Image classifier ML system design на собеседовании Data Scientist

Постановка задачи

Image classifier для product photos / medical / content moderation.

Constraints:

• 1M+ images training.
• p99 latency на inference.
• Continuous data flow (new categories, new images).

Data pipeline

Raw images → annotation → preprocessing → train / val / test split → augmentation → model.

Annotation. Internal labelers (Label Studio, V7) или external (Toloka).

Preprocessing. Resize, normalize, format consistency.

Quality check. Label disagreement, mis-labeled, near-duplicates.

Augmentation

Increase effective dataset.

• Random crop / resize.
• Flips, rotations.
• Color jitter. Brightness, contrast, saturation.
• CutMix, MixUp. Stronger.
• AutoAugment. Learned policy.
• RandAugment. Simple, effective.

Improves generalization 1-5%.

Transfer learning

Pre-trained backbone. ResNet, EfficientNet, ViT, ConvNeXt.

Pre-trained on ImageNet (1M images, 1000 classes).
Fine-tune на your domain (10k-100k images).

Strategies:

• Linear probe. Freeze backbone, train classifier head. Fast, OK.
• Full fine-tune. All weights. Best, более slow.
• Discriminative fine-tune. Different LR per layer (high для head, low для early).

Часто DINO / CLIP backbone better чем ImageNet — modern self-supervised.

Deployment

Model formats. ONNX → TensorRT для GPU, CoreML для iOS.

Quantization. INT8 — 2-4× speedup, < 1% accuracy drop.

Batching. Triton Inference Server. Dynamic batching для throughput.

Edge. MobileNet / EfficientNet-Lite для phones.

Cloud. Bigger models, better accuracy. Latency network-bound.

Edge cases

Class imbalance. Rare classes underperform. Class weights / oversampling.

Domain shift. Train на stock photos → real noisy. Augmentation, fine-tune на target.

Adversarial. Robustness — adversarial training.

OOD detection. Confidence calibration. Reject если low confidence (vs misclassify confidently).

Concept drift. New categories появляются. Continual learning или periodic retrain.

Связанные темы

• CNN-архитектуры для DS
• Image segmentation для DS
• YOLO и DETR для DS
• Inference optimization для DS
• Подготовка к собесу Data Scientist

FAQ

Это официальная информация?

Нет. Статья основана на индустриальных computer vision practices.

Тренируйте Data Science — откройте тренажёр с 1500+ вопросами для собесов.