search

capstone2019-neuralsearch / AC297r_2019_NAS

Harvard IACS Data Science Capstone: Neural Architecture Search (NAS) with Google
5 stars 1 forks source link

Adapt DARTS to general image datasets (test on MNIST/FashionMNIST first) #2

Closed JiaweiZhuang closed 4 years ago

JiaweiZhuang commented 4 years ago

We should make the DARTS code (https://github.com/quark0/darts) data-independent so it can easily run on all kinds of scientific datasets.

Currently the darts/cnn/train_search.py script has CIFAR10 hard-coded at this line:

train_data = dset.CIFAR10(root=args.data, train=True, download=True, transform=train_transform)

Also hard-coded is CIFAR_CLASSES = 10 the at here. Not sure if there are other hard-coded parameters.

Ref: quark0/darts#57

JiaweiZhuang commented 4 years ago

There would be two steps:

  1. Make it work with general image classification problems. The standard MNIST/FashionMNIST would be enough. Maybe useful to create a small unit test for different datasets. Their current code doesn't seem to have any tests.
  2. Generalize to regression problems, which are more relevant to many scientific data. This requires more tweaks to the model.