How to Obtain Runtime Data

OneFlow support oneflow.watch and oneflow.watch_diff. We can use them to register a callback function to get data and gradient tensor in job functions at runtime.

Using Guidance

To get data or gradient tensor in job function, we need to follow these steps:

  • Write a callback function and the parameters of the callback function should be annotated to indicate the data type. The logic of the callback function need to be set up by user themselves.

  • When defining the job functions, we use oneflow.watch or oneflow.watch_diff to register callback function. We obtain data tensor from the former one and their corresponding gradient from the latter one.

  • At the appropriate time when the job function is running, OneFlow will call the previous callback function which was registered earlier and pass the monitored data to the callback function then execute the logic in the callback function.

Take oneflow.watch as example:

  1. def my_watch(x: T):
  2.     #process x
  4. @global_function()
  5. def foo() -> T:
  6.     #define network ...
  7.     oneflow.watch(x, my_watch)
  8.     #...

The T in the code above is the data type in oneflow.typing. Like oneflow.typing.Numpy. Please refer to this article.

We will use the following examples to demonstrate how to use watch and watch_diff.

Use watch to Obtain the Data when Running

The following is an example to demonstrate how to use oneflow.watch to obtain the data from middle layer in OneFlow. Code:test_watch.py

Run above code:

  1. python3 test_watch.py

We can get results like the followings:

  1. in: [ 0.15727027  0.45887455  0.10939325  0.66666406 -0.62354755]
  2. out: [0.15727027 0.45887455 0.10939325 0.66666406 0.        ]

Code Explanation

In the example, we focus on y in ReluJob. Thus, we call flow.watch(y, watch_handler) to monitor y. The function oneflow.watch needs two parameters:

  • The first parameter is y which we focus on.

  • The second parameter is a callback function. When OneFlow use device resources to execute ReluJob, it will send y as a parameter to callback function. We define our callback function watch_handler to print out its parameters.

User can use customized callback function to process the data from OneFlow according to their own requirements.

Use watch_diff to Obtain Gradient when Running

The following is an example to demonstrate how to use oneflow.watch_diff to obtain the gradient at runtime.

Code: test_watch_diff.py

Run above code:

  1. python3 test_watch.py

We should have the following results:

  1. [ ...
  2.  [ 1.39966095e-03  3.49164731e-03  3.31605263e-02  4.50417027e-03
  3.    7.73609674e-04  4.89911772e-02  2.47627571e-02  7.65468649e-05
  4.   -1.18361652e-01  1.20161276e-03]] (100, 10) float32

Code Explanation

In the example above, we use oneflow.watch_diff to obtain the gradient. The processe is the same as the example which using oneflow.watch to obtain data tensor.

First, we define the callback function:

  1. def watch_diff_handler(blob: tp.Numpy):
  2.     print("watch_diff_handler:", blob, blob.shape, blob.dtype)

Then we use oneflow.watch_diff to register the callback function in job function:

  1. flow.watch_diff(logits, watch_diff_handler)

When running, OneFlow framework will call watch_diff_handler and send the gradient corresponding with logits to watch_diff_handler.

Please activate JavaScript for write a comment in LiveRe