Jul 25, 2017 · 1 min read
You may want to add:
- Use standard data (e.g. mnist, cifar10)
When testing new network architecture or writing a new piece of code, instead of trying with your application data, use the standard datasets first. This is because there are many reference results for these datasets and they are proved to be ‘solvable’. There will be no issues of label noise, train/test distribution difference , too much difficulty in dataset, etc
