It’s pretty easy to train models deterministically. Just use the same batch size, float type, initial seed, and flip this flag: https://www.tensorflow.org/api_docs/python/tf/config/experim...