Pycon 9 - Deep learning from zero to hero

Here both the slides and demos of my talk "Deep learning from zero to hero" at pycon9 2018

System requirements

The only dependence of the demos is docker. All demos run in the docker image rucka/deeplearning

The demos

• 1: The simpson transfer learning

  • Train the model with 30 epochs (accuracy ~50%): ./exec_cmd.sh /code/retrain.fast.sh
  • Train the model with 4000 epochs (accuracy ~90%): ./exec_cmd.sh /code/retrain.sh
  • Evaluate the fast model: ./exec_cmd.sh /code/evaluate.fast.sh /data/simpson/test_set/0.jpg
  • Evaluate the accurated model: ./exec_cmd.sh /code/evaluate.sh /data/simpson/test_set/0.jpg

• 2: Stock price regression forecast: execute the script ./run_book.sh, open the jupyter home link showed in the command logs and select the notebook 1. stock regression.ipynb

• 3: Stock trend classification forecast: execute the script ./run_book.sh, open the jupyter home link showed in the command logs and select the notebook 2. stock classification.ipynb. This notebook contains:

  • Multi layer perceptron network
  • Convolutional network
  • CNN vs MLP
  • Multi value classification

Available scripts

• exec_cmd.sh to run a command inside the container e.g. ./exec_cmd.sh "/code/evaluate.fast.sh /data/simpson/test_set/0.jpg"

• run_book.sh to run a jupyter notebook so you can write and test your code from your browser at http://localhost:8888. If you don't know jupyter, please take a look at the official documentation

• run_board.sh to run the tensorflow board and connect to it from the browser at http://localhost:6006. You can look at the Tensorflow documentation for further informations

• run_shell.sh to open a container shell session

Slides

The slide markup has been compiled with Deckset app for OSX. If you need, you can give a try of the trial version available at the website.

The pdf slides are available here

The slides are also published on slideshare.

Abstract (italian)

Avete sentito parlare di Deep Learning ma credete che la teoria alla base sia troppo complessa? Non avete una laurea in matematica e statistica e pensate che il machine learning non faccia per voi? Niente paura: avrete solo bisogno di una conoscenze di base di Python.

Conoscete la regola dell’80/20? Con il 20% delle conoscenze potete raggiungere l’80% dei risultati: in questo talk vi mostrerò in modo pratico tramite delle demo - alcuni trucchi per costruire dei buoni modelli predittivi, evitando di perdere (tanto) tempo nella scelta dei tools e delle librerie necessarie al vostro scopo.

L’obbiettivo è fornirvi le basi pratiche con cui scegliere un modello di rete neurale, farne training e ottimizzarlo nel modo più adatto alla tipologia del problema che dovete affrontare.