引導式測驗 · 中等 等級
用于训练可视化测验的 TensorBoard
使用 TensorBoard 標量、圖像、直方圖和配置文件,同時保持步驟、運行比較、隱私和評估定義清晰。
問題 1 的 8
What does TensorBoard display from a training program?
預覽每個問題8 問題
當您參加測驗時,就會出現答案和解釋。
- What does TensorBoard display from a training program?
- It reconstructs missing metrics from the model architecture
- It automatically logs every intermediate tensor without instrumentation
- Logged summaries and traces emitted by training or profiling code
- It reruns training from prior event files
- Why use a consistent global step for scalar logs?
- To make each run take the same wall-clock time
- To align measurements across training runs
- To force each experiment to log every batch
- To make the underlying metrics identical
- What can image summaries help diagnose?
- Whether predictions are calibrated
- Whether every class is balanced
- Whether the evaluation split is unbiased by itself
- Input and output preprocessing or visual error patterns
- What does a histogram of weights or activations show?
- A distribution of logged tensor values
- The sequence order of training examples
- A guarantee that learned features are correct
- The calibrated probabilities of every class
- How can scalar smoothing mislead debugging?
- It changes the model weights
- It removes the raw event file
- It can hide short-lived spikes or changes
- It corrects mislabeled data
- Why should run comparisons include configuration and metric definitions?
- TensorBoard infers every hidden parameter
- Different settings or definitions can make curves incomparable
- It makes all runs use one dataset
- It removes the need to log steps
- What risk arises when a dashboard logs raw user images or text?
- Event files automatically strip identifiers before sharing
- TensorBoard never stores image or text summaries
- Only run names are exposed when an event directory is shared
- Event files may expose sensitive content when shared
- What does TensorBoard not automatically provide by itself?
- Complete experiment lineage such as data version and code revision
- Scalar plots
- Logged images
- Run names