Decorator to define a function with a custom gradient.

This decorator allows fine grained control over the gradients of a sequence for operations. This may be useful for multiple reasons, including providing a more efficient or numerically stable gradient for a sequence of operations.

Usage

op_custom_gradient(f)

Arguments

f

Function f(...) that returns a tuple (output, grad_fn) where:

... is a sequence of unnamed arguments, each a tensor input or nested structure of tensor inputs to the function.
output is a (potentially nested structure of) tensor outputs of applying operations in forward_fn f() to ....
grad_fn is a function with the signature grad_fn(..., upstream) which returns a list of tensors the same size as (flattened) ...: the derivatives of tensors in output with respect to the tensors in .... upstream is a tensor or sequence of tensors holding the initial value gradients for each tensor in output.

Value

A function h(...) which returns the same value as f(...)[[1]] and whose gradient is determined by f(...)[[2]].

Note

Note that the grad function that returns gradient computation requires ... as well as an upstream named argument, depending on the backend being set. With the JAX and TensorFlow backends, it requires only one argument, whereas it might use the upstream argument in the case of the PyTorch backend.

When working with TensorFlow/JAX backend, grad(upstream) is sufficient. With PyTorch, the grad function requires ... as well as upstream, e.g. grad <- \(..., upstream). Follow the example above to use op_custom_gradient() in a way that is compatible with all backends.

Example

Backend-agnostic example.

log1pexp <- op_custom_gradient(\(x) {

    e <- op_exp(x)

    grad <- function(..., upstream = NULL) {
      upstream <- upstream %||% ..1
      op_multiply(upstream, 1.0 - 1.0 / op_add(1, e))
    }

    tuple(op_log(1 + e), grad)

})

if(config_backend() == "tensorflow") {
  tf <- tensorflow::tf
  x <- op_convert_to_tensor(100.0)
  with(tf$GradientTape() %as% tape, {
    tape$watch(x)
    y <- log1pexp(x)
  })
  dy_dx <- tape$gradient(y, x)
  stopifnot(as.numeric(dy_dx) == 1)
}

Other core ops:
op_associative_scan()
op_cast()
op_cond()
op_convert_to_numpy()
op_convert_to_tensor()
op_dtype()
op_fori_loop()
op_is_tensor()
op_map()
op_scan()
op_scatter()
op_scatter_update()
op_searchsorted()
op_shape()
op_slice()
op_slice_update()
op_stop_gradient()
op_switch()
op_unstack()
op_vectorized_map()
op_while_loop()

Other ops:
op_abs()
op_add()
op_all()
op_any()
op_append()
op_arange()
op_arccos()
op_arccosh()
op_arcsin()
op_arcsinh()
op_arctan()
op_arctan2()
op_arctanh()
op_argmax()
op_argmin()
op_argpartition()
op_argsort()
op_array()
op_associative_scan()
op_average()
op_average_pool()
op_batch_normalization()
op_binary_crossentropy()
op_bincount()
op_broadcast_to()
op_cast()
op_categorical_crossentropy()
op_ceil()
op_cholesky()
op_clip()
op_concatenate()
op_cond()
op_conj()
op_conv()
op_conv_transpose()
op_convert_to_numpy()
op_convert_to_tensor()
op_copy()
op_correlate()
op_cos()
op_cosh()
op_count_nonzero()
op_cross()
op_ctc_decode()
op_ctc_loss()
op_cumprod()
op_cumsum()
op_depthwise_conv()
op_det()
op_diag()
op_diagonal()
op_diff()
op_digitize()
op_divide()
op_divide_no_nan()
op_dot()
op_dtype()
op_eig()
op_eigh()
op_einsum()
op_elu()
op_empty()
op_equal()
op_erf()
op_erfinv()
op_exp()
op_expand_dims()
op_expm1()
op_extract_sequences()
op_eye()
op_fft()
op_fft2()
op_flip()
op_floor()
op_floor_divide()
op_fori_loop()
op_full()
op_full_like()
op_gelu()
op_get_item()
op_greater()
op_greater_equal()
op_hard_sigmoid()
op_hard_silu()
op_hstack()
op_identity()
op_imag()
op_image_affine_transform()
op_image_crop()
op_image_extract_patches()
op_image_hsv_to_rgb()
op_image_map_coordinates()
op_image_pad()
op_image_resize()
op_image_rgb_to_grayscale()
op_image_rgb_to_hsv()
op_in_top_k()
op_inv()
op_irfft()
op_is_tensor()
op_isclose()
op_isfinite()
op_isinf()
op_isnan()
op_istft()
op_leaky_relu()
op_less()
op_less_equal()
op_linspace()
op_log()
op_log10()
op_log1p()
op_log2()
op_log_sigmoid()
op_log_softmax()
op_logaddexp()
op_logical_and()
op_logical_not()
op_logical_or()
op_logical_xor()
op_logspace()
op_logsumexp()
op_lstsq()
op_lu_factor()
op_map()
op_matmul()
op_max()
op_max_pool()
op_maximum()
op_mean()
op_median()
op_meshgrid()
op_min()
op_minimum()
op_mod()
op_moments()
op_moveaxis()
op_multi_hot()
op_multiply()
op_nan_to_num()
op_ndim()
op_negative()
op_nonzero()
op_norm()
op_normalize()
op_not_equal()
op_one_hot()
op_ones()
op_ones_like()
op_outer()
op_pad()
op_power()
op_prod()
op_psnr()
op_qr()
op_quantile()
op_ravel()
op_real()
op_reciprocal()
op_relu()
op_relu6()
op_repeat()
op_reshape()
op_rfft()
op_roll()
op_round()
op_rsqrt()
op_scan()
op_scatter()
op_scatter_update()
op_searchsorted()
op_segment_max()
op_segment_sum()
op_select()
op_selu()
op_separable_conv()
op_shape()
op_sigmoid()
op_sign()
op_silu()
op_sin()
op_sinh()
op_size()
op_slice()
op_slice_update()
op_slogdet()
op_softmax()
op_softplus()
op_softsign()
op_solve()
op_solve_triangular()
op_sort()
op_sparse_categorical_crossentropy()
op_split()
op_sqrt()
op_square()
op_squeeze()
op_stack()
op_std()
op_stft()
op_stop_gradient()
op_subtract()
op_sum()
op_svd()
op_swapaxes()
op_switch()
op_take()
op_take_along_axis()
op_tan()
op_tanh()
op_tensordot()
op_tile()
op_top_k()
op_trace()
op_transpose()
op_tri()
op_tril()
op_triu()
op_unstack()
op_var()
op_vdot()
op_vectorize()
op_vectorized_map()
op_vstack()
op_where()
op_while_loop()
op_zeros()
op_zeros_like()