Missed TensorFlow Dev Summit? Check out the video playlist. Watch recordings

tf.recompute_grad

TensorFlow 1 version View source on GitHub

An eager-compatible version of recompute_grad.

tf.recompute_grad(
    f
)

For f(*args, **kwargs), this supports gradients with respect to args or kwargs, but kwargs are currently only supported in eager-mode. Note that for keras layer and model objects, this is handled automatically.

Args:

  • f: function f(*x) that returns a Tensor or sequence of Tensor outputs.

Returns:

A function g that wraps f, but which recomputes f on the backwards pass of a gradient call.