Finish setting up network on CSUA + team check-ins

Meeting 3 recap

• Setting up network on CSUA

Meeting 4 agenda

What we’ll do at this meeting is finish setting up the network on CSUA, do a Pytorch tutorial so that everyone can get some experience writing Pytorch code from scratch, and discuss what we want to work on in teams.

My hope is that every team will be able to complete an experiment before Spring Break, and write a post about it on the website.

• Finish CSUA setup (30min)
• Pytorch tutorial (45min)
• Team check-ins (30min)
• Work in teams (60min)

CSUA setup

1. Clone the project github and go to the music-translation subfolder
2. Create and source a virtualenv
3. Add the virtual env to Jupyter as a kernel edit 2 of this answer
4. Install the requirements.txt and go to src/nv-wavenet and python3 setup.py install to install the CUDA kernel
5. Create symlinks with ln -s (see https://stackoverflow.com/questions/1951742/how-can-i-symlink-a-file-in-linux)

   One symlink from /datasets/musicnet/musicnet to musicnet in the music-translation subfolder Another from /datasets/muscinet/checkpoints to checkpoints

Please be careful not to delete or overwrite checkpoints, as it’s shared. Also, it’s probably good practice to back up important checkpoints to your home folder.

1. Try running Generate_modified.ipynb in notebooks (might have to change things to get it to work. Particularly change from 3 to 0 the last argument of NVWavenetGenerator constructor)
2. Or try running train2.sh or sample2.sh [decoder][path/to/checkpoint/(lastmodel|bestmodel)] I forget what train2 expects in terms of parameters. They’re basically just wrapper scripts that call the training/evaluation with parameters. I just modified the provided train.sh and sample.sh.
3. Be sure to run any training or evaluation in tmux if you want it to run in the background. Just open a tmux and then you can close the ssh, and return to it later with tmux a #.

Pytorch tutorial

Link

Team check-ins

Team check-ins this week will be

• What experiment do you want to do this week?
• What steps toward completing that experiment do you want to complete at this meeting, and when/how do you plan to finish the experiment outside of the meeting?
• What would the website post about the experiment look like? Who will write it?
• When do you expect the experiment to be finished by?

Transfomers (James, Alicia)

Network training (Praveen, Andrew)

Pretrained Model (Jasmine, Jim)