roadmap
- 01
- 02
- 03
- 04
- 05
- 06
- 07
- 08
Prepare for TensorFlow interview questions on training and deploying models. Covers tf.function and tf.data, Keras models, tf.distribute, and SavedModel serving and TFLite on device.
77 of 77 questions in your plan: 0 known, 0 in progress, 0 due, 0 skipped, 77 new. 0 cut, outside your plan.
TensorFlow comes up in machine learning engineer and applied scientist loops at teams that train with it, in MLOps interviews about getting a model into serving, and in mobile interviews for apps that run a model on the phone. Its usual home is another round: a project deep-dive, where you defend a model you trained, or a practical exercise built on a notebook, where you write a training loop with tf.GradientTape or explain why a job leaves the GPU idle for most of each step.
The questions are mostly symptoms: a tf.function that slows down because it retraces whenever it is passed a new Python number, or a model that scores well in the notebook and badly behind the server, because its preprocessing lived in Python and never reached the export.
Answers are told apart by whether they stop at the symptom or reach the line that causes it. "The input pipeline is slow" is a start; the answer is that map handles one element at a time without num_parallel_calls, and nothing readies the next batch during the current step until prefetch is added. A failed fine-tune gets the same treatment: trainable = False on a BatchNormalization layer also switches it to its stored statistics, which freezing does not do to most layers. Higher up the level range the talk turns to costs: which distribution strategy suits the hardware, how batch size and learning rate move with the number of replicas, and how much accuracy a full-integer model gives up on a phone.
Read top to bottom, the plan follows one model from its first tensor to a phone. Tensors and graphs lead because nearly everything later runs through them: fit compiles its training step with tf.function, and a SavedModel stores traced functions beside the weights. The input pipeline comes before the model, since a model trains only as fast as it is fed. Keras makes its first appearance next; then come the training step itself, its spread across devices once it is understood, and the SavedModel that the later stages start from.
The second module returns to Keras as a library in its own right, one level deeper: what each of the ways to define a model makes harder to inspect or save, how build lets a custom layer create its weights from the first input shape, and how overriding train_step keeps what fit provides around a custom loss. Callbacks and saving close it, and saving is where multi-backend Keras matters, because a .keras file and an exported SavedModel are different artifacts.
TFLite comes last, in the order the work is done. Conversion and quantization start from an exported model and decide which numeric types the file holds; the interpreter's delegates follow, because whether a GPU or NPU delegate can run an operation depends on what conversion left behind, and anything it cannot run falls back to the CPU.
A PyTorch user will find eager tensors and tape-recorded gradients familiar. The new idea is tracing, where Python stops running and a graph starts, so begin at the first topic and keep the printed sequence by moving study order to by topics. Senior candidates for ML platform roles should put their hours into distribution and serving. Android and iOS engineers adding a model to an app can go straight to the last module after one detour through the SavedModel topic, since conversion usually starts from that export. When the date is close and the job is server-side, spend it on the first two modules with importance raised to must know; the on-device stretch earns its hours when the posting names a phone or an embedded target. The plan teaches the framework, not the modeling decisions it carries out: which architecture and loss fit a problem, which metric to trust when one class is rare, and when a gradient-boosted tree beats a network on tabular data. That is the fundamentals round of an ML interview, and the [Machine Learning Engineer roadmap](/roadmaps/machine-learning) prepares for it. For a team that trains in PyTorch, the [PyTorch roadmap](/roadmaps/pytorch) is the counterpart in the other framework.
A PyTorch user will find eager tensors and tape-recorded gradients familiar. The new idea is tracing, where Python stops running and a graph starts, so begin at the first topic and keep the printed sequence by moving study order to by topics. Senior candidates for ML platform roles should put their hours into distribution and serving.
Android and iOS engineers adding a model to an app can go straight to the last module after one detour through the SavedModel topic, since conversion usually starts from that export. When the date is close and the job is server-side, spend it on the first two modules with importance raised to must know; the on-device stretch earns its hours when the posting names a phone or an embedded target.
The plan teaches the framework, not the modeling decisions it carries out: which architecture and loss fit a problem, which metric to trust when one class is rare, and when a gradient-boosted tree beats a network on tabular data. That is the fundamentals round of an ML interview, and the Machine Learning Engineer roadmap prepares for it. For a team that trains in PyTorch, the PyTorch roadmap is the counterpart in the other framework.