checks.rst

Checks Available in Kepler

Code	Description	Applied
K1	Minibatches
K101	Bad minibatch size	Before training begins
K1010	Minibatch too small	Before training begins
K1011	Minibatch too large	Before training begins
K102	Data not shuffled	Before training begins
K103	Train / Dev test not stratified	Before training begins
K104	Train / Dev test normalized separetely	Before training begins
K2	Model Weights
K201	Weights too small	At model definition
K202	Weights outside the unit sphere	At model definition
K203	Uniform weight initialization	At model definition
K204	Data not compatible with chosen initialization	Before training begins
K3	Training Data Statistics
K301	Training inputs correlated	Before training begins
K302	Duplicate training samples	Before training begins
K303	Data not normalized	Before training begins
K4	Activations
K401	Sigmoid activation	At model definition
K402	Bad label coding for tanh	Before training begins
K403	Bad label coding for the last activation	Before training begins
K5	Training
K501	More parameters than training examples	Before training begins
K502	Vanishing or exploding gradients	During training
K503	Bad batchnorm layer position	At model definiton
K504	Model started overfitting	During training
K505	Noisy weight updates	During training