Hyperparameter Optimization through Sweeps
Created on September 30|Last edited on September 30
Comment
Sweep Results
Sweep: fxzqsh70 1
48
Sweep: fxzqsh70 2
0
The sweep figure above shows the results of the hyper parameter optimization. We span across resnet18, regnet_x_400mf, and mobilenet_v3_small architectures. We do not use the convnext_tiny architecture mainly because it requires more computational resources which were not available. The metric that we are aiming to maximize is the mean IoU.
Best Model
The best model has the following hyperparameter values:
- arch = resnet18
- batch_size = 4
- img_size = 240
- lr = 0.00040686107262288287
Run set
1
We see an improvement from 34.7% from the baseline model to 38.3%. Although this is not a huge improvement, considering that we have limited computational resources, this is decent.
Add a comment