notebooks

Course notebooks

Here you will find all the Jupyter notebooks used in the course. They will be added as the course progresses.

Notebook	Colab link
1: Digit classification Deep learning 101: Classifying handwritten numbers. This notebook serves as a quick intro to the Keras deep learning framework.
2: Fashion MNIST classification In this notebook we go a little further in computer vision.
3: Advanced image classification Time to do more advanced stuff, which requires a more powerful approach to constructing neural networks in Keras, namely the functional model API. Here we also start using TensorFlow datasets, image augmentation, and look at fine-tuning of existing models.
4: Visualising ConvNets Plot activations of the different filters in a convnet, to visualise how patterns are encoded from the first to the last layer.
5: Activations and initialisers Plot different activation functions and try them out along with different parameter initialisation schemes.
6: Callbacks and schedulers Dynamically modify the optimiser settings during training.
7: Optimisers Compare and evaluate different optimisation methods.
8: Data loading in TensorFlow Intro to efficient data loading with tf.data.Dataset.
9: Augmentation Test different augmentation methods for computer vision.
10: ResNets and modern convolutional networks Implement modern, non-sequential network architectures.
11: Image segmentation Train a model for semantic segmentation.
12: YOLO models Optional exercise: Try out pre-trained segmentation models from the YOLO family.
13: Image denoising with autoencoders Yet another computer vision task: Enhance images by removing noise.
14: Process sequences with RNNs Test different methods for forecasting passenger numbers on public transport, following the approach in Ch 14.
15: Weather forecasting with RNNs Another forecasting task, where we try even further deep learning approaches.
16: Translate languages with a sequence-to-sequence model Here we train a recurrent network to predict entire sequences. For a given input sentence in English, we train a model to predict the Frensh translation.
17: Anomaly detection Train an autoencoder to detect anomalous data in time series.
18: Hacking CNNs with adversarial examples Try to fool an advanced pre-trained convolutional network into making bad predictions, by sprinkling some magic at the input images.
19: Compare neural networks to tree-based models on tabular data Compare the predictive performance of neural networks and decision trees, when applied to tabular datasets.
20: Preprocessing tabular data with Keras Try out the different preprocessing layers in Keras.
21: Embeddings and modern networks for tabular data Skip the one-hot encodings and replace them with embeddings. Then train a transformer model on it 👩‍💻
22: Text classificatiom Build an NLP model for sentiment analysis, in this case, classifying film reviews.
23: Word embeddings Train word embeddings and investigate the high-dimensional embedding space.
24: Fine-tune a model on pretrained embeddings Benefit from the work others have done, by downloading pretrained word embeddings.
25: Code classification Can you tell the difference between Java, JavaScript, C# and Python?
26: Tokenisers Modern language models rely on different tokenisation algorithms to split up the text into useful tokens. Here we try out a few of them.
27: Transformer encoder for classification We continue doing text classification, but with a more powerful tool: The transformer.
28: Attention visualisation Visualise the attention mechanism in pretrained large language models.
29: Machine translation with transformers Translate to the language of your choice using an encoder-decoder transformer model.
30: Text generation Try out different sampling strategies to generate text just like ChatGPT (or at least almost).
31: Run LLMs from Hugging Face Hub Run state-of-the-art open-source models with almost zero effort.