The lecture content, including references to study materials. The main study material is the Deep Learning Book by Ian Goodfellow, Yoshua Bengio and Aaron Courville, (referred to as DLB).

References to study materials cover all theory required at the exam, and sometimes even more – the references in italics cover topics not required for the exam.

1. Introduction to Deep Learning

 Feb 19 Slides PDF Slides CZ Lecture CZ UniApprox CZ Practicals EN Lecture EN UniApprox EN Practicals Questions numpy_entropy pca_first mnist_layers_activations

• Random variables, probability distributions, expectation, variance, Bernoulli distribution, Categorical distribution [Sections 3.2, 3.3, 3.8, 3.9.1 and 3.9.2 of DLB]
• Self-information, entropy, cross-entropy, KL-divergence [Section 3.13 of DBL]
• Gaussian distribution [Section 3.9.3 of DLB]
• Machine Learning Basics [Section 5.1-5.1.3 of DLB]
• History of Deep Learning [Section 1.2 of DLB]
• Linear regression [Section 5.1.4 of DLB]
• Challenges Motivating Deep Learning [Section 5.11 of DLB]
• Neural network basics
  • Neural networks as graphs [Chapter 6 before Section 6.1 of DLB]
  • Output activation functions [Section 6.2.2 of DLB, excluding Section 6.2.2.4]
  • Hidden activation functions [Section 6.3 of DLB, excluding Section 6.3.3]
  • Basic network architectures [Section 6.4 of DLB, excluding Section 6.4.2]
• Universal approximation theorem

2. Training Neural Networks

 Feb 26 Slides PDF Slides CZ Lecture CZ Practicals EN Lecture EN Practicals Questions sgd_backpropagation sgd_manual mnist_training gym_cartpole

• Capacity, overfitting, underfitting, regularization [Section 5.2 of DLB]
• Hyperparameters and validation sets [Section 5.3 of DLB]
• Maximum Likelihood Estimation [Section 5.5 of DLB]
• Neural network training
  • Gradient Descent and Stochastic Gradient Descent [Sections 4.3 and 5.9 of DLB]
  • Backpropagation algorithm [Section 6.5 to 6.5.3 of DLB, especially Algorithms 6.1 and 6.2; note that Algorithms 6.5 and 6.6 are used in practice]
  • SGD algorithm [Section 8.3.1 and Algorithm 8.1 of DLB]
  • SGD with Momentum algorithm [Section 8.3.2 and Algorithm 8.2 of DLB]
  • SGD with Nestorov Momentum algorithm [Section 8.3.3 and Algorithm 8.3 of DLB]
  • Optimization algorithms with adaptive gradients
    • AdaGrad algorithm [Section 8.5.1 and Algorithm 8.4 of DLB]
    • RMSProp algorithm [Section 8.5.2 and Algorithm 8.5 of DLB]
    • Adam algorithm [Section 8.5.3 and Algorithm 8.7 of DLB]

3. Training Neural Networks II

 Mar 4 Slides PDF Slides CZ Lecture CZ Convergence CZ Practicals EN Lecture EN Convergence EN Practicals EN sgd_manual Gradients Questions mnist_regularization mnist_ensemble uppercase

• Softmax with NLL (negative log likelihood) as a loss function [Section 6.2.2.3 of DLB, notably equation (6.30); plus slides 10-12]
• Regularization [Chapter 7 until Section 7.1 of DLB]
  • Early stopping [Section 7.8 of DLB, without the How early stopping acts as a regularizer part]
  • L2 and L1 regularization [Sections 7.1 and 5.6.1 of DLB; plus slides 17-18]
  • Dataset augmentation [Section 7.4 of DLB]
  • Ensembling [Section 7.11 of DLB]
  • Dropout [Section 7.12 of DLB]
  • Label smoothing [Section 7.5.1 of DLB]
• Saturating non-linearities [Section 6.3.2 and second half of Section 6.2.2.2 of DLB]
• Parameter initialization strategies [Section 8.4 of DLB]
• Gradient clipping [Section 10.11.1 of DLB]

4. Convolutional Neural Networks

 Mar 11 Slides PDF Slides CZ Lecture CZ Practicals EN Lecture EN Practicals Questions mnist_cnn torch_dataset mnist_multiple cifar_competition

• Introduction to convolutional networks [Chapter 9 and Sections 9.1-9.3 of DLB]
• Convolution as operation on 4D tensors [Section 9.5 of DLB, notably Equations (9.7) and (9.8)]
• Max pooling and average pooling [Section 9.3 of DLB]
• Stride and Padding schemes [Section 9.5 of DLB]
• AlexNet [ImageNet Classification with Deep Convolutional Neural Networks]
• VGG [Very Deep Convolutional Networks for Large-Scale Image Recognition]
• GoogLeNet (aka Inception) [Going Deeper with Convolutions]
• Batch normalization [Section 8.7.1 of DLB, optionally the paper Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift]
• Inception v2 and v3 [Rethinking the Inception Architecture for Computer Vision]
• ResNet [Deep Residual Learning for Image Recognition]

5. Convolutional Neural Networks II

 Mar 18 Slides PDF Slides CZ Lecture CZ Practicals EN Lecture EN Transposed Convolution EN Practicals Questions cnn_manual cags_classification cags_segmentation

• Residual CNN Networks
  • ResNet [Deep Residual Learning for Image Recognition]
  • WideNet [Wide Residual Network]
  • DenseNet [Densely Connected Convolutional Networks]
  • PyramidNet [Deep Pyramidal Residual Networks]
  • ResNeXt [Aggregated Residual Transformations for Deep Neural Networks]
• Regularizing CNN Networks
  • DropLayer [Deep Networks with Stochastic Depth]
  • CutOut [Improved Regularization of Convolutional Neural Networks with Cutout]
  • DropBlock [DropBlock: A regularization method for convolutional networks]
• SENet [Squeeze-and-Excitation Networks]
• EfficientNet [EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks]
• EfficientNetV2 [EfficientNetV2: Smaller Models and Faster Training]
• Transposed convolution
• U-Net [U-Net: Convolutional Networks for Biomedical Image Segmentation]

6. Object Detection

 Mar 25 Slides PDF Slides CZ Lecture CZ Practicals EN Lecture EN Practicals Questions bboxes_utils svhn_competition

• R-CNN [R-CNN]
• Fast R-CNN [Fast R-CNN]
• Proposing RoIs using Faster R-CNN [Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks]
• Mask R-CNN [Mask R-CNN]
• Feature Pyramid Networks [Feature Pyramid Networks for Object Detection]
• Focal Loss, RetinaNet [Focal Loss for Dense Object Detection]
• EfficientDet [EfficientDet: Scalable and Efficient Object Detection]
• Group Normalization [Group Normalization]

7. Easter Monday

 Apr 01 CZ Practicals EN Practicals 3d_recognition

8. Recurrent Neural Networks

 Apr 8 Slides PDF Slides CZ Lecture EN Lecture EN SVHN Competition EN Practicals Questions sequence_classification tagger_we tagger_cle tagger_competition

• Sequence modelling using Recurrent Neural Networks (RNN) [Chapter 10 until Section 10.2.1 (excluding) of DLB]
• The challenge of long-term dependencies [Section 10.7 of DLB]
• Long Short-Term Memory (LSTM) [Section 10.10.1 of DLB, Sepp Hochreiter, Jürgen Schmidhuber (1997): Long short-term memory, Felix A. Gers, Jürgen Schmidhuber, Fred Cummins (2000): Learning to Forget: Continual Prediction with LSTM]
• Gated Recurrent Unit (GRU) [Section 10.10.2 of DLB, Kyunghyun Cho, Bart van Merrienboer, Caglar Gulcehre, Dzmitry Bahdanau, Fethi Bougares, Holger Schwenk, Yoshua Bengio: Learning Phrase Representations using RNN Encoder-Decoder for Statistical Machine Translation]
• Highway Networks [Training Very Deep Networks]
• RNN Regularization
  • Variational Dropout [A Theoretically Grounded Application of Dropout in Recurrent Neural Networks]
  • Layer Normalization [Layer Normalization]
• Bidirectional RNN [Section 10.3 of DLB]
• Word Embeddings [Section 14.2.4 of DLB]
• Character-level embeddings using Recurrent neural networks [C2W model from Finding Function in Form: Compositional Character Models for Open Vocabulary Word Representation]
• Character-level embeddings using Convolutional neural networks [CharCNN from Character-Aware Neural Language Models]

9. Structured Prediction, CTC, Word2Vec

 Apr 15 Slides PDF Slides CZ Lecture CZ Practicals EN Lecture EN Practicals Questions tensorboard_projector tagger_ner ctc_loss speech_recognition

• Structured prediction
• Connectionist Temporal Classification (CTC) loss [Connectionist Temporal Classification: Labelling Unsegmented Sequence Data with Recurrent Neural Networks]
• Word2vec word embeddings, notably the CBOW and Skip-gram architectures [Efficient Estimation of Word Representations in Vector Space]
  • Hierarchical softmax [Section 12.4.3.2 of DLB or Distributed Representations of Words and Phrases and their Compositionality]
  • Negative sampling Distributed Representations of Words and Phrases and their Compositionality]
• Character-level embeddings using character n-grams [Described simultaneously in several papers as Charagram (Charagram: Embedding Words and Sentences via Character n-grams), Subword Information (Enriching Word Vectors with Subword Information or SubGram (SubGram: Extending Skip-Gram Word Representation with Substrings)]
• ELMO [Matthew E. Peters, Mark Neumann, Mohit Iyyer, Matt Gardner, Christopher Clark, Kenton Lee, Luke Zettlemoyer: Deep contextualized word representations]

10. Seq2seq, NMT, Transformer

 Apr 22 Slides PDF Slides CZ Lecture CZ Practicals EN Lecture EN Practicals Questions lemmatizer_noattn lemmatizer_attn lemmatizer_competition

• Neural Machine Translation using Encoder-Decoder or Sequence-to-Sequence architecture [Section 12.5.4 of DLB, Ilya Sutskever, Oriol Vinyals, Quoc V. Le: Sequence to Sequence Learning with Neural Networks and Kyunghyun Cho et al.: Learning Phrase Representations using RNN Encoder-Decoder for Statistical Machine Translation]
• Using Attention mechanism in Neural Machine Translation [Section 12.4.5.1 of DLB, Dzmitry Bahdanau, Kyunghyun Cho, Yoshua Bengio: Neural Machine Translation by Jointly Learning to Align and Translate]
• Translating Subword Units [Rico Sennrich, Barry Haddow, Alexandra Birch: Neural Machine Translation of Rare Words with Subword Units]
• Google NMT [Yonghui Wu et al.: Google's Neural Machine Translation System: Bridging the Gap between Human and Machine Translation]
• Transformer architecture [Attention Is All You Need]

Requirements

To pass the practicals, you need to obtain at least 80 points, excluding the bonus points. Note that all surplus points (both bonus and non-bonus) will be transfered to the exam. In total, assignments for at least 120 points (not including the bonus points) will be available, and if you solve all the assignments (any non-zero amount of points counts as solved), you automatically pass the exam with grade 1.

Environment

The tasks are evaluated automatically using the ReCodEx Code Examiner.

The evaluation is performed using Python 3.11, Keras 3.0.5, PyTorch 2.2.0, HF Transformers 4.37.2, and Gymnasium 1.0.0a. You should install the exact version of these packages yourselves.

Teamwork

Solving assignments in teams (of size at most 3) is encouraged, but everyone has to participate (it is forbidden not to work on an assignment and then submit a solution created by other team members). All members of the team must submit in ReCodEx individually, but can have exactly the same sources/models/results. Each such solution must explicitly list all members of the team to allow plagiarism detection using this template.

No Cheating

Cheating is strictly prohibited and any student found cheating will be punished. The punishment can involve failing the whole course, or, in grave cases, being expelled from the faculty. While discussing assignments with any classmate is fine, each team must complete the assignments themselves, without using code they did not write (unless explicitly allowed). Of course, inside a team you are allowed to share code and submit identical solutions. Note that all students involved in cheating will be punished, so if you share your source code with a friend, both you and your friend will be punished. That also means that you should never publish your solutions.

numpy_entropy

 Deadline: Mar 05, 22:00  3 points  Tests  Examples

The goal of this exercise is to familiarize with Python, NumPy and ReCodEx submission system. Start with the numpy_entropy.py.

Load a file specified in args.data_path, whose lines consist of data points of our dataset, and load a file specified in args.model_path, which describes a model probability distribution, with each line being a tab-separated pair of (data point, probability).

Then compute the following quantities using NumPy, and print them each on a separate line rounded on two decimal places (or inf for positive infinity, which happens when an element of data distribution has zero probability under the model distribution):

• entropy H(data distribution)
• cross-entropy H(data distribution, model distribution)
• KL-divergence D_KL(data distribution, model distribution)

Use natural logarithms to compute the entropies and the divergence.

Entropy: 0.96 nats
Crossentropy: 0.99 nats
KL divergence: 0.03 nats

Entropy: 0.96 nats
Crossentropy: inf nats
KL divergence: inf nats

Entropy: 0.96 nats
Crossentropy: 0.99 nats
KL divergence: 0.03 nats

Entropy: 0.96 nats
Crossentropy: inf nats
KL divergence: inf nats

Entropy: 4.15 nats
Crossentropy: 4.23 nats
KL divergence: 0.08 nats

Entropy: 4.99 nats
Crossentropy: 5.03 nats
KL divergence: 0.04 nats

pca_first

 Deadline: Mar 05, 22:00  2 points  Tests

The goal of this exercise is to familiarize with PyTorch torch.Tensors, shapes and basic tensor manipulation methods. Start with the pca_first.py (and you will also need the mnist.py module).

