notebooks

Dataflowr: Deep Learning DIY

Code and notebooks for the deep learning course dataflowr. Here is the schedule followed at école polytechnique in 2023:

:sunflower:Session:one: Finetuning VGG

• Module 1 - Introduction & General Overview Slides + notebook Dogs and Cats with VGG + Practicals (more dogs and cats)

Things to remember

> - you do not need to understand everything to run a deep learning model! But the main goal of this course will be to come back to each step done today and understand them... > - to use the dataloader from Pytorch, you need to follow the API (i.e. for classification store your dataset in folders) > - using a pretrained model and modifying it to adapt it to a similar task is easy. > - if you do not understand why we take this loss, that's fine, we'll cover that in Module 3. > - even with a GPU, avoid unnecessary computations!

:sunflower:Session:two: PyTorch tensors and Autodiff

• Module 2a - PyTorch tensors
• Module 2b - Automatic differentiation + Practicals
• MLP from scratch start of HW1
• another look at autodiff with dual numbers and Julia

Things to remember

>- Pytorch tensors = Numpy on GPU + gradients! >- in deep learning, [broadcasting](https://numpy.org/doc/stable/user/basics.broadcasting.html) is used everywhere. The rules are the same as for Numpy. >- Automatic differentiation is not only the chain rule! Backpropagation algorithm (or dual numbers) is a clever algorithm to implement automatic differentiation...

:sunflower:Session:three:

• Module 3 - Loss function for classification
• Module 4 - Optimization for deep learning
• Module 5 - Stacking layers and overfitting a MLP on CIFAR10: Stacking_layers_MLP_CIFAR10.ipynb
• Module 6: Convolutional neural network
• how to regularize with dropout and uncertainty estimation with MC Dropout: Module 15 - Dropout

Things to remember

>- Loss vs Accuracy. Know your loss for a classification task! >- know your optimizer (Module 4) >- know how to build a neural net with torch.nn.module (Module 5) >- know how to use convolution and pooling layers (kernel, stride, padding) >- know how to use dropout

:sunflower:Session:four:

• Module 7 - Dataloading
• Module 8a - Embedding layers
• Module 8b - Collaborative filtering and build your own recommender system: 08_collaborative_filtering_empty.ipynb (on a larger dataset 08_collaborative_filtering_1M.ipynb)
• Module 8c - Word2vec and build your own word embedding 08_Word2vec_pytorch_empty.ipynb
• Module 16 - Batchnorm and check your understanding with 16_simple_batchnorm_eval.ipynb and more 16_batchnorm_simple.ipynb
• Module 17 - Resnets and transform your classifier into an out-of-distribution detector with ODIN_mobilenet_empty.ipynb
• start of Homework 2: Class Activation Map and adversarial examples

Things to remember

> - know how to use dataloader > - to deal with categorical variables in deep learning, use embeddings > - in the case of word embedding, starting in an unsupervised setting, we built a supervised task (i.e. predicting central / context words in a window) and learned the representation thanks to negative sampling > - know your batchnorm > - architectures with skip connections allows deeper models

:sunflower:Session:five:

• Module 9a: Autoencoders and code your noisy autoencoder 09_AE_NoisyAE.ipynb
• Module 10: Generative Adversarial Networks and code your GAN, Conditional GAN and InfoGAN 10_GAN_double_moon.ipynb
• Module 13: Siamese Networks and Representation Learning
• start of Homework 3: VAE for MNIST clustering and generation

:sunflower:Session:six:

• Module 11a - Recurrent Neural Networks theory
• Module 11b - Recurrent Neural Networks practice and predict engine failure with 11_predictions_RNN_empty.ipynb
• Module 11c - Batches with sequences in Pytorch

:sunflower:Session:seven:

• Module 12 - Attention and Transformers
• Correcting the PyTorch tutorial on attention in seq2seq: 12_seq2seq_attention.ipynb
• Build your own microGPT: GPT_hist.ipynb

  :sunflower:Session:eight:

• Module 9b - UNets
• Module 9c - Flows
• Build your own Real NVP: Normalizing_flows_empty.ipynb

  :sunflower:Session:nine:

• Module 18a - Denoising Diffusion Probabilistic Models
• Train your own DDPM on MNIST: ddpm_nano_empty.ipynb
• Finetuning on CIFAR10: ddpm_micro_sol.ipynb

For more updates:

