s. One cell added that scores the models on the 418 held-out rows. Results differ from the earlier outputs
world_alcohol_pandas_revision.ipynb	pandas selection and filtering	WHO Global Health Observatory alcohol consumption by beverage type (CC BY 4.0), downloaded by the notebook; substitute for the lost world_alcohol.csv	yes (31/32), no errors		Verified (re-run 2026-10-15, WHO GHO substitute data): 23 s. Has about 41,500 rows instead of 100, so printed rows and counts differ

Status labels: Verified (re-run date) means the notebook was executed top to bottom and its outputs were kept. Verified (re-run date, substitute data) means the same, but the notebook's original data file is not available, so it ran on the named public dataset instead and its results are not the original results. Not re-run (reason) means it was only edited as described, or not touched, and its stored outputs come from earlier sessions. The Colab links use the master branch, which is this repository's default branch.

Results

Numbers come from the re-runs on 2026-10-13 (log: docs/runs/2026-10-13-notebook-reruns.md), 2026-10-14 (log: docs/runs/2026-10-14-cnn-mnist.md) and 2026-10-15 (log: docs/runs/2026-10-15-selfcontained-notebooks.md). Most of the notebooks do not set a random seed, so repeated runs differ slightly. The rows dated 2026-10-15 for the five substituted datasets are results on the substitute data; they are not the results of the original data files and must not be compared with the earlier stored outputs.

Metric	Value	Status	How measured	Date
Banknote, random forest accuracy	0.99125 (0.9912536443148688)	Verified (re-run 2026-10-13)	banknote_randomforest.ipynb, 343 held-out rows, confusion matrix [[184, 3], [0, 156]]	2026-10-13
California housing, random forest R²	0.8094 (0.8094088629585355), MSE 2616.38	Verified (re-run 2026-10-13)	california_regression.ipynb, as printed by the notebook	2026-10-13
MNIST, small CNN, 1 epoch at batch size 1	training accuracy 0.8985	Verified (re-run 2026-10-13)	mnist_opencv.ipynb. Training accuracy, not a test score	2026-10-13
MNIST, CNN (2 conv + dense), 5 epochs at batch size 10, full 60,000 training images	test accuracy 0.9905 (0.9904999732971191), test loss 0.0297	Verified (re-run 2026-10-14, full settings, CPU)	CNN_Identify_the_Digits_(MNIST).ipynb, 10,000 MNIST test images (also used as the validation set; no tuning was done on it). Log: docs/runs/2026-10-14-cnn-mnist.md	2026-10-14
Banknote, PyTorch network (4-8-4-1, 3,500 full-batch epochs), test accuracy	1.000 (343 of 343), final training loss 0.0106	Verified (re-run 2026-10-15, UCI banknote data)	ANN_using_pytorch.ipynb, 25% held-out rows (343), confusion matrix [[191, 0], [0, 152]]	2026-10-15
Banknote, Keras networks, test accuracy	0.9621 (6 hidden units, SGD, 10 epochs) to 1.000 (four hidden layers with dropout, 20 epochs of Adam)	Verified (re-run 2026-10-15, UCI banknote data)	ANN_using_tensorflow.ipynb, 343 held-out rows. Unseeded	2026-10-15
SMS spam, Multinomial Naive Bayes on TF-IDF (1-2 grams), validation accuracy	0.9426 (spam recall 0.48, precision 1.00)	Verified (re-run 2026-10-15, SMS Spam Collection substitute data)	Natural_Language_Processing_with_Disaster_Tweets.ipynb, 836 validation messages	2026-10-15
Pima diabetes, random forest (20 trees), test accuracy	0.779 (180 of 231; confusion matrix [[129, 17], [34, 51]])	Verified (re-run 2026-10-15, Pima substitute data)	diabetes.ipynb, 231 held-out rows. Unseeded trees: the single decision trees vary by a few points between runs	2026-10-15
Titanic, logistic regression, held-out accuracy	0.780 (best of nine models; random forest 0.773)	Verified (re-run 2026-10-15, OpenML Titanic substitute data)	titanic-tutorial.ipynb, 418 held-out rows. The notebook's own scores are training accuracy	2026-10-15
Restaurant tips, one-feature linear regression, training RMSE after 10 epochs	1.20 US dollars	Verified (re-run 2026-10-15, restaurant tips substitute data)	tip_amount_tensorflow.ipynb, 244 rows, as printed by the notebook. Unseeded; training error, not a test score	2026-10-15
WHO alcohol table	41,521 rows, 5 columns	Verified (re-run 2026-10-15, WHO GHO substitute data)	world_alcohol_pandas_revision.ipynb, shape printed by the notebook. A pandas exercise, so no model metric	2026-10-15