Alternatively, you can instead use the pca_first.keras.py template, which uses backend-agnostic keras.ops operations instead of PyTorch operations – both templates can be used to solve the assignment.

In this assignment, you should compute the covariance matrix of several examples from the MNIST dataset, then compute the first principal component, and quantify the explained variance of it. It is fine if you are not familiar with terms like covariance matrix or principal component – the template contains a detailed description of what you have to do.

Finally, you might want to read the Introduction to PyTorch Tensors.

Total variance: 53.12
Explained variance: 9.64%

Total variance: 53.05
Explained variance: 9.89%

Total variance: 52.74
Explained variance: 9.71%

mnist_layers_activations

 Deadline: Mar 05, 22:00  2 points  Tests  Examples

Before solving the assignment, start by playing with example_keras_tensorboard.py, in order to familiarize with Keras and TensorBoard. Run it, and when it finishes, run TensorBoard using tensorboard --logdir logs. Then open http://localhost:6006 in a browser and explore the active tabs.

Your goal is to modify the mnist_layers_activations.py template such that a user-specified neural network is constructed:

• A number of hidden layers (including zero) can be specified on the command line using parameter hidden_layers.
• Activation function of these hidden layers can be also specified as a command line parameter activation, with supported values of none, relu, tanh and sigmoid.

accuracy: 0.7801 - loss: 0.8405 - val_accuracy: 0.9300 - val_loss: 0.2716

accuracy: 0.8483 - loss: 0.5230 - val_accuracy: 0.9352 - val_loss: 0.2422

accuracy: 0.8503 - loss: 0.5286 - val_accuracy: 0.9604 - val_loss: 0.1432

accuracy: 0.8529 - loss: 0.5183 - val_accuracy: 0.9564 - val_loss: 0.1632

accuracy: 0.7851 - loss: 0.8650 - val_accuracy: 0.9414 - val_loss: 0.2196

accuracy: 0.8497 - loss: 0.5011 - val_accuracy: 0.9664 - val_loss: 0.1225

Epoch  1/10 accuracy: 0.7801 - loss: 0.8405 - val_accuracy: 0.9300 - val_loss: 0.2716
Epoch  5/10 accuracy: 0.9222 - loss: 0.2792 - val_accuracy: 0.9406 - val_loss: 0.2203
Epoch 10/10 accuracy: 0.9304 - loss: 0.2515 - val_accuracy: 0.9432 - val_loss: 0.2159

Epoch  1/10 accuracy: 0.8483 - loss: 0.5230 - val_accuracy: 0.9352 - val_loss: 0.2422
Epoch  5/10 accuracy: 0.9236 - loss: 0.2758 - val_accuracy: 0.9360 - val_loss: 0.2325
Epoch 10/10 accuracy: 0.9298 - loss: 0.2517 - val_accuracy: 0.9354 - val_loss: 0.2439

Epoch  1/10 accuracy: 0.8503 - loss: 0.5286 - val_accuracy: 0.9604 - val_loss: 0.1432
Epoch  5/10 accuracy: 0.9824 - loss: 0.0613 - val_accuracy: 0.9808 - val_loss: 0.0740
Epoch 10/10 accuracy: 0.9948 - loss: 0.0202 - val_accuracy: 0.9788 - val_loss: 0.0821

Epoch  1/10 accuracy: 0.8529 - loss: 0.5183 - val_accuracy: 0.9564 - val_loss: 0.1632
Epoch  5/10 accuracy: 0.9800 - loss: 0.0728 - val_accuracy: 0.9740 - val_loss: 0.0853
Epoch 10/10 accuracy: 0.9948 - loss: 0.0244 - val_accuracy: 0.9782 - val_loss: 0.0772

Epoch  1/10 accuracy: 0.7851 - loss: 0.8650 - val_accuracy: 0.9414 - val_loss: 0.2196
Epoch  5/10 accuracy: 0.9647 - loss: 0.1270 - val_accuracy: 0.9704 - val_loss: 0.1079
Epoch 10/10 accuracy: 0.9852 - loss: 0.0583 - val_accuracy: 0.9756 - val_loss: 0.0837

Epoch  1/10 accuracy: 0.8497 - loss: 0.5011 - val_accuracy: 0.9664 - val_loss: 0.1225
Epoch  5/10 accuracy: 0.9862 - loss: 0.0438 - val_accuracy: 0.9734 - val_loss: 0.1026
Epoch 10/10 accuracy: 0.9932 - loss: 0.0202 - val_accuracy: 0.9818 - val_loss: 0.0865

Epoch  1/10 accuracy: 0.7710 - loss: 0.6793 - val_accuracy: 0.9570 - val_loss: 0.1479
Epoch  5/10 accuracy: 0.9780 - loss: 0.0783 - val_accuracy: 0.9786 - val_loss: 0.0808
Epoch 10/10 accuracy: 0.9869 - loss: 0.0481 - val_accuracy: 0.9724 - val_loss: 0.1163

Epoch  1/10 accuracy: 0.1072 - loss: 2.3068 - val_accuracy: 0.1784 - val_loss: 2.1247
Epoch  5/10 accuracy: 0.8825 - loss: 0.4776 - val_accuracy: 0.9164 - val_loss: 0.3686
Epoch 10/10 accuracy: 0.9294 - loss: 0.2994 - val_accuracy: 0.9386 - val_loss: 0.2671

sgd_backpropagation

 Deadline: Mar 12, 22:00  3 points  Tests  Examples

In this exercise you will learn how to compute gradients using the so-called automatic differentiation, which allows to automatically run backpropagation algorithm for a given computation. You can read the Automatic Differentiation with torch.autograd tutorial if interested. After computing the gradient, you should then perform training by running manually implemented minibatch stochastic gradient descent.

Starting with the sgd_backpropagation.py template, you should:

• implement a neural network with a single tanh hidden layer and categorical output layer;
• compute the crossentropy loss;
• use .backward() to automatically compute the gradient of the loss with respect to all variables;
• perform the SGD update.

Dev accuracy after epoch 1 is 93.30
Dev accuracy after epoch 2 is 94.38
Test accuracy after epoch 2 is 93.15

Dev accuracy after epoch 1 is 93.64
Dev accuracy after epoch 2 is 94.80
Test accuracy after epoch 2 is 93.54

Dev accuracy after epoch 1 is 93.30
Dev accuracy after epoch 2 is 94.38
Dev accuracy after epoch 3 is 95.16
Dev accuracy after epoch 4 is 95.50
Dev accuracy after epoch 5 is 95.96
Dev accuracy after epoch 6 is 96.04
Dev accuracy after epoch 7 is 95.82
Dev accuracy after epoch 8 is 95.92
Dev accuracy after epoch 9 is 95.96
Dev accuracy after epoch 10 is 96.16
Test accuracy after epoch 10 is 95.26

Dev accuracy after epoch 1 is 93.64
Dev accuracy after epoch 2 is 94.80
Dev accuracy after epoch 3 is 95.56
Dev accuracy after epoch 4 is 95.98
Dev accuracy after epoch 5 is 96.24
Dev accuracy after epoch 6 is 96.74
Dev accuracy after epoch 7 is 96.52
Dev accuracy after epoch 8 is 96.54
Dev accuracy after epoch 9 is 97.04
Dev accuracy after epoch 10 is 97.02
Test accuracy after epoch 10 is 96.16

sgd_manual

 Deadline: Mar 12, 22:00  2 points  Tests  Examples

The goal in this exercise is to extend your solution to the sgd_backpropagation assignment by manually computing the gradient.

While in this assignment we compute the gradient manually, we will nearly always use the automatic differentiation. Therefore, the assignment is more of a mathematical exercise than a real-world application. Furthermore, we will compute the derivatives together on the Mar 06 practicals.

Start with the sgd_manual.py template, which is based on sgd_backpropagation.py one.

Note that ReCodEx disables the PyTorch automatic differentiation during evaluation.

Dev accuracy after epoch 1 is 93.30
Dev accuracy after epoch 2 is 94.38
Test accuracy after epoch 2 is 93.15

Dev accuracy after epoch 1 is 93.64
Dev accuracy after epoch 2 is 94.80
Test accuracy after epoch 2 is 93.54

Dev accuracy after epoch 1 is 93.30
Dev accuracy after epoch 2 is 94.38
Dev accuracy after epoch 3 is 95.16
Dev accuracy after epoch 4 is 95.50
Dev accuracy after epoch 5 is 95.96
Dev accuracy after epoch 6 is 96.04
Dev accuracy after epoch 7 is 95.82
Dev accuracy after epoch 8 is 95.92
Dev accuracy after epoch 9 is 95.96
Dev accuracy after epoch 10 is 96.16
Test accuracy after epoch 10 is 95.26

Dev accuracy after epoch 1 is 93.64
Dev accuracy after epoch 2 is 94.80
Dev accuracy after epoch 3 is 95.56
Dev accuracy after epoch 4 is 95.98
Dev accuracy after epoch 5 is 96.24
Dev accuracy after epoch 6 is 96.74
Dev accuracy after epoch 7 is 96.52
Dev accuracy after epoch 8 is 96.54
Dev accuracy after epoch 9 is 97.04
Dev accuracy after epoch 10 is 97.02
Test accuracy after epoch 10 is 96.16

mnist_training

 Deadline: Mar 12, 22:00  2 points  Tests  Examples

This exercise should teach you using different optimizers, learning rates, and learning rate decays. Your goal is to modify the mnist_training.py template and implement the following:

• Using specified optimizer (either SGD or Adam).
• Optionally using momentum for the SGD optimizer.
• Using specified learning rate for the optimizer.
• Optionally use a given learning rate schedule. The schedule can be either linear, exponential, or cosine. If a schedule is specified, you also get a final learning rate, and the learning rate should be gradually decresed during training to reach the final learning rate just after the training (i.e., the first update after the training would use exactly the final learning rate).

accuracy: 0.6537 - loss: 1.2786 - val_accuracy: 0.9098 - val_loss: 0.3743

accuracy: 0.8221 - loss: 0.6138 - val_accuracy: 0.9492 - val_loss: 0.1873

accuracy: 0.8400 - loss: 0.5742 - val_accuracy: 0.9528 - val_loss: 0.1800

accuracy: 0.8548 - loss: 0.5121 - val_accuracy: 0.9640 - val_loss: 0.1327

accuracy: 0.8858 - loss: 0.3598 - val_accuracy: 0.9564 - val_loss: 0.1393

Epoch 1/2 accuracy: 0.8889 - loss: 0.3520 - val_accuracy: 0.9682 - val_loss: 0.1107
Epoch 2/2 accuracy: 0.9715 - loss: 0.0956 - val_accuracy: 0.9792 - val_loss: 0.0688
Next learning rate to be used: 0.0001

Epoch 1/2 accuracy: 0.8912 - loss: 0.3447 - val_accuracy: 0.9702 - val_loss: 0.0997
Epoch 2/2 accuracy: 0.9746 - loss: 0.0824 - val_accuracy: 0.9778 - val_loss: 0.0776
Next learning rate to be used: 0.001

Epoch 1/2 accuracy: 0.8875 - loss: 0.3548 - val_accuracy: 0.9726 - val_loss: 0.0976
Epoch 2/2 accuracy: 0.9742 - loss: 0.0851 - val_accuracy: 0.9764 - val_loss: 0.0740
Next learning rate to be used: 0.0001

Epoch  1/10 accuracy: 0.6537 - loss: 1.2786 - val_accuracy: 0.9098 - val_loss: 0.3743
Epoch  2/10 accuracy: 0.8848 - loss: 0.4316 - val_accuracy: 0.9222 - val_loss: 0.2895
Epoch  3/10 accuracy: 0.9057 - loss: 0.3450 - val_accuracy: 0.9308 - val_loss: 0.2539
Epoch  4/10 accuracy: 0.9118 - loss: 0.3131 - val_accuracy: 0.9372 - val_loss: 0.2368
Epoch  5/10 accuracy: 0.9188 - loss: 0.2924 - val_accuracy: 0.9406 - val_loss: 0.2202
Epoch  6/10 accuracy: 0.9235 - loss: 0.2750 - val_accuracy: 0.9426 - val_loss: 0.2076
Epoch  7/10 accuracy: 0.9291 - loss: 0.2572 - val_accuracy: 0.9464 - val_loss: 0.1997
Epoch  8/10 accuracy: 0.9304 - loss: 0.2456 - val_accuracy: 0.9494 - val_loss: 0.1909
Epoch  9/10 accuracy: 0.9339 - loss: 0.2340 - val_accuracy: 0.9536 - val_loss: 0.1813
Epoch 10/10 accuracy: 0.9377 - loss: 0.2199 - val_accuracy: 0.9534 - val_loss: 0.1756

Epoch  1/10 accuracy: 0.8221 - loss: 0.6138 - val_accuracy: 0.9492 - val_loss: 0.1873
Epoch  2/10 accuracy: 0.9370 - loss: 0.2173 - val_accuracy: 0.9646 - val_loss: 0.1385
Epoch  3/10 accuracy: 0.9599 - loss: 0.1453 - val_accuracy: 0.9716 - val_loss: 0.1076
Epoch  4/10 accuracy: 0.9673 - loss: 0.1127 - val_accuracy: 0.9746 - val_loss: 0.0961
Epoch  5/10 accuracy: 0.9740 - loss: 0.0933 - val_accuracy: 0.9774 - val_loss: 0.0875
Epoch  6/10 accuracy: 0.9778 - loss: 0.0811 - val_accuracy: 0.9746 - val_loss: 0.0856
Epoch  7/10 accuracy: 0.9821 - loss: 0.0680 - val_accuracy: 0.9774 - val_loss: 0.0803
Epoch  8/10 accuracy: 0.9825 - loss: 0.0632 - val_accuracy: 0.9776 - val_loss: 0.0780
Epoch  9/10 accuracy: 0.9849 - loss: 0.0552 - val_accuracy: 0.9804 - val_loss: 0.0725
Epoch 10/10 accuracy: 0.9877 - loss: 0.0463 - val_accuracy: 0.9780 - val_loss: 0.0735

