Homework 3

Submission date: January 16th, 2020

Topics

• Sequence models for text generation
• Image generation with a Variational Autoencoder
• Generative adversarial networks

Downloading

The assignment code is available here.

Updates

1. 2020-01-06
   • Part 1: Fixed order of arguments in the post_epoch_fn for training.
   • Part 2: Correction in the formula for the reparametrization trick: $\sigma^2$ replaced with $\sigma$.

   To update, simply replace the original notebooks with the new ones. No code modifications are necessary, but make sure you implemented the reparametrization trick using the correct formula.

2. 2020-01-08
   • Part 3: There was an unnecessary break statement in the GAN training block in the notebook.

   To update, replace the original notebook. No code modifications are necessary.

FAQ

Make sure to read the getting started page and the guide for using course servers.

Q: What is the checkpoint_file_final for?
A: You can use this to create your final submission with result images from your best-trained model. Just train with checkpoints enabled, and when you get results that your happy with rename the checkpoint file with _final. You don’t need to submit the checkpoint files (the main.py script will ignore them).

Q: How can we run long training blocks in the notebooks without running them interactively in jupyter-lab (e.g. from command line on the server)?
A: The easiest way is to simply copy the block (and relevant import statements) into a new python script and run that (with srun/sbatch on the server). A more automated way is to convert the whole notebook to a python script, for example:

jupyter nbconvert Part1_Sequence.ipynb --to python

And then run it with ipython within srun or sbatch, for example:

srun -c 2 --gres=gpu:1 ipython Part1_Sequence.py