CNN on MNIST with a separate validation split

CNN_Identify_the_Digits_(MNIST) above passes the test images to model.fit as validation data. The companion notebook CNN_Identify_the_Digits_(MNIST)_validation_split.ipynb trains the same network (batch size 10, 5 epochs) on the first 50,000 training images, uses the last 10,000 training images as validation data, fixes the random seeds, and evaluates the test set once at the end. Over seeds 0, 1 and 2 the test accuracy was 0.9907, 0.9890 and 0.9887 (mean 0.9895): Verified (re-run 2026-10-14, seeds 0, 1, 2, validation split = last 10,000 of the 60,000 training images). The difference from 0.9905 is about 0.1 percentage points, the same size as the spread between seeds, so these runs show no large inflation but cannot settle a small one; the new runs also use 50,000 rather than 60,000 training images. Details, per-epoch numbers and what the split changes: docs/runs/2026-10-14-mnist-validation-split.md. The three seeds were run with scripts/mnist_validation_split.py (python scripts/mnist_validation_split.py --seed 0 --out seed0.json, about 6 minutes each on an 8-core CPU).

Tech stack

Python 3.11, Jupyter, TensorFlow 2.15.1 with Keras 2.15.0, scikit-learn 1.5.2, NumPy 1.26.4, pandas 2.2.3, matplotlib 3.9.2, seaborn 0.13.2. ANN_using_pytorch.ipynb uses PyTorch 2.5.1 (CPU build). The NLP notebook uses gensim 4.3.3, NLTK 3.9.1 and SciPy 1.13.1. All pins are in requirements.txt.

Getting started

Prerequisites: Python 3.11 (TensorFlow 2.15 supports 3.9 to 3.11), about 3 GB of disk space for the packages (TensorFlow and the CPU build of PyTorch), and an internet connection for the dataset downloads. The commands below were tested on Windows; on Linux or macOS use python3.11 -m venv .venv and source .venv/bin/activate instead of the two Windows lines.

git clone https://github.com/Janeferdinant/Machine-learning-Course.git
cd Machine-learning-Course
py -3.11 -m venv .venv
.venv\Scripts\activate
pip install -r requirements.txt
python -m jupyter nbconvert --to notebook --execute --output-dir run-output banknote_randomforest.ipynb

Expected: the last command writes run-output/banknote_randomforest.ipynb in well under a minute, and the cell that prints the random forest accuracy shows 0.9912536443148688. To work interactively, install JupyterLab (pip install jupyterlab), run python -m jupyter lab and open any notebook. TensorFlow is a large install and can take a long time on a slow connection.

Project structure

*.ipynb                    The 13 course notebooks plus the validation-split companion of the MNIST CNN
requirements.txt           Package versions used for the re-runs (2026-10-13 to 2026-10-15)
docs/runs/                 Logs of the re-runs
scripts/                   mnist_validation_split.py: the seeded validation-split run (three seeds)
finance.py                 A Colab export of a Streamlit stock demo; not valid Python (see Limitations)
THIRD_PARTY_NOTICES.md     Dataset and code credits and licenses
LICENSE                    CC BY-NC-ND 4.0
data/                      Created by the notebooks; git-ignored

Real-world applications

Fraud and forgery detection (banknotes), property price estimation (California housing), digit and image recognition (MNIST), segmentation (k-means), and text classification (short messages).