Epoch  1/10 accuracy: 0.8400 - loss: 0.5742 - val_accuracy: 0.9528 - val_loss: 0.1800
Epoch  2/10 accuracy: 0.9389 - loss: 0.2123 - val_accuracy: 0.9670 - val_loss: 0.1335
Epoch  3/10 accuracy: 0.9602 - loss: 0.1431 - val_accuracy: 0.9728 - val_loss: 0.1052
Epoch  4/10 accuracy: 0.9685 - loss: 0.1115 - val_accuracy: 0.9770 - val_loss: 0.0946
Epoch  5/10 accuracy: 0.9747 - loss: 0.0927 - val_accuracy: 0.9754 - val_loss: 0.0878
Epoch  6/10 accuracy: 0.9775 - loss: 0.0798 - val_accuracy: 0.9754 - val_loss: 0.0852
Epoch  7/10 accuracy: 0.9813 - loss: 0.0680 - val_accuracy: 0.9780 - val_loss: 0.0797
Epoch  8/10 accuracy: 0.9828 - loss: 0.0621 - val_accuracy: 0.9796 - val_loss: 0.0757
Epoch  9/10 accuracy: 0.9847 - loss: 0.0550 - val_accuracy: 0.9804 - val_loss: 0.0731
Epoch 10/10 accuracy: 0.9875 - loss: 0.0464 - val_accuracy: 0.9782 - val_loss: 0.0731

Epoch  1/10 accuracy: 0.8548 - loss: 0.5121 - val_accuracy: 0.9640 - val_loss: 0.1327
Epoch  2/10 accuracy: 0.9552 - loss: 0.1505 - val_accuracy: 0.9706 - val_loss: 0.1118
Epoch  3/10 accuracy: 0.9744 - loss: 0.0900 - val_accuracy: 0.9770 - val_loss: 0.0833
Epoch  4/10 accuracy: 0.9808 - loss: 0.0658 - val_accuracy: 0.9778 - val_loss: 0.0786
Epoch  5/10 accuracy: 0.9836 - loss: 0.0533 - val_accuracy: 0.9804 - val_loss: 0.0735
Epoch  6/10 accuracy: 0.9890 - loss: 0.0403 - val_accuracy: 0.9782 - val_loss: 0.0772
Epoch  7/10 accuracy: 0.9911 - loss: 0.0311 - val_accuracy: 0.9792 - val_loss: 0.0756
Epoch  8/10 accuracy: 0.9922 - loss: 0.0257 - val_accuracy: 0.9818 - val_loss: 0.0717
Epoch  9/10 accuracy: 0.9947 - loss: 0.0202 - val_accuracy: 0.9806 - val_loss: 0.0734
Epoch 10/10 accuracy: 0.9953 - loss: 0.0167 - val_accuracy: 0.9802 - val_loss: 0.0779

Epoch  1/10 accuracy: 0.8858 - loss: 0.3598 - val_accuracy: 0.9564 - val_loss: 0.1393
Epoch  2/10 accuracy: 0.9565 - loss: 0.1478 - val_accuracy: 0.9622 - val_loss: 0.1445
Epoch  3/10 accuracy: 0.9688 - loss: 0.1041 - val_accuracy: 0.9686 - val_loss: 0.1184
Epoch  4/10 accuracy: 0.9717 - loss: 0.1016 - val_accuracy: 0.9644 - val_loss: 0.1538
Epoch  5/10 accuracy: 0.9749 - loss: 0.0914 - val_accuracy: 0.9642 - val_loss: 0.1477
Epoch  6/10 accuracy: 0.9754 - loss: 0.0878 - val_accuracy: 0.9714 - val_loss: 0.1375
Epoch  7/10 accuracy: 0.9779 - loss: 0.0804 - val_accuracy: 0.9684 - val_loss: 0.1510
Epoch  8/10 accuracy: 0.9793 - loss: 0.0764 - val_accuracy: 0.9696 - val_loss: 0.1803
Epoch  9/10 accuracy: 0.9808 - loss: 0.0747 - val_accuracy: 0.9708 - val_loss: 0.1576
Epoch 10/10 accuracy: 0.9812 - loss: 0.0750 - val_accuracy: 0.9716 - val_loss: 0.1556

Epoch  1/10 accuracy: 0.8862 - loss: 0.3582 - val_accuracy: 0.9636 - val_loss: 0.1395
Epoch  2/10 accuracy: 0.9603 - loss: 0.1313 - val_accuracy: 0.9684 - val_loss: 0.1056
Epoch  3/10 accuracy: 0.9730 - loss: 0.0899 - val_accuracy: 0.9718 - val_loss: 0.1089
Epoch  4/10 accuracy: 0.9780 - loss: 0.0701 - val_accuracy: 0.9676 - val_loss: 0.1250
Epoch  5/10 accuracy: 0.9818 - loss: 0.0528 - val_accuracy: 0.9744 - val_loss: 0.1001
Epoch  6/10 accuracy: 0.9876 - loss: 0.0389 - val_accuracy: 0.9738 - val_loss: 0.1233
Epoch  7/10 accuracy: 0.9907 - loss: 0.0255 - val_accuracy: 0.9780 - val_loss: 0.0989
Epoch  8/10 accuracy: 0.9954 - loss: 0.0141 - val_accuracy: 0.9802 - val_loss: 0.0909
Epoch  9/10 accuracy: 0.9976 - loss: 0.0079 - val_accuracy: 0.9814 - val_loss: 0.0923
Epoch 10/10 accuracy: 0.9995 - loss: 0.0033 - val_accuracy: 0.9818 - val_loss: 0.0946
Next learning rate to be used: 0.0001

Epoch  1/10 accuracy: 0.8877 - loss: 0.3564 - val_accuracy: 0.9616 - val_loss: 0.1278
Epoch  2/10 accuracy: 0.9642 - loss: 0.1228 - val_accuracy: 0.9624 - val_loss: 0.1149
Epoch  3/10 accuracy: 0.9778 - loss: 0.0720 - val_accuracy: 0.9748 - val_loss: 0.0781
Epoch  4/10 accuracy: 0.9844 - loss: 0.0500 - val_accuracy: 0.9750 - val_loss: 0.0973
Epoch  5/10 accuracy: 0.9884 - loss: 0.0356 - val_accuracy: 0.9800 - val_loss: 0.0709
Epoch  6/10 accuracy: 0.9933 - loss: 0.0228 - val_accuracy: 0.9792 - val_loss: 0.0810
Epoch  7/10 accuracy: 0.9956 - loss: 0.0150 - val_accuracy: 0.9806 - val_loss: 0.0785
Epoch  8/10 accuracy: 0.9969 - loss: 0.0095 - val_accuracy: 0.9826 - val_loss: 0.0746
Epoch  9/10 accuracy: 0.9985 - loss: 0.0069 - val_accuracy: 0.9808 - val_loss: 0.0783
Epoch 10/10 accuracy: 0.9994 - loss: 0.0036 - val_accuracy: 0.9818 - val_loss: 0.0783
Next learning rate to be used: 0.001

Epoch  1/10 accuracy: 0.8858 - loss: 0.3601 - val_accuracy: 0.9624 - val_loss: 0.1311
Epoch  2/10 accuracy: 0.9566 - loss: 0.1461 - val_accuracy: 0.9654 - val_loss: 0.1270
Epoch  3/10 accuracy: 0.9695 - loss: 0.1023 - val_accuracy: 0.9740 - val_loss: 0.0965
Epoch  4/10 accuracy: 0.9755 - loss: 0.0790 - val_accuracy: 0.9710 - val_loss: 0.1152
Epoch  5/10 accuracy: 0.9831 - loss: 0.0562 - val_accuracy: 0.9748 - val_loss: 0.1004
Epoch  6/10 accuracy: 0.9889 - loss: 0.0353 - val_accuracy: 0.9758 - val_loss: 0.1003
Epoch  7/10 accuracy: 0.9930 - loss: 0.0206 - val_accuracy: 0.9786 - val_loss: 0.0864
Epoch  8/10 accuracy: 0.9972 - loss: 0.0096 - val_accuracy: 0.9790 - val_loss: 0.0958
Epoch  9/10 accuracy: 0.9985 - loss: 0.0068 - val_accuracy: 0.9802 - val_loss: 0.0880
Epoch 10/10 accuracy: 0.9992 - loss: 0.0042 - val_accuracy: 0.9802 - val_loss: 0.0891
Next learning rate to be used: 0.0001

gym_cartpole

 Deadline: Mar 12, 22:00  3 points

Solve the CartPole-v1 environment from the Gymnasium library, utilizing only provided supervised training data. The data is available in gym_cartpole_data.txt file, each line containing one observation (four space separated floats) and a corresponding action (the last space separated integer). Start with the gym_cartpole.py.

The solution to this task should be a model which passes evaluation on random inputs. This evaluation can be performed by running the gym_cartpole.py with --evaluate argument (optionally rendering if --render option is provided), or directly calling the evaluate_model method. In order to pass, you must achieve an average reward of at least 475 on 100 episodes. Your model should have either one or two outputs (i.e., using either sigmoid or softmax output function).

When designing the model, you should consider that the size of the training data is very small and the data is quite noisy.

When submitting to ReCodEx, do not forget to also submit the trained model.

mnist_regularization

 Deadline: Mar 19, 22:00  3 points  Tests

You will learn how to implement three regularization methods in this assignment. Start with the mnist_regularization.py template and implement the following:

• Allow using dropout with rate args.dropout. Add a dropout layer after the first Flatten and also after all Dense hidden layers (but not after the output layer).
• Allow using AdamW with weight decay with strength of args.weight_decay, making sure the weight decay is not applied on bias.
• Allow using label smoothing with weight args.label_smoothing. Instead of SparseCategoricalCrossentropy, you will need to use CategoricalCrossentropy which offers label_smoothing argument.

In addition to submitting the task in ReCodEx, also run the following variations and observe the results in TensorBoard, notably the training, development and test set accuracy and loss:

• dropout rate 0, 0.3, 0.5, 0.6, 0.8;
• weight decay 0, 0.1, 0.3, 0.5, 1.0;
• label smoothing 0, 0.1, 0.3, 0.5.

accuracy: 0.5981 - loss: 1.2688 - val_accuracy: 0.9174 - val_loss: 0.3051

accuracy: 0.3429 - loss: 1.9163 - val_accuracy: 0.8826 - val_loss: 0.4937

accuracy: 0.7014 - loss: 1.0412 - val_accuracy: 0.9236 - val_loss: 0.2776

accuracy: 0.7006 - loss: 1.0429 - val_accuracy: 0.9232 - val_loss: 0.2801

accuracy: 0.7102 - loss: 1.3015 - val_accuracy: 0.9276 - val_loss: 0.7656

accuracy: 0.7113 - loss: 1.6854 - val_accuracy: 0.9332 - val_loss: 1.3709

mnist_ensemble

 Deadline: Mar 19, 22:00  2 points  Tests  Examples

Your goal in this assignment is to implement model ensembling. The mnist_ensemble.py template trains args.models individual models, and your goal is to perform an ensemble of the first model, first two models, first three models, …, all models, and evaluate their accuracy on the development set.

Model 1, individual accuracy 96.04, ensemble accuracy 96.04
Model 2, individual accuracy 96.28, ensemble accuracy 96.56
Model 3, individual accuracy 96.12, ensemble accuracy 96.58
Model 4, individual accuracy 95.92, ensemble accuracy 96.70
Model 5, individual accuracy 96.38, ensemble accuracy 96.72

Model 1, individual accuracy 96.46, ensemble accuracy 96.46
Model 2, individual accuracy 96.86, ensemble accuracy 96.88
Model 3, individual accuracy 96.54, ensemble accuracy 97.04
Model 4, individual accuracy 96.54, ensemble accuracy 97.06
Model 5, individual accuracy 96.82, ensemble accuracy 97.20

Model 1, individual accuracy 97.82, ensemble accuracy 97.82
Model 2, individual accuracy 97.80, ensemble accuracy 98.08
Model 3, individual accuracy 98.02, ensemble accuracy 98.20
Model 4, individual accuracy 98.20, ensemble accuracy 98.28
Model 5, individual accuracy 97.64, ensemble accuracy 98.28

Model 1, individual accuracy 98.12, ensemble accuracy 98.12
Model 2, individual accuracy 98.22, ensemble accuracy 98.42
Model 3, individual accuracy 98.26, ensemble accuracy 98.52
Model 4, individual accuracy 98.32, ensemble accuracy 98.62
Model 5, individual accuracy 97.98, ensemble accuracy 98.70

uppercase

 Deadline: Mar 19, 22:00  4 points+5 bonus

This assignment introduces first NLP task. Your goal is to implement a model which is given Czech lowercased text and tries to uppercase appropriate letters. To load the dataset, use uppercase_data.py module which loads (and if required also downloads) the data. While the training and the development sets are in correct case, the test set is lowercased.

This is an open-data task, where you submit only the uppercased test set together with the training script (which will not be executed, it will be only used to understand the approach you took, and to indicate teams). Explicitly, submit exactly one .txt file and at least one .py/ipynb file.

The task is also a competition. Everyone who submits a solution achieving at least 98.5% accuracy gets 4 basic points; the remaining 5 bonus points are distributed depending on relative ordering of your solutions. The accuracy is computed per-character and can be evaluated by running uppercase_data.py with --evaluate argument, or using its evaluate_file method.

