<p dir="ltr" style="line-height: 1.38; text-indent: 36pt; margin-top: 0pt; margin-bottom: 0pt;"><span id="docs-internal-guid-7137b2dc-7fff-d811-9c5c-985bb19de0f5">[size= 11pt; font-family: Lexend, sans-serif; background-color: transparent; font-variant: normal; vertical-align: baseline; white-space: pre-wrap]Keras Services doesn't say "see you later" when it returns a model. It does all kinds of optimizations behind the scenes, like scheduling learning rates, mixed-precision training and early stopping callbacks via tf.data and building input pipelines. All together, these training hacks that reduce memory (training in 16-bit Floating Point precision), training fast with multiple batch feeding (not limiting the number of batches that can be processed) and training until it's not random anymore (stopping training when we hit a plateau) are magical. Unless your training takes forever, people don't talk about callbacks and tf.data. [/size]</span>