Review this PyTorch validation function.
Validation accuracy is noisy across runs and lower than expected for a model with dropout and batchnorm. What's wrong?