You may want to start with the uppercase.py template, which uses the uppercase_data.py to load the data, generate an alphabet of given size containing most frequent characters, and generate sliding window view on the data. The template also comments on possibilities of character representation.

Do not use RNNs, CNNs, or Transformer in this task (if you have doubts, contact me); fully connected layers (and therefore also embedding layers), any activations, residual connections, and any regularization layers are fine.

mnist_cnn

 Deadline: Mar 26, 22:00  3 points  Tests  Examples

To pass this assignment, you will learn to construct basic convolutional neural network layers. Start with the mnist_cnn.py template and assume the requested architecture is described by the cnn argument, which contains comma-separated specifications of the following layers:

• C-filters-kernel_size-stride-padding: Add a convolutional layer with ReLU activation and specified number of filters, kernel size, stride and padding. Example: C-10-3-1-same
• CB-filters-kernel_size-stride-padding: Same as C-filters-kernel_size-stride-padding, but use batch normalization. In detail, start with a convolutional layer without bias and activation, then add batch normalization layer, and finally ReLU activation. Example: CB-10-3-1-same
• M-pool_size-stride: Add max pooling with specified size and stride, using the default "valid" padding. Example: M-3-2
• R-[layers]: Add a residual connection. The layers contain a specification of at least one convolutional layer (but not a recursive residual connection R). The input to the R layer should be processed sequentially by layers, and the produced output (after the ReLU nonlinearity of the last layer) should be added to the input (of this R layer). Example: R-[C-16-3-1-same,C-16-3-1-same]
• F: Flatten inputs. Must appear exactly once in the architecture.
• H-hidden_layer_size: Add a dense layer with ReLU activation and specified size. Example: H-100
• D-dropout_rate: Apply dropout with the given dropout rate. Example: D-0.5

An example architecture might be --cnn=CB-16-5-2-same,M-3-2,F,H-100,D-0.5. You can assume the resulting network is valid; it is fine to crash if it is not.

After a successful ReCodEx submission, you can try obtaining the best accuracy on MNIST and then advance to cifar_competition.

accuracy: 0.8503 - loss: 0.5286 - val_accuracy: 0.9604 - val_loss: 0.1432

accuracy: 0.7706 - loss: 0.7444 - val_accuracy: 0.9572 - val_loss: 0.1606

accuracy: 0.6630 - loss: 1.0703 - val_accuracy: 0.8798 - val_loss: 0.3894

accuracy: 0.5898 - loss: 1.2535 - val_accuracy: 0.8774 - val_loss: 0.4079

accuracy: 0.6822 - loss: 1.0011 - val_accuracy: 0.9284 - val_loss: 0.2537

accuracy: 0.7562 - loss: 0.7717 - val_accuracy: 0.9486 - val_loss: 0.1734

Epoch  1/10 accuracy: 0.8503 - loss: 0.5286 - val_accuracy: 0.9604 - val_loss: 0.1432
Epoch  2/10 accuracy: 0.9508 - loss: 0.1654 - val_accuracy: 0.9650 - val_loss: 0.1245
Epoch  3/10 accuracy: 0.9710 - loss: 0.1034 - val_accuracy: 0.9738 - val_loss: 0.0916
Epoch  4/10 accuracy: 0.9773 - loss: 0.0774 - val_accuracy: 0.9762 - val_loss: 0.0848
Epoch  5/10 accuracy: 0.9824 - loss: 0.0613 - val_accuracy: 0.9808 - val_loss: 0.0740
Epoch  6/10 accuracy: 0.9857 - loss: 0.0485 - val_accuracy: 0.9760 - val_loss: 0.0761
Epoch  7/10 accuracy: 0.9893 - loss: 0.0373 - val_accuracy: 0.9770 - val_loss: 0.0774
Epoch  8/10 accuracy: 0.9911 - loss: 0.0323 - val_accuracy: 0.9774 - val_loss: 0.0813
Epoch  9/10 accuracy: 0.9922 - loss: 0.0271 - val_accuracy: 0.9794 - val_loss: 0.0819
Epoch 10/10 accuracy: 0.9948 - loss: 0.0202 - val_accuracy: 0.9788 - val_loss: 0.0821

Epoch  1/10 accuracy: 0.7706 - loss: 0.7444 - val_accuracy: 0.9572 - val_loss: 0.1606
Epoch  2/10 accuracy: 0.9177 - loss: 0.2808 - val_accuracy: 0.9646 - val_loss: 0.1286
Epoch  3/10 accuracy: 0.9313 - loss: 0.2340 - val_accuracy: 0.9732 - val_loss: 0.1038
Epoch  4/10 accuracy: 0.9389 - loss: 0.2025 - val_accuracy: 0.9730 - val_loss: 0.0951
Epoch  5/10 accuracy: 0.9409 - loss: 0.1919 - val_accuracy: 0.9752 - val_loss: 0.0927
Epoch  6/10 accuracy: 0.9448 - loss: 0.1784 - val_accuracy: 0.9768 - val_loss: 0.0864
Epoch  7/10 accuracy: 0.9495 - loss: 0.1649 - val_accuracy: 0.9758 - val_loss: 0.0833
Epoch  8/10 accuracy: 0.9506 - loss: 0.1577 - val_accuracy: 0.9768 - val_loss: 0.0826
Epoch  9/10 accuracy: 0.9544 - loss: 0.1496 - val_accuracy: 0.9778 - val_loss: 0.0806
Epoch 10/10 accuracy: 0.9560 - loss: 0.1413 - val_accuracy: 0.9754 - val_loss: 0.0792

Epoch  1/10 accuracy: 0.8109 - loss: 0.6191 - val_accuracy: 0.9654 - val_loss: 0.1286
Epoch  2/10 accuracy: 0.9382 - loss: 0.2101 - val_accuracy: 0.9718 - val_loss: 0.0995
Epoch  3/10 accuracy: 0.9530 - loss: 0.1598 - val_accuracy: 0.9752 - val_loss: 0.0820
Epoch  4/10 accuracy: 0.9586 - loss: 0.1377 - val_accuracy: 0.9792 - val_loss: 0.0758
Epoch  5/10 accuracy: 0.9635 - loss: 0.1233 - val_accuracy: 0.9792 - val_loss: 0.0684
Epoch  6/10 accuracy: 0.9639 - loss: 0.1133 - val_accuracy: 0.9800 - val_loss: 0.0709
Epoch  7/10 accuracy: 0.9698 - loss: 0.1003 - val_accuracy: 0.9822 - val_loss: 0.0647
Epoch  8/10 accuracy: 0.9701 - loss: 0.0945 - val_accuracy: 0.9814 - val_loss: 0.0626
Epoch  9/10 accuracy: 0.9720 - loss: 0.0886 - val_accuracy: 0.9810 - val_loss: 0.0658
Epoch 10/10 accuracy: 0.9727 - loss: 0.0843 - val_accuracy: 0.9816 - val_loss: 0.0643

Epoch  1/10 accuracy: 0.8549 - loss: 0.4564 - val_accuracy: 0.9836 - val_loss: 0.0529
Epoch  2/10 accuracy: 0.9809 - loss: 0.0610 - val_accuracy: 0.9830 - val_loss: 0.0527
Epoch  3/10 accuracy: 0.9878 - loss: 0.0406 - val_accuracy: 0.9902 - val_loss: 0.0303
Epoch  4/10 accuracy: 0.9905 - loss: 0.0309 - val_accuracy: 0.9872 - val_loss: 0.0444
Epoch  5/10 accuracy: 0.9916 - loss: 0.0247 - val_accuracy: 0.9918 - val_loss: 0.0286
Epoch  6/10 accuracy: 0.9930 - loss: 0.0214 - val_accuracy: 0.9924 - val_loss: 0.0286
Epoch  7/10 accuracy: 0.9941 - loss: 0.0184 - val_accuracy: 0.9910 - val_loss: 0.0318
Epoch  8/10 accuracy: 0.9955 - loss: 0.0135 - val_accuracy: 0.9944 - val_loss: 0.0236
Epoch  9/10 accuracy: 0.9963 - loss: 0.0116 - val_accuracy: 0.9928 - val_loss: 0.0262
Epoch 10/10 accuracy: 0.9953 - loss: 0.0126 - val_accuracy: 0.9916 - val_loss: 0.0309

Epoch  1/10 accuracy: 0.8951 - loss: 0.3258 - val_accuracy: 0.9868 - val_loss: 0.0435
Epoch  2/10 accuracy: 0.9834 - loss: 0.0514 - val_accuracy: 0.9866 - val_loss: 0.0479
Epoch  3/10 accuracy: 0.9879 - loss: 0.0401 - val_accuracy: 0.9898 - val_loss: 0.0351
Epoch  4/10 accuracy: 0.9904 - loss: 0.0297 - val_accuracy: 0.9886 - val_loss: 0.0441
Epoch  5/10 accuracy: 0.9918 - loss: 0.0245 - val_accuracy: 0.9940 - val_loss: 0.0233
Epoch  6/10 accuracy: 0.9937 - loss: 0.0195 - val_accuracy: 0.9898 - val_loss: 0.0336
Epoch  7/10 accuracy: 0.9934 - loss: 0.0203 - val_accuracy: 0.9934 - val_loss: 0.0229
Epoch  8/10 accuracy: 0.9951 - loss: 0.0139 - val_accuracy: 0.9938 - val_loss: 0.0260
Epoch  9/10 accuracy: 0.9958 - loss: 0.0127 - val_accuracy: 0.9938 - val_loss: 0.0248
Epoch 10/10 accuracy: 0.9954 - loss: 0.0132 - val_accuracy: 0.9934 - val_loss: 0.0217

torch_dataset

 Deadline: Mar 26, 22:00  2 points  Tests

In this assignment you will familiarize yourselves with torch.utils.data, which is a PyTorch way of constructing training datasets. If you want, you can read the Dataset and DataLoaders tutorial.

The goal of this assignment is to start with the torch_dataset.py template and implement a simple image augmentation preprocessing.

accuracy: 0.1297 - loss: 2.2519 - val_accuracy: 0.2710 - val_loss: 1.9796

accuracy: 0.1354 - loss: 2.2565 - val_accuracy: 0.2690 - val_loss: 1.9889

mnist_multiple

 Deadline: Mar 26, 22:00  3 points  Tests

In this assignment you will implement a model with multiple inputs and outputs. Start with the mnist_multiple.py template and:

• The goal is to create a model, which given two input MNIST images, compares if the digit on the first one is greater than on the second one.
• We perform this comparison in two different ways:
  • first by directly predicting the comparison by the network (direct comparison),
  • then by first classifying the images into digits and then comparing these predictions (indirect comparison).
• The model has four outputs:
  • direct comparison whether the first digit is greater than the second one,
  • digit classification for the first image,
  • digit classification for the second image,
  • indirect comparison comparing the digits predicted by the above two outputs.
• You need to implement:
  • the model, using multiple inputs, outputs, losses and metrics;
  • construction of two-image dataset examples using regular MNIST data via the PyTorch datasets.

direct_comparison_accuracy: 0.7993 - indirect_comparison_accuracy: 0.8930 - loss: 1.6710 - val_direct_comparison_accuracy: 0.9508 - val_indirect_comparison_accuracy: 0.9836 - val_loss: 0.2984

direct_comparison_accuracy: 0.7680 - indirect_comparison_accuracy: 0.8637 - loss: 2.1429 - val_direct_comparison_accuracy: 0.9288 - val_indirect_comparison_accuracy: 0.9772 - val_loss: 0.4157

cifar_competition

 Deadline: Mar 26, 22:00  4 points+5 bonus

The goal of this assignment is to devise the best possible model for CIFAR-10. You can load the data using the cifar10.py module. Note that the test set is different than that of official CIFAR-10.

The task is a competition. Everyone who submits a solution achieving at least 70% test set accuracy gets 4 points; the remaining 5 bonus points are distributed depending on relative ordering of your solutions.

Note that my solutions usually need to achieve around ~85% on the development set to score 70% on the test set.

You may want to start with the cifar_competition.py template which generates the test set annotation in the required format.

Note that in this assignment, you cannot use the keras.applications module.

cnn_manual

 Deadline: Apr 02, 22:00  3 points Slides  Tests

To pass this assignment, you need to manually implement the forward and backward pass through a 2D convolutional layer. Start with the cnn_manual.py template, which constructs a series of 2D convolutional layers with ReLU activation and valid padding, specified in the args.cnn option. The args.cnn contains comma-separated layer specifications in the format filters-kernel_size-stride.

Of course, you cannot use any PyTorch convolutional operation (instead, implement the forward and backward pass using matrix multiplication and other operations), nor the .backward() for gradient computation.

To make debugging easier, the template supports a --verify option, which allows comparing the forward pass and the three gradients you compute in the backward pass to correct values.

Dev accuracy after epoch 1 is 90.42
Test accuracy after epoch 1 is 88.55

Dev accuracy after epoch 1 is 92.26
Test accuracy after epoch 1 is 90.59

Dev accuracy after epoch 1 is 90.82
Test accuracy after epoch 1 is 88.78

Dev accuracy after epoch 1 is 92.92
Test accuracy after epoch 1 is 90.97

cags_classification

 Deadline: Apr 02, 22:00  4 points+5 bonus

The goal of this assignment is to use a pretrained model, for example the EfficientNetV2-B0, to achieve best accuracy in CAGS classification.

The CAGS dataset consists of images of cats and dogs of size $224×224$, each classified in one of the 34 breeds and each containing a mask indicating the presence of the animal. To load the dataset, use the cags_dataset.py module.

