Placeholder Image

字幕表 動画を再生する


  • SPEAKER: In this video, I'll show you

  • how you can use Autograph to write complex, high performance

  • TensorFlow code using normal Python.

  • Autograph is available in the new TF2 function API,

  • makes it easy to run TensorFlow computations in a way

  • that's efficient and portable.

  • When you annotate a Python function with tf.function,

  • Autograph will automatically convert its Python code

  • to TensorFlow graph code.

  • The code is then compiled into a graph

  • and executed when you call the function.

  • Let's look at an example.

  • This simple function calculates the square of a scalar input

  • if it's positive.

  • In TensorFlow 2.0, you don't have to use tf.cond anymore.

  • You can just write a normal if statement,

  • and Autograph will generate a tf.cond operation

  • so that the entire computation runs as a graph.

  • This is the generated code that Autograph writes for you.

  • Notice that we're writing true and false functions that

  • would normally be fed into a tf.cond statement.

  • Instead of writing these, you can simply

  • use Python if statements.

  • Let's take a look at a more complicated example.

  • This is a bare bones RNN cell.

  • Note that it contains a data dependent for loop,

  • and it also contains a data independent if statement.

  • Autograph will only run the data dependent loop in the graph

  • and leave the data independent if statement untouched.

  • Simply adding a tf.function as a decorator

  • still lets you call the function directly and get results

  • immediately.

  • But the function runs in graph mode.

  • It prints results.

  • And we can also time it.

  • Now, if we remove the tf.function decorator, which

  • I've preemptively done here, and run the function in eager mode,

  • we get the same results out.

  • However, it's going to be a little bit slower because we

  • won't have coalesced the entire function into a single tf.graph

  • op.

  • We can time both options with tf.function in Autograph

  • and without.

  • You'll note that using tf.function, which

  • requires only a single function decorator,

  • is significantly faster than the eager mode version

  • without tf.function.




動画の操作 ここで「動画」の調整と「字幕」の表示を設定することができます

B2 中上級

AutoGraph。グラフのための簡単な制御フロー(TensorFlow Tip of the Week (AutoGraph: Easy control flow for graphs (TensorFlow Tip of the Week))

  • 0 0
    林宜悉 に公開 2021 年 01 月 14 日