Python and TensorFlow 2 for Finland, the course that left only notebooks

deepkapha.ai had been to Finland three times for Ari Hovi Oy, the Finnish data training firm. Two public days in June 2018, two more that November, and a third pair in June 2019. The fourth was booked for 26 and 27 March 2020. The outline was drawn up on 18 January and titled Python Machine Learning with TensorFlow 2.0. Day one promised NumPy, Pandas, Matplotlib, Seaborn and Scikit-Learn, then hands-on work in TensorFlow 2.1. Day two promised neural networks, computer vision, natural language processing, best practice for putting models into production, and a live demo of a working cancer classifier.

No photographs were taken, and nothing in the course material names a meeting platform. The machines changed, though: in 2018 and 2019 every Finland working file was a local Jupyter notebook, and here fifteen of the twenty-one notebooks are Google Colab files, two of them asking for a high-memory runtime. The class ran on hosted runtimes rather than local installs.

It ran on the days it was booked for. NumPy was saved at 10:00 on 26 March and Pandas at 11:01. The first Series typed in front of the class held three city names: San Francisco, Helsinki, Amsterdam. By early afternoon the work was a neural network written in raw TensorFlow, saved at 14:15, then a multi-layer perceptron at 14:32 which finished 100 epochs on MNIST at 0.9319. The same problem went through Keras at 15:14 and came back at 1.0000 on the training set against 0.9836 on test, which settled the question of overfitting without anyone needing to argue it. A CIFAR-10 convolutional exercise followed at 15:30, and principal component analysis on the Wisconsin breast cancer data, thirty features pressed down to two, closed the day at 16:15.

Day two was built that night. Between 21:26 and 22:46 the exercises and solutions went in: a convolutional network, Fashion-MNIST, cats and dogs trained from scratch and then again on top of a pre-trained network, word embeddings, a recurrent network, text classification. The learner-facing Solutions folder was created at 21:23 and left empty; the three solution notebooks sat in the instructor folder instead. Only one notebook changed during the second day itself, the text classification one, at 14:16 on 27 March.

The reading had gone out first. A supplementary pack assembled from 22 March carried twenty-four cheatsheets, six books running from Bishop and Hastie to Goodfellow, and three papers on the ADAM optimiser, the Swish activation and one-pixel attacks. One more was added on 25 March, the day before the first session: Gehrmann and colleagues on patient phenotyping from clinical narratives, setting convolutional networks against concept extraction over 1,610 discharge summaries from MIMIC-III. The teaching deck behind all of it runs to 507 slides with 448 pages of notes, and its agenda slides still carry dates from earlier rooms, one reading 7 January 2019 and another 21 May 2018. Production and deployment were on the brochure for day two and are not in the notebooks. Finland came back for them on 28 April 2021, in a session on MLOps that built on this course.