flax.linen.vjp#

flax.linen.vjp(fn, mdl, *primals, has_aux=False, reduce_axes=(), vjp_variables='params', variables=True, rngs=True)[source]#

A lifted version of jax.vjp.

See jax.vjp for the unlifted vector-Jacobiam product (backward gradient).

Note that a gradient is returned for all variables in the collections specified by vjp_variables. However, the backward funtion only expects a cotangent for the return value of fn. If variables require a co-tangent as well they can be returned from fn using Module.variables.

Example:

class LearnScale(nn.Module):
  @nn.compact
  def __call__(self, x):
    p = self.param('scale', nn.initializers.zeros, ())
    return p * x

class Foo(nn.Module):
  @nn.compact
  def __call__(self, x):
    y, bwd = nn.vjp(lambda mdl, x: mdl(x), LearnScale(), x)
    params_grad, x_grad = bwd(jnp.ones(y.shape))
    return y, params_grad, x_grad
Parameters

  • fn – Function to be differentiated. Its arguments should be arrays, scalars, or standard Python containers of arrays or scalars. It should return an array, scalar, or standard Python container of arrays or scalars. It will receive the scope and primals as arguments.

  • mdl – The module of which the variables will be differentiated.

  • *primals – A sequence of primal values at which the Jacobian of fn should be evaluated. The length of primals should be equal to the number of positional parameters to fn. Each primal value should be a tuple of arrays, scalar, or standard Python containers thereof.

  • has_aux – Optional, bool. Indicates whether fn returns a pair where the first element is considered the output of the mathematical function to be differentiated and the second element is auxiliary data. Default False.

  • reduce_axes – Optional, tuple of axis names. If an axis is listed here, and fn implicitly broadcasts a value over that axis, the backward pass will perform a psum of the corresponding gradient. Otherwise, the VJP will be per-example over named axes. For example, if 'batch' is a named batch axis, vjp(f, *args, reduce_axes=('batch',)) will create a VJP function that sums over the batch while vjp(f, *args) will create a per-example VJP.

  • vjp_variables – The vjpfun will return a cotangent vector for all variable collections specified by this filter.

  • variables – other variables collections that are available inside fn but do not receive a cotangent.

  • rngs – the prngs that are available inside fn.

Returns

If has_aux is False, returns a (primals_out, vjpfun) pair, where primals_out is fn(*primals). vjpfun is a function from a cotangent vector with the same shape as primals_out to a tuple of cotangent vectors with the same shape as primals, representing the vector-Jacobian product of fn evaluated at primals. If has_aux is True, returns a (primals_out, vjpfun, aux) tuple where aux is the auxiliary data returned by fn.