Limitations

• All thirteen notebooks have been re-run: four on 2026-10-13, CNN_Identify_the_Digits_(MNIST) and SigmoidNeuron on 2026-10-14, and the other seven on 2026-10-15. Five of those seven (diabetes, titanic-tutorial, Natural_Language_Processing_with_Disaster_Tweets, tip_amount_tensorflow, world_alcohol_pandas_revision) now run on public substitute data because their original files are not available. Their results are not the results of the original data, and the NLP notebook now classifies SMS messages although its file name still says "Disaster Tweets" (the name is kept so the Colab links keep working).
• The substitute datasets are downloaded from third-party servers. OpenML in particular sometimes answers with an error under load (seen during the 2026-10-15 runs); the Pima notebook then falls back to a GitHub copy of the same table, and the Titanic notebook retries a few times. If a download fails, run the cell again later.
• CNN_Identify_the_Digits_(MNIST) was re-run on 2026-10-14 with its original training settings (30,000 training steps, about 10 minutes of training on an 8-core CPU). The network diagram cell needs Graphviz and prints a note without it, and the personal-image cells (2.jpeg) are skipped because that file is not in the repository.
• The fixes to tip_amount_tensorflow (undefined variables, feature[-1]) and ANN_using_tensorflow (undefined predictions, empty Dense()) are listed in the note at the top of each notebook. The network diagrams in ANN_using_tensorflow are now drawn with matplotlib and show at most 10 nodes per layer.
• diabetes draws its trees with scikit-learn's plot_tree instead of pydotplus and Graphviz, and no longer writes diabetes.png.
• Results vary between runs because most notebooks set no seed. The mnist_opencv accuracy of 0.8985 is a training accuracy, not a test score; the CNN notebook's 0.9905 is measured on the 10,000 test images.
• mnist_opencv is named after OpenCV but does not use it.
• The versions in requirements.txt were tested locally, not in Colab. Colab ships newer TensorFlow, Keras and NumPy, so a notebook may behave differently there.
• finance.py contains !pip install lines, so it cannot run as a Python file; it is left as uploaded.
• The versions in requirements.txt were tested on Windows with Python 3.11 only. On Linux the extra index line in requirements.txt selects the CPU build of PyTorch; that was not tested here.

Future scope

Set random seeds, add a Colab smoke test for the self-contained notebooks, rename the NLP notebook now that it no longer uses tweets, and map notebooks to lecture topics.

Credits & citation

• Origin. This repository was created from the GitHub Learning Lab course "Get Started with TensorFlow" (everydeveloper) under the name TensorFlow-Pokemon-Course, and later renamed. No course content remains. See THIRD_PARTY_NOTICES.md.
• Code. plot_kmeans_silhouette_analysis follows the scikit-learn example gallery (BSD-3-Clause). california_regression and tip_amount_tensorflow build on Google's Machine Learning Crash Course exercise "Linear regression with a real dataset" (Apache-2.0). Others follow public tutorials, listed with their confidence level in THIRD_PARTY_NOTICES.md.
• Data. Lohweg, V. (2012). Banknote Authentication [Dataset]. UCI Machine Learning Repository. https://doi.org/10.24432/C55P57. California housing data via Google's Machine Learning Crash Course. MNIST by LeCun, Cortes and Burges. Substitute data used from 2026-10-15: Pima Indians Diabetes Database (Sigillito; original owner NIDDK) via OpenML and UCI; Titanic passenger list (Harrell and Cason, Vanderbilt Biostatistics) via OpenML; Almeida, T. and Hidalgo, J. (2011). SMS Spam Collection [Dataset]. UCI Machine Learning Repository. https://doi.org/10.24432/C5CC84 (CC BY 4.0); restaurant tips table (Bryant and Smith, 1995) via the plotly/datasets repository; World Health Organization, Global Health Observatory, recorded alcohol per capita consumption by beverage type (CC BY 4.0). Licenses and URLs are in THIRD_PARTY_NOTICES.md.

License

Original work © Jane Rose Ferdinant, licensed CC BY-NC-ND 4.0. Third-party components retain their own licenses; see THIRD_PARTY_NOTICES.md.

Author

Jane Rose Ferdinant: Portfolio · LinkedIn