model failed fidelity test

faseelmo / noc_latency_predictor

MIT License

1 stars 0 forks source link

model failed fidelity test #32

Closed faseelmo closed 9 months ago

faseelmo commented 10 months ago

Try Implementing relative loss
I feel like the model somewhat captures the task latency but not the effect of mapping on latency
Add a shit ton of data. Especially mapping. See when they start to converge.
Increase the parameters in the Model
Maybe add something like Resnet in GCN?

faseelmo commented 10 months ago

Think if task having task id as a feature is useful. Find a better way to show mapping on the feature than using Manhattan distance. Learn how XY routing works.

faseelmo commented 10 months ago

Train on one single task, but a large dataset of different mappings.

use those weights then to train on multiple tasks. Maybe freeze some weights?