To load the EfficientNetV2-B0, use the keras.applications.EfficientNetV2B0 function, which constructs a Keras model, downloading the weights automatically. However, you can use any model from keras.applications in this assignment.

An example performing classification of given images is available in image_classification.py.

A note on finetuning: each keras.layers.Layer has a mutable trainable property indicating whether its variables should be updated – however, after changing it, you need to call .compile again (or otherwise make sure the list of trainable variables for the optimizer is updated). Furthermore, training argument passed to the invocation call decides whether the layer is executed in training regime (neurons gets dropped in dropout, batch normalization computes estimates on the batch) or in inference regime. There is one exception though – if trainable == False on a batch normalization layer, it runs in the inference regime even when training == True.

The task is a competition. Everyone who submits a solution achieving at least 93% test set accuracy gets 4 points; the remaining 5 bonus points are distributed depending on relative ordering of your solutions.

You may want to start with the cags_classification.py template which generates the test set annotation in the required format.

cags_segmentation

 Deadline: Apr 02, 22:00  4 points+5 bonus

The goal of this assignment is to use a pretrained model, for example the EfficientNetV2-B0, to achieve best image segmentation IoU score on the CAGS dataset. The dataset and the EfficientNetV2-B0 is described in the cags_classification assignment. Nevertheless, you can again use any model from keras.applications in this assignment.

A mask is evaluated using intersection over union (IoU) metric, which is the intersection of the gold and predicted mask divided by their union, and the whole test set score is the average of its masks' IoU. A Keras-compatible metric is implemented by the class MaskIoUMetric of the cags_dataset.py module, which can also evaluate your predictions (either by running with --task=segmentation --evaluate=path arguments, or using its evaluate_segmentation_file method).

The task is a competition. Everyone who submits a solution achieving at least 87% test set IoU gets 4 points; the remaining 5 bonus points are distributed depending on relative ordering of your solutions.

You may want to start with the cags_segmentation.py template, which generates the test set annotation in the required format – each mask should be encoded on a single line as a space separated sequence of integers indicating the length of alternating runs of zeros and ones.

bboxes_utils

 Deadline: Apr 09, 22:00  2 points

This is a preparatory assignment for svhn_competition. The goal is to implement several bounding box manipulation routines in the bboxes_utils.py module. Notably, you need to implement the following methods:

• bboxes_to_rcnn: convert given bounding boxes to a R-CNN-like representation relative to the given anchors;
• bboxes_from_rcnn: convert R-CNN-like representations relative to given anchors back to bounding boxes;
• bboxes_training: given a list of anchors and gold objects, assign gold objects to anchors and generate suitable training data (the exact algorithm is described in the template).

The bboxes_utils.py contains simple unit tests, which are evaluated when executing the module, which you can use to check the validity of your implementation. Note that the template does not contain type annotations because Python typing system is not flexible enough to describe the tensor shape changes.

When submitting to ReCodEx, the method main is executed, returning the implemented bboxes_to_rcnn, bboxes_from_rcnn and bboxes_training methods. These methods are then executed and compared to the reference implementation.

svhn_competition

 Deadline: Apr 09, 22:00  5 points+5 bonus

The goal of this assignment is to implement a system performing object recognition, optionally utilizing the pretrained EfficientNetV2-B0 backbone (or any other model from keras.applications).

The Street View House Numbers (SVHN) dataset annotates for every photo all digits appearing on it, including their bounding boxes. The dataset can be loaded using the svhn_dataset.py module. Similarly to the CAGS dataset, the train/dev/test are PyTorch torch.utils.data.Datasets, and every element is a dictionary with the following keys:

• "image": a square 3-channel image stored using PyTorch tensor of type torch.uint8,
• "classes": a 1D np.ndarray with all digit labels appearing in the image,
• "bboxes": a [num_digits, 4] 2D np.ndarray with bounding boxes of every digit in the image, each represented as [TOP, LEFT, BOTTOM, RIGHT].

Each test set image annotation consists of a sequence of space separated five-tuples label top left bottom right, and the annotation is considered correct, if exactly the gold digits are predicted, each with IoU at least 0.5. The whole test set score is then the prediction accuracy of individual images. You can again evaluate your predictions using the svhn_dataset.py module, either by running with --evaluate=path arguments, or using its evaluate_file method.

The task is a competition. Everyone who submits a solution achieving at least 20% test set accuracy gets 5 points; the remaining 5 bonus points are distributed depending on relative ordering of your solutions. Note that I usually need at least 35% development set accuracy to achieve the required test set performance.

You should start with the svhn_competition.py template, which generates the test set annotation in the required format.

A baseline solution can use RetinaNet-like single stage detector, using only a single level of convolutional features (no FPN) with single-scale and single-aspect anchors. Focal loss is available as keras.losses.BinaryFocalCrossentropy and non-maximum suppression as torchvision.ops.nms or torchvision.ops.batched_nms.

3d_recognition

 Deadline: Apr 16, 22:00  3 points+4 bonus

Your goal in this assignment is to perform 3D object recognition. The input is voxelized representation of an object, stored as a 3D grid of either empty or occupied voxels, and your goal is to classify the object into one of 10 classes. The data is available in two resolutions, either as 20×20×20 data or 32×32×32 data. To load the dataset, use the modelnet.py module.

The official dataset offers only train and test sets, with the test set having a different distributions of labels. Our dataset contains also a development set, which has nearly the same label distribution as the test set.

If you want, it is possible to use any model from keras.applications in this assignment; however, the only way I know how to utilize such a pre-trained model is to render the objects to a set of 2D images and classify them instead.

The task is a competition. Everyone who submits a solution achieving at least 88% test set accuracy gets 3 points; the remaining 4 bonus points are distributed depending on relative ordering of your solutions.

You can start with the 3d_recognition.py template, which among others generates test set annotations in the required format.

sequence_classification

 Deadline: Apr 23, 22:00  2 points  Tests  Examples

The goal of this assignment is to introduce recurrent neural networks, show their convergence speed, and illustrate exploding gradient issue. The network should process sequences of 50 small integers and compute parity for each prefix of the sequence. The inputs are either 0/1, or vectors with one-hot representation of small integer.

Your goal is to modify the sequence_classification.py template and implement the following:

• Use the specified RNN type (SimpleRNN, GRU, and LSTM) and dimensionality.
• Process the sequence using the required RNN.
• Use additional hidden layer on the RNN outputs if requested.
• Implement gradient clipping if requested.

In addition to submitting the task in ReCodEx, please also run the following variations and observe the results in TensorBoard. Concentrate on the way how the RNNs converge, convergence speed, exploding gradient issues and how gradient clipping helps:

• --rnn=SimpleRNN --sequence_dim=1, --rnn=GRU --sequence_dim=1, --rnn=LSTM --sequence_dim=1
• the same as above but with --sequence_dim=3
• the same as above but with --sequence_dim=10
• --rnn=SimpleRNN --hidden_layer=85 --rnn_dim=30 --sequence_dim=30 and the same with --clip_gradient=1
• the same as above but with --rnn=GRU with and without --clip_gradient=1
• the same as above but with --rnn=LSTM with and without --clip_gradient=1

Epoch 1/5 accuracy: 0.4854 - loss: 0.7253 - val_accuracy: 0.5092 - val_loss: 0.6971
Epoch 2/5 accuracy: 0.5101 - loss: 0.6944 - val_accuracy: 0.4990 - val_loss: 0.6914
Epoch 3/5 accuracy: 0.5000 - loss: 0.6904 - val_accuracy: 0.5198 - val_loss: 0.6892
Epoch 4/5 accuracy: 0.5200 - loss: 0.6887 - val_accuracy: 0.5328 - val_loss: 0.6875
Epoch 5/5 accuracy: 0.5326 - loss: 0.6869 - val_accuracy: 0.5362 - val_loss: 0.6857

Epoch 1/5 accuracy: 0.5277 - loss: 0.6925 - val_accuracy: 0.5217 - val_loss: 0.6921
Epoch 2/5 accuracy: 0.5183 - loss: 0.6921 - val_accuracy: 0.5217 - val_loss: 0.6918
Epoch 3/5 accuracy: 0.5185 - loss: 0.6919 - val_accuracy: 0.5217 - val_loss: 0.6914
Epoch 4/5 accuracy: 0.5212 - loss: 0.6914 - val_accuracy: 0.5282 - val_loss: 0.6910
Epoch 5/5 accuracy: 0.5320 - loss: 0.6904 - val_accuracy: 0.5355 - val_loss: 0.6905

Epoch 1/5 accuracy: 0.5359 - loss: 0.6926 - val_accuracy: 0.5361 - val_loss: 0.6925
Epoch 2/5 accuracy: 0.5358 - loss: 0.6925 - val_accuracy: 0.5333 - val_loss: 0.6923
Epoch 3/5 accuracy: 0.5370 - loss: 0.6923 - val_accuracy: 0.5369 - val_loss: 0.6920
Epoch 4/5 accuracy: 0.5342 - loss: 0.6919 - val_accuracy: 0.5366 - val_loss: 0.6917
Epoch 5/5 accuracy: 0.5378 - loss: 0.6915 - val_accuracy: 0.5444 - val_loss: 0.6914

Epoch 1/5 accuracy: 0.5377 - loss: 0.6923 - val_accuracy: 0.5414 - val_loss: 0.6911
Epoch 2/5 accuracy: 0.5465 - loss: 0.6902 - val_accuracy: 0.5577 - val_loss: 0.6878
Epoch 3/5 accuracy: 0.5600 - loss: 0.6862 - val_accuracy: 0.5450 - val_loss: 0.6811
Epoch 4/5 accuracy: 0.5491 - loss: 0.6783 - val_accuracy: 0.5590 - val_loss: 0.6707
Epoch 5/5 accuracy: 0.5539 - loss: 0.6678 - val_accuracy: 0.5433 - val_loss: 0.6591

Epoch 1/5 accuracy: 0.5421 - loss: 0.6923 - val_accuracy: 0.5409 - val_loss: 0.6910
Epoch 2/5 accuracy: 0.5504 - loss: 0.6900 - val_accuracy: 0.5511 - val_loss: 0.6875
Epoch 3/5 accuracy: 0.5566 - loss: 0.6860 - val_accuracy: 0.5494 - val_loss: 0.6816
Epoch 4/5 accuracy: 0.5504 - loss: 0.6788 - val_accuracy: 0.5398 - val_loss: 0.6721
Epoch 5/5 accuracy: 0.5539 - loss: 0.6699 - val_accuracy: 0.5494 - val_loss: 0.6624

Epoch 1/5 accuracy: 0.4984 - loss: 0.7004 - val_accuracy: 0.5223 - val_loss: 0.6884
Epoch 2/5 accuracy: 0.5198 - loss: 0.6862 - val_accuracy: 0.5117 - val_loss: 0.6794
Epoch 3/5 accuracy: 0.5132 - loss: 0.6784 - val_accuracy: 0.5121 - val_loss: 0.6732
Epoch 4/5 accuracy: 0.5160 - loss: 0.6723 - val_accuracy: 0.5191 - val_loss: 0.6683
Epoch 5/5 accuracy: 0.5235 - loss: 0.6680 - val_accuracy: 0.5276 - val_loss: 0.6639

Epoch 1/5 accuracy: 0.5109 - loss: 0.6929 - val_accuracy: 0.5128 - val_loss: 0.6915
Epoch 2/5 accuracy: 0.5174 - loss: 0.6894 - val_accuracy: 0.5155 - val_loss: 0.6785
Epoch 3/5 accuracy: 0.5446 - loss: 0.6630 - val_accuracy: 0.9538 - val_loss: 0.2142
Epoch 4/5 accuracy: 0.9812 - loss: 0.1270 - val_accuracy: 0.9987 - val_loss: 0.0304
Epoch 5/5 accuracy: 0.9985 - loss: 0.0270 - val_accuracy: 0.9995 - val_loss: 0.0135

Epoch 1/5 accuracy: 0.5131 - loss: 0.6930 - val_accuracy: 0.5187 - val_loss: 0.6918
Epoch 2/5 accuracy: 0.5187 - loss: 0.6892 - val_accuracy: 0.5340 - val_loss: 0.6760
Epoch 3/5 accuracy: 0.6401 - loss: 0.5744 - val_accuracy: 1.0000 - val_loss: 0.0845
Epoch 4/5 accuracy: 1.0000 - loss: 0.0585 - val_accuracy: 1.0000 - val_loss: 0.0194
Epoch 5/5 accuracy: 1.0000 - loss: 0.0154 - val_accuracy: 1.0000 - val_loss: 0.0082

Epoch 1/5 accuracy: 0.5151 - loss: 0.6888 - val_accuracy: 0.5323 - val_loss: 0.6571
Epoch 2/5 accuracy: 0.5387 - loss: 0.6497 - val_accuracy: 0.5575 - val_loss: 0.6321
Epoch 3/5 accuracy: 0.5570 - loss: 0.6242 - val_accuracy: 0.6199 - val_loss: 0.5854
Epoch 4/5 accuracy: 0.8367 - loss: 0.2854 - val_accuracy: 0.9897 - val_loss: 0.0503
Epoch 5/5 accuracy: 0.9995 - loss: 0.0058 - val_accuracy: 0.9999 - val_loss: 0.0014

Epoch 1/5 accuracy: 0.5151 - loss: 0.6888 - val_accuracy: 0.5323 - val_loss: 0.6571
Epoch 2/5 accuracy: 0.5387 - loss: 0.6497 - val_accuracy: 0.5582 - val_loss: 0.6321
Epoch 3/5 accuracy: 0.5576 - loss: 0.6237 - val_accuracy: 0.6542 - val_loss: 0.5625
Epoch 4/5 accuracy: 0.9033 - loss: 0.1909 - val_accuracy: 0.9999 - val_loss: 0.0014
Epoch 5/5 accuracy: 0.9997 - loss: 0.0029 - val_accuracy: 1.0000 - val_loss: 4.4711e-04

tagger_we

 Deadline: Apr 23, 22:00  3 points  Tests  Examples

In this assignment you will create a simple part-of-speech tagger. For training and evaluation, we will use Czech dataset containing tokenized sentences, each word annotated by gold lemma and part-of-speech tag. The morpho_dataset.py module (down)loads the dataset and provides mappings between strings and integers.

