Computer Vision Interview 20 essential Q&A Updated 2026
TF Vision

TensorFlow Vision: 20 Essential Q&A

tf.data pipelines, Keras applications, and deployment to mobile with TFLite.

~11 min read 20 questions Intermediate
tf.dataKerasKerasCVTFLite
Quick Navigation
1. TF vision stack
2. tf.data.Dataset
3. Preprocessing layers
4. keras.applications
5. Finetuning
6. TPU strategy
7. Distributed
8. SavedModel
9. TFLite conversion
10. Quantization
11. TensorFlow Serving
12. KerasCV
13. vs PyTorch
14. TFRecord
15. tf.function
16. tf.io image decode
17. Losses & metrics
18. Profiler
19. TF Hub
20. Debugging
1 TensorFlow vision stack? ⚡ easy
Answer: Keras high-level API + tf.data for input pipelines + optional KerasCV for detection/segmentation blocks.
2 Why tf.data? đź“Š medium
Answer: Declarative, parallel map/prefetch/batch—keeps GPU fed; cache() and prefetch(AUTOTUNE) are standard patterns. 
ds = tf.data.Dataset.list_files("*.jpg").map(decode_fn).batch(32).prefetch(tf.data.AUTOTUNE)
3 Preprocessing layers? đź“Š medium
Answer: Augment inside model (RandomFlip, RandomRotation) for exportable training graph—same code path in TFLite if supported.
4 keras.applications? ⚡ easy
Answer: Pretrained ResNet, EfficientNet, etc.—set include_top false for feature extraction; load imagenet weights.
5 Finetuning recipe? 🔥 hard
Answer: Freeze base, train head; unfreeze top layers with small LR; use early stopping—watch catastrophic forgetting on tiny data.
6 TPU strategy? 🔥 hard
Answer: TPUStrategy scope + dataset stored in GCS—batch size and image size constraints differ from GPU pipelines.
7 MirroredStrategy? đź“Š medium
Answer: Data-parallel multi-GPU on one host—simple path to scale batch without manual gradient sync code.
8 SavedModel? đź“Š medium
Answer: Directory with graph + variables + signatures—TF Serving and TFLite consume this interchange format.
9 TFLite conversion? đź“Š medium
Answer: TFLiteConverter.from_keras_model—select ops, FP16/INT8 quantization for mobile/NPU deployment.
10 Quantization-aware training? 🔥 hard
Answer: Simulate low precision during training—better INT8 accuracy than post-training quant on hard models.
11 TensorFlow Serving? đź“Š medium
Answer: Model server with REST/gRPC—batching scheduler for production latency/throughput tradeoffs.
12 What is KerasCV? đź“Š medium
Answer: Modular OD/seg components (YOLO, RetinaNet building blocks), augmentations—accelerates TF-native vision research.
13 TF vs PyTorch (interview angle)? đź“Š medium
Answer: TF: deployment/serving story, TPU; PyTorch: research ergonomics—both mature for vision; know your team stack.
14 TFRecord? ⚡ easy
Answer: Serialized Example protos for large datasets—efficient sequential read on GCS; not mandatory for small local image folders.
15 @tf.function? đź“Š medium
Answer: Trace graph for performance—avoid Python side effects; use Tensor ops inside for autograph success.
16 tf.io.decode_image? ⚡ easy
Answer: Decode JPEG/PNG to uint8 tensor—pair with convert_image_dtype for float pipeline.
17 Losses for segmentation? đź“Š medium
Answer: SparseCategoricalCrossentropy with logits, Dice/Focal in add-ons—match activation (sigmoid vs softmax) and mask handling.
18 Profiler? ⚡ easy
Answer: TensorBoard profiler traces input bottleneck vs GPU kernel gaps—optimize prefetch and augmentation placement.
19 TF Hub? đź“Š medium
Answer: Reusable SavedModel modules (image feature vectors)—quick baseline without training from scratch.
20 Debugging? ⚡ easy
Answer: tf.data.take(1), eager execution default in TF2, assert shapes—Static vs dynamic shapes affect XLA.

TF Vision Cheat Sheet

Input
  • tf.data
Models
  • applications
  • KerasCV
Ship
  • TFLite / Serving

đź’ˇ Pro tip: tf.data prefetch; SavedModel for serving and mobile.

Full tutorial track

Go deeper with the matching tutorial chapter and code examples.

TensorFlow Vision Tutorial
Previous Next