引導式測驗 · 硬 等級
機器學習的 JAX 與 XLA 測驗
將 JAX 轉換連接到自動微分、向量化、編譯和並行執行,同時考慮追蹤行為。
問題 1 的 8
哪一個 JAX 轉換計算可微函數的梯度?
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。
- 哪一個 JAX 轉換計算可微函數的梯度?
- jax.vmap
- jax.jit
- jax.grad
- jax.device_put
- jax.vmap 提供什麼?
- 即時編譯
- 映射軸上的自動向量化
- 自動微分
- 顯式隨機密鑰分割
- jax.jit 扮演什麼角色?
- 計算標量目標的梯度
- 跨批次向量化函數
- 將陣列放置在選定的設備上
- Compiles a compatible function for efficient execution through XLA
- Why can a Python if statement fail inside a jitted function when its condition uses an array value?
- Tracing cannot use a runtime array value as an ordinary Python boolean
- Python evaluates every array-dependent if during tracing
- An array value always converts to a Python bool inside jit
- The compiled function specializes on runtime values automatically
- How does explicit random-key passing fit JAX's programming style?
- It stores one hidden global generator
- It removes random variation from training
- It makes randomness an explicit input to computations
- It fixes package versions
- Why may changing input shapes cause additional JAX compilation?
- The model must retrain for each input
- Compilation can specialize to shapes and static arguments
- XLA discards the previous dataset
- Gradients cannot use arrays
- What should be done when timing asynchronous device execution?
- Measure only Python dispatch time
- Skip warmup for compiled functions
- Change the random seed during timing
- Synchronize the result before stopping the timer
- Which statement about JAX versus PyTorch is sound?
- Choose based on workload, ecosystem and team needs
- JAX is always faster for every model
- PyTorch cannot run on accelerators
- The frameworks have identical APIs