Batch Normalization: Inference With Running Statistics
Intermediate
The problem, from first principles
Inference runs on one example at a time, so batch statistics are unavailable or unstable. Batch norm instead uses running averages collected during training, fixed at inference time.
From theory to code
Implement batchnorm_inference(x, gamma, beta, running_mean, running_var, eps), which normalizes with the stored running statistics.
Constraints
running_meanandrunning_varhave shape(D,).- The output is deterministic for a given input.
Hints
Hint 1
Use running_mean and running_var in place of the batch mean and variance, then scale and shift.
Keywords
research-papersnormalizationbatchnorm
Python
Run your code to see results
No results yet
Run your code to check it against the examples, or submit to run every test.