Your goal is to modify the tagger_we.py template and implement the following:

• Use specified RNN layer type (GRU and LSTM) and dimensionality.
• Create word embeddings for training vocabulary.
• Process the sentences using bidirectional RNN.
• Predict part-of-speech tags. Note that you need to properly handle sentences of different lengths in one batch.

Epoch=1/1 3.1s loss=2.3541 accuracy=0.3138 dev_loss=2.0320 dev_accuracy=0.3611

Epoch=1/1 3.2s loss=2.1970 accuracy=0.4233 dev_loss=1.5569 dev_accuracy=0.5121

Epoch=1/5 21.1s loss=0.9776 accuracy=0.7080 dev_loss=0.3744 dev_accuracy=0.8814
Epoch=2/5 19.2s loss=0.1060 accuracy=0.9736 dev_loss=0.2947 dev_accuracy=0.9013
Epoch=3/5 19.4s loss=0.0291 accuracy=0.9921 dev_loss=0.2794 dev_accuracy=0.9057
Epoch=4/5 19.7s loss=0.0166 accuracy=0.9960 dev_loss=0.2976 dev_accuracy=0.9015
Epoch=5/5 19.7s loss=0.0096 accuracy=0.9978 dev_loss=0.3159 dev_accuracy=0.8957

Epoch=1/5 20.5s loss=0.7698 accuracy=0.7703 dev_loss=0.3432 dev_accuracy=0.8903
Epoch=2/5 18.9s loss=0.0735 accuracy=0.9807 dev_loss=0.2999 dev_accuracy=0.8969
Epoch=3/5 19.0s loss=0.0245 accuracy=0.9923 dev_loss=0.3244 dev_accuracy=0.8965
Epoch=4/5 19.2s loss=0.0153 accuracy=0.9955 dev_loss=0.3302 dev_accuracy=0.8929
Epoch=5/5 19.0s loss=0.0088 accuracy=0.9977 dev_loss=0.3641 dev_accuracy=0.8923

tagger_cle

 Deadline: Apr 23, 22:00  3 points  Tests  Examples

This assignment is a continuation of tagger_we. Using the tagger_cle.py template, implement character-level word embedding computation using a bidirectional character-level GRU.

Once submitted to ReCodEx, you should experiment with the effect of CLEs compared to a plain tagger_we, and the influence of their dimensionality. Note that tagger_cle has by default smaller word embeddings so that the size of word representation (64 + 32 + 32) is the same as in the tagger_we assignment.

Epoch=1/1 4.0s loss=2.2871 accuracy=0.2909 dev_loss=1.8784 dev_accuracy=0.4275

Epoch=1/1 4.4s loss=2.2846 accuracy=0.2875 dev_loss=1.8835 dev_accuracy=0.4289

Epoch=1/5 22.6s loss=1.0757 accuracy=0.6784 dev_loss=0.3678 dev_accuracy=0.8969
Epoch=2/5 21.5s loss=0.1476 accuracy=0.9684 dev_loss=0.1978 dev_accuracy=0.9375
Epoch=3/5 22.1s loss=0.0490 accuracy=0.9881 dev_loss=0.1722 dev_accuracy=0.9488
Epoch=4/5 21.3s loss=0.0303 accuracy=0.9912 dev_loss=0.1651 dev_accuracy=0.9470
Epoch=5/5 21.1s loss=0.0201 accuracy=0.9942 dev_loss=0.1630 dev_accuracy=0.9479

Epoch=1/5 22.2s loss=1.1264 accuracy=0.6594 dev_loss=0.3980 dev_accuracy=0.8977
Epoch=2/5 21.4s loss=0.2340 accuracy=0.9408 dev_loss=0.2175 dev_accuracy=0.9377
Epoch=3/5 24.1s loss=0.1163 accuracy=0.9690 dev_loss=0.1624 dev_accuracy=0.9525
Epoch=4/5 26.6s loss=0.0852 accuracy=0.9745 dev_loss=0.1493 dev_accuracy=0.9560
Epoch=5/5 24.9s loss=0.0718 accuracy=0.9778 dev_loss=0.1450 dev_accuracy=0.9563

tagger_competition

 Deadline: Apr 23, 22:00  4 points+5 bonus

In this assignment, you should extend tagger_cle into a real-world Czech part-of-speech tagger. We will use Czech PDT dataset loadable using the morpho_dataset.py module. Note that the dataset contains more than 1500 unique POS tags and that the POS tags have a fixed structure of 15 positions (so it is possible to generate the POS tag characters independently).

You can use the following additional data in this assignment:

• You can use outputs of a morphological analyzer loadable with morpho_analyzer.py. If a word form in train, dev or test PDT data is known to the analyzer, all its (lemma, POS tag) pairs are returned.
• You can use any unannotated text data (Wikipedia, Czech National Corpus, …), and also any pre-trained word embeddings (assuming they were trained on plain texts).

The task is a competition. Everyone who submits a solution with at least 92.5% label accuracy gets 4 points; the remaining 5 bonus points are distributed depending on relative ordering of your solutions. Lastly, 3 bonus points will be given to anyone surpassing pre-neural-network state-of-the-art of 96.35%.

You can start with the tagger_competition.py template, which among others generates test set annotations in the required format. Note that you can evaluate the predictions as usual using the morpho_dataset.py module, either by running with --task=tagger --evaluate=path arguments, or using its evaluate_file method.

tensorboard_projector

You can try exploring the TensorBoard Projector with pre-trained embeddings for 20k most frequent lemmas in Czech and English – after extracting the archive, start tensorboard --logdir dir_where_the_archive_is_extracted.

In order to use the Projector tab yourself, you can take inspiration from the projector_export.py script, which was used to export the above pre-trained embeddings from the Word2vec format.

tagger_ner

 Deadline: Apr 30, 22:00  2 points  Tests  Examples

This assignment is an extension of tagger_we task. Using the tagger_ner.py template, implement optimal decoding of named entity spans from BIO-encoded tags. In a valid sequence, the I-TYPE tag must follow either B-TYPE or I-TYPE tags.

The evaluation is performed using the provided metric computing F1 score of the span prediction (i.e., a recognized possibly-multiword named entity is true positive if both the entity type and the span exactly match).

In practice, character-level embeddings (and also pre-trained word embeddings) would be used to obtain superior results.

Epoch=1/2 10.3s loss=1.0373 accuracy=0.8046 dev_loss=0.7787 dev_accuracy=0.8225 dev_f1_constrained=0.0067 dev_f1_greedy=0.0057
Epoch=2/2 9.3s loss=0.6481 accuracy=0.8179 dev_loss=0.6709 dev_accuracy=0.8331 dev_f1_constrained=0.0910 dev_f1_greedy=0.0834

Epoch=1/2 10.2s loss=1.8772 accuracy=0.8045 dev_loss=1.7571 dev_accuracy=0.8224 dev_f1_constrained=0.0039 dev_f1_greedy=0.0039
Epoch=2/2 9.3s loss=1.7125 accuracy=0.8181 dev_loss=1.7097 dev_accuracy=0.8308 dev_f1_constrained=0.0546 dev_f1_greedy=0.0413

Epoch=1/5 47.6s loss=0.7167 accuracy=0.8339 dev_loss=0.5699 dev_accuracy=0.8422 dev_f1_constrained=0.1915 dev_f1_greedy=0.1468
Epoch=2/5 48.0s loss=0.3642 accuracy=0.8915 dev_loss=0.4467 dev_accuracy=0.8785 dev_f1_constrained=0.4204 dev_f1_greedy=0.3684
Epoch=3/5 48.3s loss=0.1773 accuracy=0.9476 dev_loss=0.4161 dev_accuracy=0.8847 dev_f1_constrained=0.4826 dev_f1_greedy=0.4307
Epoch=4/5 49.4s loss=0.0852 accuracy=0.9755 dev_loss=0.4318 dev_accuracy=0.8877 dev_f1_constrained=0.4878 dev_f1_greedy=0.4427
Epoch=5/5 49.0s loss=0.0490 accuracy=0.9860 dev_loss=0.4354 dev_accuracy=0.8948 dev_f1_constrained=0.5214 dev_f1_greedy=0.4909

Epoch=1/5 48.6s loss=1.7265 accuracy=0.8357 dev_loss=1.6516 dev_accuracy=0.8541 dev_f1_constrained=0.2550 dev_f1_greedy=0.2130
Epoch=2/5 49.5s loss=1.5508 accuracy=0.9061 dev_loss=1.6000 dev_accuracy=0.8925 dev_f1_constrained=0.4974 dev_f1_greedy=0.4591
Epoch=3/5 50.4s loss=1.4624 accuracy=0.9618 dev_loss=1.5795 dev_accuracy=0.8993 dev_f1_constrained=0.5369 dev_f1_greedy=0.5035
Epoch=4/5 50.4s loss=1.4269 accuracy=0.9806 dev_loss=1.5778 dev_accuracy=0.9003 dev_f1_constrained=0.5475 dev_f1_greedy=0.5202
Epoch=5/5 50.0s loss=1.4109 accuracy=0.9890 dev_loss=1.5757 dev_accuracy=0.9026 dev_f1_constrained=0.5546 dev_f1_greedy=0.5368

ctc_loss

 Deadline: Apr 30 May 7, 22:00  2 points  Tests  Examples

This assignment is an extension of tagger_we task. Using the ctc_loss.py template, manually implement the CTC loss computation and also greedy CTC decoding. You can use torch.nn.CTCLoss during development as a reference, but it is not available during ReCodEx evaluation.

Epoch=1/1 0.4s loss=27.2595 edit_distance=2.3694 dev_loss=17.1484 dev_edit_distance=0.6000

Epoch=1/1 8.0s loss=6.5798 edit_distance=0.6902 dev_loss=2.3089 dev_edit_distance=0.5864

Epoch=1/5 67.0s loss=2.4850 edit_distance=0.6043 dev_loss=1.6261 dev_edit_distance=0.5635
Epoch=2/5 67.6s loss=1.2934 edit_distance=0.4832 dev_loss=1.3653 dev_edit_distance=0.4375
Epoch=3/5 68.0s loss=0.7368 edit_distance=0.3033 dev_loss=1.2962 dev_edit_distance=0.3980
Epoch=4/5 68.4s loss=0.4250 edit_distance=0.1754 dev_loss=1.5679 dev_edit_distance=0.3999
Epoch=5/5 68.4s loss=0.2656 edit_distance=0.1082 dev_loss=1.7975 dev_edit_distance=0.4054

speech_recognition

 Deadline: Apr 30, 22:00  5 points+5 bonus

This assignment is a competition task in speech recognition area. Specifically, your goal is to predict a sequence of letters given a spoken utterance. We will be using Czech recordings from the Common Voice, with input sound waves passed through the usual preprocessing – computing Mel-frequency cepstral coefficients (MFCCs). You can repeat this preprocessing on a given audio using the load_audio and mfcc_extract methods from the common_voice_cs.py module. This module can also load the dataset, downloading it when necessary (note that it has 200MB, so it might take a while). Furthermore, you can listen to the development portion of the dataset. Lastly, the whole dataset is available for download in MP3 format (but you are not expected to download that, only if you would like to perform some custom preprocessing).

Additional following data can be utilized in this assignment:

• You can use any unannotated text data (Wikipedia, Czech National Corpus, …), and also any pre-trained word embeddings or language models (assuming they were trained on plain texts).
• You can use any unannotated speech data.

The task is a competition. The evaluation is performed by computing the edit distance to the gold letter sequence, normalized by its length (a corresponding metric EditDistanceMetric is provided by the common_voice_cs.py). Everyone who submits a solution with at most 50% test set edit distance gets 5 points; the remaining 5 bonus points are distributed depending on relative ordering of your solutions. Note that you can evaluate the predictions as usual using the common_voice_cs.py module, either by running with --evaluate=path arguments, or using its evaluate_file method.

Start with the speech_recognition.py template containing a structure suitable for computing the CTC loss and perform CTC decoding. You can use torch.nn.CTCLoss to compute the loss and you can use torchaudio.models.decoder.CTCDecoder/torchaudio.models.decoder.CUCTCDecoder to perform beam-search decoding.

lemmatizer_noattn

 Deadline: May 7, 22:00  3 points  Tests  Examples

The goal of this assignment is to create a simple lemmatizer. For training and evaluation, we use the same dataset as in tagger_we loadable by the updated morpho_dataset.py module.

Your goal is to modify the lemmatizer_noattn.py template and implement the following:

• Embed characters of source forms and run a bidirectional GRU encoder.
• Embed characters of target lemmas.
• Implement a training time decoder which uses gold target characters as inputs.
• Implement an inference time decoder which uses previous predictions as inputs.
• The initial state of both decoders is the output state of the corresponding GRU encoded form.
• If requested, tie the embeddings in the decoder.

