The repository includes modified copies of the original Jupyter notebooks and Python files from the excellent (and really unique) deep learning course "Practical Deep Learning For Coders" Part 1 and Part 2, v.1, created by fast.ai.

The original files require Keras 1. One main goal has been to modify the original files to the minimum extent possible. The comments added to the modules generally start with "# -" when they are not just "# Keras 2".

The current version of the repository has been tested with Keras 2.1.2. The previous version, tested with Keras 2.0.6, is available here.

Located in the nbs folder. Tested on Ubuntu 16.04 and Python 3.5, installed through Anaconda, using the Theano 1.0.1 backend.

Located in the nbs2 folder. Tested on Ubuntu 16.04 and Python 3.5, installed through Anaconda, using the TensorFlow 1.3.0 backend. A few modules requiring PyTorch were also tested, using PyTorch 0.3.0.

The files keras.json.for_TensorFlow and keras.json.for_Theano provide a template for the appropriate keras.json file, based on which one of the two backends needs to be used by Keras.

An environment.yml file for creating a suitable conda environment is provided.

neural-style.ipynb: due to a function parameter change in Keras 2.1, the VGG16 provided by Keras 2.1 has been used instead of the original custom module vgg16_avg.py

rossman.ipynb: section "Using 3rd place data" has been left out for lack of the required data

spelling_bee_RNN.ipynb and attention_wrapper.py: due to the changed implementation of the recurrent.py module in Keras 2.1, the attention part of the notebook doesn't work anymore

taxi_data_prep_and_mlp.ipynb: section "Uh oh ..." has been left out. Caveat: running all the notebook at once exhausted 128 GB RAM; I was able to run each section individually only after resetting the notebook kernel each time

tiramisu-keras.ipynb: in order to run the larger size model I had to reset the notebook kernel in order to free up enough GPU memory (almost 12 GB) and jump directly to the model

neural-style-pytorch.ipynb (found no way to load the VGG weights; it looks like some version compatibility issue)

rossman_exp.py

seq2seq-translation.ipynb

taxi.ipynb

tiramisu-pytorch.ipynb