quiz bu lëkkale ak tegtal · Diggante Niveau

Jàng PyTorch: ab laaj ak tontu ci kàrtu yoon

Tabax PyTorch xam-xam ci toppalante ci tensor ak autograd ba ci module, tàggat, jàngat, saytu ak natt gëna xéewale.

Yooni njiit yu jëm ciJàng kàrtu yoonu Pytorch
Laaj 1 de 8

What does PyTorch autograd compute for differentiable tensor operations?

Xoolaatal laaj bu nekk8 laaj

Sooy def quiz bi, tontu yi ak leeral yi dañuy feeñ.

  1. What does PyTorch autograd compute for differentiable tensor operations?
    • Predictions and class labels only
    • Parameter updates performed by the optimizer
    • Gradients of tracked differentiable operations
    • DataLoader batch order
  2. Why clear gradients between optimizer updates in a typical training loop?
    • Gradients are overwritten on every backward call
    • Gradients accumulate by default across backward calls
    • Gradients are cleared automatically by optimizer.step
    • Each parameter receives a gradient only once per epoch
  3. What does model.eval() change for mode-dependent modules?
    • It permanently freezes every parameter
    • It disables autograd for all code that follows
    • It resets optimizer state and gradients
    • It changes mode-dependent modules such as dropout and batch normalization to evaluation behavior
  4. Which division of responsibility between Dataset and DataLoader is correct?
    • Dataset defines per-example access; DataLoader batches and iterates over examples
    • Dataset defines model layers; DataLoader computes the loss
    • Dataset stores optimizer state; DataLoader updates gradients
    • Dataset selects a checkpoint; DataLoader changes evaluation mode
  5. What does a model state dictionary contain?
    • A serialized Python architecture and all source code
    • Parameter and buffer values to load into a recreated module
    • The entire training dataset and split definitions
    • The final test metric and evaluation report
  6. Why profile data loading and transfer along with GPU computation?
    • The pipeline may be limited by input preparation or transfer, not just model kernels
    • GPU time always dominates when the model is on a device
    • Data loading cannot affect end-to-end throughput
    • Only the number of layers determines request latency
  7. Which evaluation design is appropriate when tuning PyTorch checkpoints?
    • Choose the best checkpoint on final test data repeatedly
    • Train and report on the same batches only
    • Use test labels to set augmentation
    • Use validation data for selection and keep a final test separate
  8. What should precede advanced optimization such as mixed precision or compilation?
    • A correct baseline and profiling evidence
    • Compile before measuring a baseline
    • Enable mixed precision for every dtype
    • Increase batch size without checking memory