Epoch=1/1 4.7s loss=3.0619 accuracy=0.0114 dev_accuracy=0.1207

Epoch=1/1 5.0s loss=2.9198 accuracy=0.0515 dev_accuracy=0.1491

Epoch=1/3 21.4s loss=2.2488 accuracy=0.1792 dev_accuracy=0.3247
Epoch=2/3 21.2s loss=0.9973 accuracy=0.4235 dev_accuracy=0.4670
Epoch=3/3 20.8s loss=0.5733 accuracy=0.5820 dev_accuracy=0.5983

Epoch=1/3 21.1s loss=1.9168 accuracy=0.2528 dev_accuracy=0.3765
Epoch=2/3 20.6s loss=0.8213 accuracy=0.4883 dev_accuracy=0.5110
Epoch=3/3 21.0s loss=0.5173 accuracy=0.6207 dev_accuracy=0.6094

lemmatizer_attn

 Deadline: May 7, 22:00  3 points  Tests  Examples

This task is a continuation of the lemmatizer_noattn assignment. Using the lemmatizer_attn.py template, implement the following features in addition to lemmatizer_noattn:

• The bidirectional GRU encoder returns outputs for all input characters, not just the last.
• Implement attention in the decoders. Notably, project the encoder outputs and current state into same-dimensionality vectors, apply non-linearity, and generate weights for every encoder output. Finally sum the encoder outputs using these weights and concatenate the computed attention to the decoder inputs.

Once submitted to ReCodEx, you should experiment with the effect of using the attention, and the influence of RNN dimensionality on network performance.

Epoch=1/1 7.3s loss=3.0203 accuracy=0.0016 dev_accuracy=0.0338

Epoch=1/1 6.9s loss=2.8839 accuracy=0.0362 dev_accuracy=0.1570

Epoch=1/3 43.9s loss=1.9892 accuracy=0.2314 dev_accuracy=0.5919
Epoch=2/3 44.3s loss=0.3911 accuracy=0.7048 dev_accuracy=0.7501
Epoch=3/3 46.0s loss=0.2234 accuracy=0.7894 dev_accuracy=0.7836

Epoch=1/3 44.3s loss=1.5970 accuracy=0.3400 dev_accuracy=0.6315
Epoch=2/3 45.0s loss=0.3174 accuracy=0.7320 dev_accuracy=0.7521
Epoch=3/3 45.2s loss=0.1880 accuracy=0.8076 dev_accuracy=0.8040

lemmatizer_competition

 Deadline: May 7, 22:00  4 points+5 bonus

In this assignment, you should extend lemmatizer_noattn or lemmatizer_attn into a real-world Czech lemmatizer. As in tagger_competition, we will use Czech PDT dataset loadable using the morpho_dataset.py module.

You can also use the same additional data as in the tagger_competition assignment.

The task is a competition. Everyone who submits a solution with at least 96.5% label accuracy gets 4 points; the remaining 5 bonus points are distributed depending on relative ordering of your solutions. Lastly, 3 bonus points will be given to anyone surpassing pre-neural-network state-of-the-art of 98.76%.

You can start with the lemmatizer_competition.py template, which among others generates test set annotations in the required format. Note that you can evaluate the predictions as usual using the morpho_dataset.py module, either by running with --task=lemmatizer --evaluate=path arguments, or using its evaluate_file method.

In the competitions, your goal is to train a model, and then predict target values on the given unannotated test set.

Submitting to ReCodEx

When submitting a competition solution to ReCodEx, you can include any number of files of any kind, and either submit them individually or compess them in a .zip file. However, there should be exactly one text file with the test set annotation (.txt) and at least one Python source (.py/ipynb) containing the model training and prediction. The Python sources are not executed, but must be included for inspection.

Competition Evaluation

• For every submission, ReCodEx checks the above conditions (exactly one .txt, at least one .py/ipynb) and whether the given annotations can be evaluated without error. If not, it will report the corresponding error in the logs.

• Before the first deadline, ReCodEx prints the exact achieved performance, but only if it is worse than the baseline.

  If you surpass the baseline, the assignment is marked as solved in ReCodEx and you immediately get regular points for the assignment. However, ReCodEx does not print the reached performance.

• After the first deadline, the latest submission of every user surpassing the required baseline participates in a competition. Additional bonus points are then awarded according to the ordering of the performance of the participating submissions.

• After the competition results announcement, ReCodEx starts to show the exact performance for all the already submitted solutions and also for the solutions submitted later.

What Is Allowed

• You can use only the given annotated data for training and evaluation.
• You can use the given annotated training data in any way.
• You can use the given annotated development data for evaluation or hyperparameter tuning, but not for the training itself.
• Additionally, you can use any unannotated or manually created data for training and evaluation.
• The test set annotations must be the result of your system (so you cannot manually correct them; but your system can contain other parts than just trained models, like hand-written rules).
• Do not use test set annotations in any way, if you somehow get access to them.
• Unless stated otherwise, you can use any architecture to solve the competition task at hand, but the implementation must be created by you and you must understand it fully. You can of course take inspiration from any paper or existing implementation, but please reference it in that case.
  • You can of course use anything from the Keras/PyTorch packages (but not models from Keras CV, torchvision, …).
  • You can use any data augmentation (even implementations not written by you).
  • You can use any optimizer and any hyperparameter optimization method (even implementations not written by you).
• If you utilize an already trained model, it must be trained only on the allowed training data, unless stated otherwise.

Install

• Installing to central user packages repository

  You can install all required packages to central user packages repository using python3 -m pip install --user --no-cache-dir keras~=3.0.5 --extra-index-url=https://download.pytorch.org/whl/cu118 torch~=2.2.0 torchaudio~=2.2.0 torchvision~=0.17.0 torchmetrics~=1.3.1 flashlight-text~=0.0.3 tensorboard~=2.16.2 transformers~=4.37.2 gymnasium~=1.0.0a1 pygame~=2.5.2.

  The above command installs CUDA 11.8 PyTorch build, but you can change cu118 to:

  • cpu to get CPU-only (smaller) version,
  • cu121 to get CUDA 12.1 build,
  • rocm5.7 to get AMD ROCm 5.7 build.

• Installing to a virtual environment

  Python supports virtual environments, which are directories containing independent sets of installed packages. You can create a virtual environment by running python3 -m venv VENV_DIR followed by VENV_DIR/bin/pip install --no-cache-dir keras~=3.0.5 --extra-index-url=https://download.pytorch.org/whl/cu118 torch~=2.2.0 torchaudio~=2.2.0 torchvision~=0.17.0 torchmetrics~=1.3.1 flashlight-text~=0.0.3 tensorboard~=2.16.2 transformers~=4.37.2 gymnasium~=1.0.0a1 pygame~=2.5.2. (or VENV_DIR/Scripts/pip on Windows).

  Again, apart from the CUDA 11.8 build, you can change cu118 to:

  • cpu to get CPU-only (smaller) version,
  • cu121 to get CUDA 12.1 build,
  • rocm5.7 to get AMD ROCm 5.7 build.

• Windows installation

  • On Windows, it can happen that python3 is not in PATH, while py command is – in that case you can use py -m venv VENV_DIR, which uses the newest Python available, or for example py -3.11 -m venv VENV_DIR, which uses Python version 3.11.

  • If you encounter a problem creating the logs in the args.logdir directory, a possible cause is that the path is longer than 260 characters, which is the default maximum length of a complete path on Windows. However, you can increase this limit on Windows 10, version 1607 or later, by following the instructions.

• GPU support on Linux and Windows

  PyTorch supports NVIDIA GPU or AMD GPU out of the box, you just need to select appropriate --extra-index-url when installing the packages.

  If you encounter problems loading CUDA or cuDNN libraries, make sure your LD_LIBRARY_PATH does not contain paths to older CUDA/cuDNN libraries.

• GPU support on macOS

  The support for Apple Silicon GPUs in PyTorch+Keras is currently not great. Apple is working on mlx backend for Keras, which might improve the situation in the future.

  You can instead use the TensorFlow backend – just install tensorflow~=2.16.1 and tensorflow-metal packages from PyPI, and run export KERAS_BACKEND=tensorflow in every terminal before running assignment scripts.

• How to install TensorFlow

  If you would like to install also TensorFlow, run pip install tensorflow-cpu for CPU-only support, and pip install tensorflow[and-cuda] for Linux/WSL2 GPU support. However, the paths to the CUDA libraries seem not to be detected correctly, so I had to run

  export LD_LIBRARY_PATH=$(echo VENV_DIR/lib/python*/site-packages/nvidia/*/lib | tr " " ":")
  

  in the terminal for the GPU support to work.

MetaCentrum

• How to apply for MetaCentrum account?

  After reading the Terms and conditions, you can apply for an account here.

  After your account is created, please make sure that the directories containing your solutions are always private.

• How to activate Python 3.10 on MetaCentrum?

  On Metacentrum, currently the newest available Python is 3.10, which you need to activate in every session by running the following command:

  module add python/python-3.10.4-intel-19.0.4-sc7snnf
  

• How to install the required virtual environment on MetaCentrum?

  To create a virtual environment, you first need to decide where it will reside. Either you can find a permanent storage, where you have large-enough quota, or you can use scratch storage for a submitted job.

  TL;DR:

  • Run an interactive CPU job, asking for 16GB scratch space:

    qsub -l select=1:ncpus=1:mem=8gb:scratch_local=16gb -I
    

  • In the job, use the allocated scratch space as the temporary directory:

    export TMPDIR=$SCRATCHDIR
    

  • You should clear the scratch space before you exit using the clean_scratch command. You can instruct the shell to call it automatically by running:

    trap 'clean_scratch' TERM EXIT
    

  • Finally, create the virtual environment and install PyTorch in it:

    module add python/python-3.10.4-intel-19.0.4-sc7snnf
    python3 -m venv CHOSEN_VENV_DIR
    CHOSEN_VENV_DIR/bin/pip install --no-cache-dir --upgrade pip setuptools
    CHOSEN_VENV_DIR/bin/pip install --no-cache-dir keras~=3.0.5 --extra-index-url=https://download.pytorch.org/whl/cu118 torch~=2.2.0 torchaudio~=2.2.0 torchvision~=0.17.0 torchmetrics~=1.3.1 flashlight-text~=0.0.3 tensorboard~=2.16.2 transformers~=4.37.2 gymnasium~=1.0.0a1 pygame~=2.5.2
    

• How to run a GPU computation on MetaCentrum?

  First, read the official MetaCentrum documentation: Basic terms, Run simple job, GPU computing, GPU clusters.

  TL;DR: To run an interactive GPU job with 1 CPU, 1 GPU, 8GB RAM, and 16GB scatch space, run:

  qsub -q gpu -l select=1:ncpus=1:ngpus=1:mem=8gb:scratch_local=16gb -I
  

  To run a script in a non-interactive way, replace the -I option with the script to be executed.

  If you want to run a CPU-only computation, remove the -q gpu and ngpus=1: from the above commands.

AIC

• How to install required packages on AIC?

  The Python 3.11.7 is available /opt/python/3.11.7/bin/python3, so you should start by creating a virtual environment using

  /opt/python/3.11.7/bin/python3 -m venv VENV_DIR
  

  and then install the required packages in it using

  VENV_DIR/bin/pip install --no-cache-dir keras~=3.0.5 --extra-index-url=https://download.pytorch.org/whl/cu118 torch~=2.2.0 torchaudio~=2.2.0 torchvision~=0.17.0 torchmetrics~=1.3.1 flashlight-text~=0.0.3 tensorboard~=2.16.2 transformers~=4.37.2 gymnasium~=1.0.0a1 pygame~=2.5.2
  

• How to run a GPU computation on AIC?

  First, read the official AIC documentation: Submitting CPU Jobs, Submitting GPU Jobs.

  TL;DR: To run an interactive GPU job with 1 CPU, 1 GPU, and 16GB RAM, run:

  srun -p gpu -c1 -G1 --mem=16G --pty bash
  

  To run a shell script requiring a GPU in a non-interactive way, use

  sbatch -p gpu -c1 -G1 --mem=16G SCRIPT_PATH
  

  If you want to run a CPU-only computation, remove the -p gpu and -G1 from the above commands.

Git

• Is it possible to keep the solutions in a Git repository?

  Definitely. Keeping the solutions in a branch of your repository, where you merge them with the course repository, is probably a good idea. However, please keep the cloned repository with your solutions private.

• On GitHub, do not create a public fork with your solutions

  If you keep your solutions in a GitHub repository, please do not create a clone of the repository by using the Fork button – this way, the cloned repository would be public.

  Of course, if you just want to create a pull request, GitHub requires a public fork and that is fine – just do not store your solutions in it.

• How to clone the course repository?

  To clone the course repository, run

  git clone https://github.com/ufal/npfl138
  

  This creates the repository in the npfl138 subdirectory; if you want a different name, add it as a last parameter.

  To update the repository, run git pull inside the repository directory.

• How to keep the course repository as a branch in your repository?

  If you want to store the course repository just in a local branch of your existing repository, you can run the following command while in it:

  git remote add upstream https://github.com/ufal/npfl138
  git fetch upstream
  git checkout -t upstream/master
  

  This creates a branch master; if you want a different name, add -b BRANCH_NAME to the last command.

  In both cases, you can update your checkout by running git pull while in it.

• How to merge the course repository with your modifications?

  If you want to store your solutions in a branch merged with the course repository, you should start by

  git remote add upstream https://github.com/ufal/npfl138
  git pull upstream master
  

  which creates a branch master; if you want a different name, change the last argument to master:BRANCH_NAME.

  You can then commit to this branch and push it to your repository.

  To merge the current course repository with your branch, run

  git merge upstream master
  

  while in your branch. Of course, it might be necessary to resolve conflicts if both you and I modified the same place in the templates.

ReCodEx

• What files can be submitted to ReCodEx?

  You can submit multiple files of any type to ReCodEx. There is a limit of 20 files per submission, with a total size of 20MB.

• What file does ReCodEx execute and what arguments does it use?

