Created
May 10, 2019 12:00
-
-
Save chicobentojr/d20dd040ff957d24d43a94cdf92e913e to your computer and use it in GitHub Desktop.
"A Recipe for Training Neural Networks" - Andrej Karpathy
- Neural net training is a leaky abstraction
- Neural net training fails silently
- Become one with the data
- Set up the end-to-end training/evaluation skeleton + get dumb baseline
- fix random seed
- simplify
- add significant digits to your eval
- verify loss @ init
- init well
- human baseline
- input-indepent baseline
- overfit one batch
- verify decreasing training loss
- visualize just before the net
- visualize prediction dynamics
- use backprop to chart dependencies
- generalize a special case
- Overfit
- picking the model
- adam is safe
- complexify only one at a time
- do not trust learning rate decay defaults
- Regularize
- get more data
- data augment
- creative augmentation
- pretrain
- stick with supervised learning
- smaller input dimensionality
- smaller model size
- decrease the batch size
- drop
- weight decay
- early stopping
- try a larger model
- Tune
- random over grid search
- hyper-parameter optimization
- Squeeze out the juice
- ensembles
- leave it training
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment