Kvíz spojený s průvodcem · Střední úroveň
TensorFlow Lite a LiteRT pro mobilní kvíz
Pochopte přechod pojmenování TensorFlow Lite na LiteRT, konverzi, kvantizaci, delegování a ověření cílového zařízení.
Otázka 1 z 8
What does LiteRT refer to in current Google AI Edge documentation?
Náhled každé otázky8 otázky
Odpovědi a vysvětlení se zobrazují, jak budete v kvízu.
- What does LiteRT refer to in current Google AI Edge documentation?
- A dataset format used to train mobile models
- A separate runtime unrelated to TensorFlow Lite
- The on-device runtime historically known as TensorFlow Lite
- A chipset-specific accelerator delegate
- What does a hardware delegate do in an on-device inference runtime?
- Repartitions a model graph during training
- Runs supported operations through an accelerator backend
- Guarantees every graph operation runs on an NPU
- Replaces the model's preprocessing and postprocessing
- Why check operator compatibility during model conversion?
- It changes model architecture and labels automatically
- It ensures all operators run on an accelerator
- It selects the app's minimum operating-system version
- Unsupported operations may block conversion or require a different runtime path
- What can quantization trade for smaller models or faster execution?
- Some numerical accuracy
- Every model operator remains available in every runtime
- The need to benchmark on target devices
- The app's minimum operating-system version
- Why test an NPU delegate on the intended phone?
- All phone NPUs use identical operator coverage
- The model artifact determines which phone chipset is installed
- Delegate availability and supported operations vary by device and runtime
- An NPU guarantees execution of every model operation
- What should match between model training and mobile inference?
- Match the model architecture while allowing mobile preprocessing to differ
- Input preprocessing and output interpretation
- Match only the model-file extension
- Match only the training framework's interface
- Which benchmark covers the mobile user experience more fully?
- A desktop conversion time alone
- Training-set accuracy alone
- The exported file size alone
- Cold load, warm latency, memory, and energy on target devices
- How should TensorFlow Lite and LiteRT references be handled during maintenance?
- Check version-specific migration and package guidance before changing APIs
- Assume every older tutorial matches current package names
- Rename each model file before any old app can load it
- Ignore runtime and model artifact compatibility