|
| 1 | +from functools import partial |
| 2 | +import jax |
| 3 | + |
| 4 | +jax.config.update("jax_enable_x64", True) # double precision |
| 5 | +from jax import Array, random |
| 6 | +from jax import numpy as jnp |
| 7 | +from jsrm.utils.numerical_jacobian import approx_derivative |
| 8 | + |
| 9 | + |
| 10 | +def test_finite_differences(method = "2-point"): |
| 11 | + def fun(x: Array): |
| 12 | + return jnp.stack([x[0] * jnp.sin(x[1]), x[0] * jnp.cos(x[1])]) |
| 13 | + |
| 14 | + jac_autodiff_fn = jax.jacfwd(fun) |
| 15 | + jac_numdiff_fn = partial(approx_derivative, fun, method=method) |
| 16 | + |
| 17 | + rng = random.PRNGKey(0) |
| 18 | + for i in range(100): |
| 19 | + rng, subrng = random.split(rng) |
| 20 | + x = random.uniform(subrng, (2,), minval=-1.0, maxval=1.0) |
| 21 | + |
| 22 | + jac_autodiff = jac_autodiff_fn(x) |
| 23 | + jac_numdiff = jac_numdiff_fn(x) |
| 24 | + print("x = ", x, "\njac_autodiff = \n", jac_autodiff, "\njac_numdiff = \n", jac_numdiff) |
| 25 | + |
| 26 | + error_jac = jnp.linalg.norm(jac_autodiff - jac_numdiff) |
| 27 | + print("error_jac = ", error_jac) |
| 28 | + |
| 29 | + if not jnp.allclose(jac_autodiff, jac_numdiff, atol=1e-6): |
| 30 | + raise ValueError("Jacobian mismatch!") |
| 31 | + |
| 32 | + |
| 33 | +if __name__ == "__main__": |
| 34 | + test_finite_differences(method="2-point") |
| 35 | + test_finite_differences(method="3-point") |
0 commit comments