kaggle dataset link: https://www.kaggle.com/fanconic/skin-cancer-malignant-vs-benign
TODO:
-
Look for other medical data sets, hopefully binary
-
Make seperate Interactive Notebook: Ideas
- show our tops losses, in case doctor shows
- leaderboard
- compares to latest papers?
- visualizes what the model looks like
- shows the heatmap of what the model focuses on
Model Tuning/Improvement:
-
fp-16 training
-
weight decay parameter increase, 0.01 - 0.01,0.2? Separate Batch norm param? a
-
hotdog, not-hotdog ensemble
-
what is spatial dropout?
-
include all validation data for training
-
try out final batchnorm layer(done, unsure)
-
slice learning rates, pct-start?(done, unsure)
-
iterative resizing (done)
-
mess with data augmentation, warp, lighting, rotate (done)
-
Get dropout in there, not sure if default (done, default 0.5)
Maybe:
- Check latest papers, architecture, tricks, data-sets? (maybe for a sequel project)