lucasdavid / algorithms-in-tensorflow

Experiments with gpu support, persistent logging and stable docker env.

Algorithms in Tensorflow

Experiments with tensorflow 2.0, gpu support, persistent logging and stable docker env.

Summary

Supervised

Notebook	Kind	Description	References
Intro	classification	Introduction over classification, using Logistic Regressison and training with Grid Search Cross-validation	dataset
Multi-class	classification	Introduction to multi-class classification and examples
Study of Churn	classification	Study over client churn behavior
Study of Facts	classification	Study over a set of true and fake news
Churn	regression	Introduction over linear regression
Study of Weather WW2	regression	Instantial study over Weather information during WW2	dataset
cifar10	fine-tuning classification	CNN fine-tuned from ImageNet to solve Cifar10	dataset
Barzinga	fine-tuning classification	Classifying objects from a low-budget WebCAM
best-artworks-of-all-time	fine-tuning classification	Art authorship attribution fine-tuned from ImageNet	BAoAT dataset
Study of Mapping Challenge	semantic segmentation	Segmentation of construction satellite images using U-NET and EfficientNetB4	Mapping Challenge
Study of Oxford IIT Pet	semantic segmentation	Segmentation of dogs and cats from the oxford_iit_pet	Oxford IIT Pet dataset

Unsupervised

Notebook	Kind	Description	References
gatys	style transfer	Style transfer between two arbitrary images	article
VaE Cifar10	variational autoencoder	VaE over Cifar10 and embedding of the set in the metric space	tutorial
VaE Fashion MNIST	variational autoencoder	VaE over Fashion MNIST	tutorial
Contrastive loss	siamese network	Siamese CNN trained with contrastive loss	article

Knowledge Distillation

Notebook	Kind	Description	References
vanilla	Response-based Distillation	A small network is trained to emulate the output of a larger one.	article
FitNet	Activation-based Distillation	A network better generalizes when regularized by a larger one.	article

Explaining

Notebook	Kind	Description	References
Activation Maximization	optimization	Network concept-representation by gradient ascending over the output value	class notes
Grad-CAM	CAM	Explaining networks' decision using gradient info and CAM	article
Grad-CAM++	CAM	Adjust Grad-CAM weights to prevent activation domiance of large regions over small ones	article
Score-CAM	CAM	CAM based on Increase of Confidence	article
Gradient Backpropagation	saliency	Gradient-based
Guided Gradient Backpropagation	saliency	Gradient-based explaining method considering positive intermediate gradients	article
Smooth Gradient Backpropagation	saliency	Gradient-based explaining method with local-level gradient correction	article
Full Gradient Representation	saliency	Explaining using function linearization with gradient-based and bias information	article

Usage

Code in this repository is kept inside jupyter notebooks, so any jupyter server will do. I added a docker-compose env to simplify things, which can be used as follows:

./actions/run.sh                                          # start jupyter notebook
./actions.run.sh {up,down,build}                          # more compose commands
./actions.run.sh exec experiments python path/to/file.py  # any commands, really
./actions/run.sh tensorboard                              # start tensorboard