snippets.md

Print network parameters' shapes

print([param.shape for param in self.network.parameters()])

Ensure that network parameters are being updated

before = torch.cat([param.clone().detach().flatten() for param in self.network.parameters()])
# forward, loss, backward, step
after = torch.cat([param.clone().detach().flatten() for param in self.network.parameters()])
assert torch.nonzero(after - before).shape[0] == before.shape[0]

Print currently running diode_clipper scripts of the user

ps -x | grep diode_clipper

Run profiler on a script and output to a file

python -m cProfile -o diode_ode_numerical_profile.bin diode_clipper\diode_ode_numerical.py -u 38 -l 1 -s 5 -i 0 -m forward_euler -f 22050

Count the number of parameters in a model

def count_parameters(model):
    return sum(p.numel() for p in model.parameters() if p.requires_grad)

Delete all files created today from a directory except for pngs

find DIR -daystart -ctime 0 -type f -not -name '*.png' -print -delete

Command for generating oscillator trajectories in the DAFX paper

python harmonic_oscillator\main.py --epochs 1 --visualize --interpolation linear -c 0.2 --nsteps 1000

Triton modules needed to run training sessions

• gcc
• cuda

module list output:

1. gcc/9.2.0
2. libiconv/1.16
3. xz/5.2.5
4. zlib/1.2.11
5. libxml2/2.9.12
6. cuda/11.2.1

Triton Interactive Shell

srun -p interactive --time=01:00:00 --gres=gpu:1 --mem=3000M --pty bash
srun -p interactive --time=01:00:00 --mem=3000M --pty bash