  Exactly one file with py suffix must contain a line starting with def main(. Such a file is imported by ReCodEx and the main method is executed (during the import, __name__ == "__recodex__").

  The file must also export an argument parser called parser. ReCodEx uses its arguments and default values, but it overwrites some of the arguments depending on the test being executed – the template should always indicate which arguments are set by ReCodEx and which are left intact.

• What are the time and memory limits?

  The memory limit during evaluation is 1.5GB. The time limit varies, but it should be at least 10 seconds and at least twice the running time of my solution.

Finetuning

• How to make a part of the network frozen, so that its weights are not updated?

  Each keras.layers.Layer/keras.Model has a mutable trainable property indicating whether its variables should be updated – however, after changing it, you need to call .compile again (or otherwise make sure the list of trainable variables for the optimizer is updated).

  Note that once trainable == False, the insides of a layer are no longer considered, even if some its sub-layers have trainable == True. Therefore, if you want to freeze only some sub-layers of a layer you use in your model, the layer itself must have trainable == True.

• How to choose whether dropout/batch normalization is executed in training or inference regime?

  When calling a keras.layers.Layer/keras.Model, a named option training can be specified, indicating whether training or inference regime should be used. For a model, this option is automatically passed to its layers which require it, and Keras automatically passes it during model.{fit,evaluate,predict}.

  However, you can manually pass for example training=False to a layer when using Functional API, meaning that layer is executed in the inference regime even when the whole model is training.

• How does trainable and training interact?

  The only layer, which is influenced by both these options, is batch normalization, for which:

  • if trainable == False, the layer is always executed in inference regime;
  • if trainable == True, the training/inference regime is chosen according to the training option.

TensorBoard

• Cannot start TensorBoard after installation

  If tensorboard executable cannot be found, make sure the directory with pip installed packages is in your PATH (that directory is either in your virtual environment if you use a virtual environment, or it should be ~/.local/bin on Linux and %UserProfile%\AppData\Roaming\Python\Python311 and %UserProfile%\AppData\Roaming\Python\Python311\Scripts on Windows).

• What can be logged in TensorBoard? See the documentation of the SummaryWriter. Common possibilities are:

  • scalar values:

    summary_writer.add_scalar(name like "train/loss", value, step)
    

  • tensor values displayed as histograms or distributions:

    summary_writer.add_histogram(name like "train/output_layer", tensor, step)
    

  • images as tensors with shape [num_images, h, w, channels], where channels can be 1 (grayscale), 2 (grayscale + alpha), 3 (RGB), 4 (RGBA):

    summary_writer.add_images(name like "train/samples", images, step, dataformats="NHWC")
    

    Other dataformats are "HWC" (shape [h, w, channels]), "HW", "NCHW", "CHW".
  • possibly large amount of text (e.g., all hyperparameter values, sample translations in MT, …) in Markdown format:

    summary_writer.add_text(name like "hyperparameters", markdown, step)
    

  • audio as tensors with shape [1, samples] and values in $[-1,1]$ range:

    summary_writer.add_audio(name like "train/samples", clip, step, [sample_rate])
    

Requirements

To pass the practicals, you need to obtain at least 80 points, excluding the bonus points. Note that all surplus points (both bonus and non-bonus) will be transfered to the exam. In total, assignments for at least 120 points (not including the bonus points) will be available, and if you solve all the assignments (any non-zero amount of points counts as solved), you automatically pass the exam with grade 1.

To pass the exam, you need to obtain at least 60, 75, or 90 points out of 100-point exam to receive a grade 3, 2, or 1, respectively. The exam consists of 100-point-worth questions from the list below (the questions are randomly generated, but in such a way that there is at least one question from every but the first lecture). In addition, you can get surplus points from the practicals and at most 10 points for community work (i.e., fixing slides or reporting issues) – but only the points you already have at the time of the exam count. You can take the exam without passing the practicals first.

Exam Questions

Lecture 1 Questions

• Considering a neural network with $D$ input neurons, a single hidden layer with $H$ neurons, $K$ output neurons, hidden activation $f$ and output activation $a$, list its parameters (including their size) and write down how the output is computed. [5]

• List the definitions of frequently used MLP output layer activations (the ones producing parameters of a Bernoulli distribution and a categorical distribution). Then write down three commonly used hidden layer activations (sigmoid, tanh, ReLU). [5]

• Formulate the Universal approximation theorem. [5]

Lecture 2 Questions

• Describe maximum likelihood estimation, as minimizing NLL, cross-entropy and KL divergence. [10]

• Define mean squared error and show how it can be derived using MLE. [5]

• Describe gradient descent and compare it to stochastic (i.e., online) gradient descent and minibatch stochastic gradient descent. [5]

• Formulate conditions on the sequence of learning rates used in SGD to converge to optimum almost surely. [5]

• Write down the backpropagation algorithm. [5]

• Write down the mini-batch SGD algorithm with momentum. [5]

• Write down the AdaGrad algorithm and show that it tends to internally decay learning rate by a factor of $1/\sqrt{t}$ in step $t$. Then write down the RMSProp algorithm and explain how it solves the problem with the involuntary learning rate decay. [10]

• Write down the Adam algorithm. Then show why the bias-correction terms $(1-\beta^t)$ make the estimation of the first and second moment unbiased. [10]

Lecture 3 Questions

• Considering a neural network with $D$ input neurons, a single ReLU hidden layer with $H$ units and softmax output layer with $K$ units, write down the explicit formulas of the gradient of all the MLP parameters (two weight matrices and two bias vectors), assuming input $\boldsymbol x$, target $g$ and negative log likelihood loss. [10]

• Assume a network with MSE loss generated a single output $o \in \mathbb{R}$, and the target output is $g$. What is the value of the loss function itself, and what is the explicit formula of the gradient of the loss function with respect to $o$? [5]

• Assume a binary-classification network with cross-entropy loss generated a single output $z \in \mathbb{R}$, which is passed through the sigmoid output activation function, producing $o = \sigma(z)$. If the target output is $g$, what is the value of the loss function itself, and what is the explicit formula of the gradient of the loss function with respect to $z$? [5]

• Assume a $K$-class-classification network with cross-entropy loss generated a $K$-element output $\boldsymbol z \in \mathbb{R}^K$, which is passed through the softmax output activation function, producing $\boldsymbol o=\operatorname{softmax}(\boldsymbol z)$. If the target distribution is $\boldsymbol g$, what is the value of the loss function itself, and what is the explicit formula of the gradient of the loss function with respect to $\boldsymbol z$? [5]

• Define $L_2$ regularization and describe its effect both on the value of the loss function and on the value of the loss function gradient. [5]

• Describe the dropout method and write down exactly how it is used during training and during inference. [5]

• Describe how label smoothing works for cross-entropy loss, both for sigmoid and softmax activations. [5]

• How are weights and biases initialized using the default Glorot initialization? [5]

Lecture 4 Questions

• Write down the equation of how convolution of a given image is computed. Assume the input is an image $I$ of size $H \times W$ with $C$ channels, the kernel $K$ has size $N \times M$, the stride is $T \times S$, the operation performed is in fact cross-correlation (as usual in convolutional neural networks) and that $O$ output channels are computed. [5]

• Explain both SAME and VALID padding schemes and write down the output size of a convolutional operation with an $N \times M$ kernel on image of size $H \times W$ for both these padding schemes (stride is 1). [5]

• Describe batch normalization including all its parameters, and write down an algorithm how it is used during training and the algorithm how it is used during inference. Be sure to explicitly write over what is being normalized in case of fully connected layers and in case of convolutional layers. [10]

• Describe overall architecture of VGG-19 (you do not need to remember the exact number of layers/filters, but you should describe which layers are used). [5]

Lecture 5 Questions

• Describe overall architecture of ResNet. You do not need to remember the exact number of layers/filters, but you should draw a bottleneck block (including the applications of BatchNorms and ReLUs) and state how residual connections work when the number of channels increases. [10]

• Draw the original ResNet block (including the exact positions of BatchNorms and ReLUs) and also the improved variant with full pre-activation. [5]

• Compare the bottleneck block of ResNet and ResNeXt architectures (draw the latter using convolutions only, i.e., do not use grouped convolutions). [5]

• Describe the CNN regularization method of networks with stochastic depth. [5]

• Compare Cutout and DropBlock. [5]

• Describe in detail how is CutMix performed. [5]

• Describe Squeeze and Excitation applied to a ResNet block. [5]

• Draw the Mobile inverted bottleneck block (including explanation of separable convolutions, the expansion factor, exact positions of BatchNorms and ReLUs, but without describing Squeeze and excitation blocks). [5]

• Assume an input image $I$ of size $H \times W$ with $C$ channels, and a convolutional kernel $K$ with size $N \times M$, stride $S$ and $O$ output channels. Write down (or derive) the equation of transposed convolution (or equivalently backpropagation through a convolution to its inputs). [5]

Lecture 6 Questions

• Describe the differences among semantic segmentation, image classification, object detection, and instance segmentation, and write down which metrics are used for these tasks. [5]

• Write down how is $\mathit{AP}_{50}$ computed. [5]

• Considering a Fast-RCNN architecture, draw overall network architecture, explain what a RoI-pooling layer is, show how the network parametrizes bounding boxes and write down the loss. Finally, describe non-maximum suppression and how the Fast-RCNN prediction is performed. [10]

• Considering a Faster-RCNN architecture, describe the region proposal network (what are anchors, architecture including both heads, how are the coordinates of proposals parametrized, what does the loss look like). [10]

• Considering Mask-RCNN architecture, describe the additions to a Faster-RCNN architecture (the RoI-Align layer, the new mask-producing head). [5]

• Write down the focal loss with class weighting, including the commonly used hyperparameter values. [5]

• Draw the overall architecture of a RetinaNet architecture (the computation of $C_1, \ldots, C_7$, the FPN architecture computing $P_1, \ldots, P_7$ including the block combining feature maps of different resolutions; the classification and bounding box generation heads, including their output size). Write down the losses for both heads. [10]

• Describe GroupNorm, and compare it to BatchNorm and LayerNorm. [5]

Lecture 8 Questions

• Write down how the Long Short-Term Memory (LSTM) cell operates, including the explicit formulas. Also mention the forget gate bias. [10]

• Write down how the Gated Recurrent Unit (GRU) operates, including the explicit formulas. [10]

• Describe Highway network computation. [5]

• Why the usual dropout cannot be used on recurrent state? Describe how the problem can be alleviated with variational dropout. [5]

• Describe layer normalization including all its parameters, and write down how it is computed (be sure to explicitly state over what is being normalized in case of fully connected layers and convolutional layers). [5]

• Draw a tagger architecture utilizing word embeddings, recurrent character-level word embeddings (including how are these computed from individual characters), and two sentence-level bidirectional RNNs (explaining the bidirectionality) with a residual connection. Where would you put the dropout layers? [10]

Lecture 9 Questions

• In the context of named entity recognition, describe what the BIO encoding is and why it is used. [5]

• Write down the dynamic programming algorithm for decoding a BIO-tag sequence, including its asymptotic complexity. [10]

• In the context of CTC loss, describe regular and extended labelings and write down the algorithm for computing the log probability of a gold label sequence $\boldsymbol y$. [10]

• Describe how CTC predictions are performed using a beam-search. [5]

• Draw the CBOW architecture from word2vec, including the sizes of the inputs and the sizes of the outputs and used non-linearities. Also make sure to indicate where the embeddings are being trained. [5]

• Draw the SkipGram architecture from word2vec, including the sizes of the inputs and the sizes of the outputs and used non-linearities. Also make sure to indicate where the embeddings are being trained. [5]

• Describe the hierarchical softmax used in word2vec. [5]

• Describe the negative sampling proposed in word2vec, including the choice of distribution of negative samples. [5]

• Explain how are ELMo embeddings trained and how are they used in downstream applications. [5]

Lecture 10 Questions

• Considering machine translation, draw a recurrent sequence-to-sequence architecture without attention, both during training and during inference (include embedding layers, recurrent cells, classification layers, argmax/softmax). [5]

• Considering machine translation, draw a recurrent sequence-to-sequence architecture with attention, used during training (include embedding layers, recurrent cells, attention, classification layers). Then write down how exactly is the attention computed. [10]

• Explain how is word embeddings tying used in a sequence-to-sequence architecture, including the necessary scaling. [5]

• Write down why are subword units used in text processing, and describe the BPE algorithm for constructing a subword dictionary from a large corpus. [5]

• Write down why are subword units used in text processing, and describe the WordPieces algorithm for constructing a subword dictionary from a large corpus. [5]

• Pinpoint the differences between the BPE and WordPieces algorithms, both during dictionary construction and during inference. [5]

• Describe the Transformer encoder architecture, including the description of self-attention (but you do not need to describe multi-head attention), FFN and positions of LNs and dropouts. [10]

• Write down the formula of Transformer self-attention, and then describe multi-head self-attention in detail. [10]

• Describe the Transformer decoder architecture, including the description of self-attention and masked self-attention (but you do not need to describe multi-head attention), FFN and positions of LNs and dropouts. Also discuss the difference between training and prediction regimes. [10]

• Why are positional embeddings needed in Transformer architecture? Write down the sinusoidal positional embeddings used in the Transformer. [5]

• Compare RNN to Transformer – what are the strengths and weaknesses of these architectures? [5]

Related Courses

Machine Learning for Greenhorns

Introductory course to machine learning, focusing both on theoretical foundations as well as on practical applications in Python.

Deep Reinforcement Learning

Course introducing reinforcement learning, from basic tabular methods to involvement of deep neural networks, focusing both on theory as well as on practical aspects.

Archive

Summer 2022/23

Summer 2021/22

Summer 2020/21

Summer 2019/20

Summer 2018/19

Summer 2017/18

Winter 2016/17