:sunflower: All notebooks

• Module 1: Introduction & General Overview
  • Intro: finetuning VGG for dogs vs cats 01_intro.ipynb
  • Practical: Using CNN for more dogs and cats 01_practical_empty.ipynb and its solution 01_practical_sol.ipynb
• Module 2: Pytorch tensors and automatic differentiation
  • Basics on PyTorch tensors and automatic differentiation 02a_basics.ipynb
  • Linear regression from numpy to pytorch 02b_linear_reg.ipynb
  • Practical: implementing backprop from scratch 02_backprop.ipynb and its solution 02_backprop_sol.ipynb
  • Bonus: intro to JAX: autodiff the functional way autodiff_functional_empty.ipynb and its solution autodiff_functional_sol.ipynb
  • Bonus: Linear regression in JAX linear_regression_jax.ipynb
  • Bonus: automatic differentiation with dual numbers AD_with_dual_numbers_Julia.ipynb
• Homework 1: MLP from scratch
  • hw1_mlp.ipynb and its solution hw1_mlp_sol.ipynb
• Module 3: Loss functions for classification
  • An explanation of underfitting and overfitting with polynomial regression 03_polynomial_regression.ipynb
• Module 4: Optimization for deep leaning
  • Practical: code Adagrad, RMSProp, Adam, AMSGrad 04_gradient_descent_optimization_algorithms_empty.ipynb and its solution 04_gradient_descent_optimization_algorithms_sol.ipynb
• Module 5: Stacking layers
  • Practical: overfitting a MLP on CIFAR10 Stacking_layers_MLP_CIFAR10.ipynb and its solution MLP_CIFAR10.ipynb
• Module 6: Convolutional neural network
  • Practical: build a simple digit recognizer with CNN 06_convolution_digit_recognizer.ipynb
• Homework 2: Class Activation Map and adversarial examples
  • HW2_CAM_Adversarial.ipynb
• Module 8: Embedding layers, Collaborative filtering and Word2vec
  • Practical: Collaborative filtering with Movielens 100k dataset 08_collaborative_filtering_empty.ipynb
  • Practical: Refactoring code, collaborative filtering with Movielens 1M dataset 08_collaborative_filtering_1M.ipynb
  • Practical: Word Embedding (word2vec) in PyTorch 08_Word2vec_pytorch_empty.ipynb
  • Finding Synonyms and Analogies with Glove 08_Playing_with_word_embedding.ipynb
• Module 9a: Autoencoders
  • Practical: denoising autoencoder (with convolutions and transposed convolutions) 09_AE_NoisyAE.ipynb
• Module 9b - UNets
  • UNet for image segmentation UNet_image_seg.ipynb
• Module 9c - Flows
  • implementing Real NVP Normalizing_flows_empty.ipynb and its solution Normalizing_flows_sol.ipynb
• Module 10 - Generative Adversarial Networks
  • Conditional GAN and InfoGAN 10_GAN_double_moon.ipynb
• Module 11 - Recurrent Neural Networks and Batches with sequences in Pytorch
  • notebook used in the theory course: 11_RNN.ipynb
  • predicting engine failure with RNN 11_predictions_RNN_empty.ipynb
• Module 12 - Attention and Transformers
  • Correcting the PyTorch tutorial on attention in seq2seq: 12_seq2seq_attention.ipynb and its solution
  • building a simple transformer block and thinking like transformers: GPT_hist.ipynb and its solution
• Module 13 - Siamese Networks and Representation Learning
  • learning embeddings with contrastive loss: 13_siamese_triplet_mnist_empty.ipynb
• Module 15 - Dropout
  • Dropout on a toy dataset: 15a_dropout_intro.ipynb
  • playing with dropout on MNIST: 15b_dropout_mnist.ipynb
• Module 16 - Batchnorm
  • impact of batchnorm: 16_batchnorm_simple.ipynb
  • Playing with batchnorm without any training: 16_simple_batchnorm_eval.ipynb
• Module 18a - Denoising Diffusion Probabilistic Models
  • Denoising Diffusion Probabilistic Models for MNIST: ddpm_nano_empty.ipynb and its solution ddpm_nano_sol.ipynb
  • Denoising Diffusion Probabilistic Models for CIFAR10: ddpm_micro_sol.ipynb
• Module - Deep Learning on graphs
  • Inductive bias in GCN: a spectral perspective GCN_inductivebias_spectral.ipynb and for colab GCN_inductivebias_spectral-colab.ipynb
  • Graph ConvNets in PyTorch spectral_gnn.ipynb
• NERF
  • PyTorch Tiny NERF tiny_nerf_extended.ipynb

Usage

If you want to run locally, follow the instructions of Module 0 - Running the notebooks locally

2020 version of the course

Archives are available on the archive-2020 branch.

This site is open source. Improve this page.