Our project for the HDR ML Challenge for anomalous gravitational wave detection.
The current (this is first version) Train.py script expects a YAML config file, giving model, dataset, optimizer, and loss hyperparameters. The best teacher is an example, so here's my initial one:
model:
d_input: 200
hidden_layers: [400, 400]
d_output: 200
dropout: 0.0
dataset:
location: "/path/to/folder/where/npy/files/live"
batch_size: 200
optimizer:
epochs: 15
lr: 1.0e-4
weight_decay: 0.0
loss:
type: "MSE"
The barest information you need in config.yaml is:
- All hyperparameters needed to instantiate your model
- The location of a folder containing your dataset files
Everything else has default values in the script, but it'd be good to include them as well, since it also serves as a form of documentation for your trainings. I even include some things (like the loss "type") that aren't used by the script, and are there for me to look at when I'm browsing through old trainings.